update input for loop blocks, do not need to include intermediate

* cfg; slg; pag; sdxl without controlnet

---------

Co-authored-by: Aryan <aryan@huggingface.co>

src/diffusers/__init__.py

@@ -34,10 +34,12 @@ from .utils import (
 
 _import_structure = {
     "configuration_utils": ["ConfigMixin"],
+    "guiders": [],
     "hooks": [],
     "loaders": ["FromOriginalModelMixin"],
     "models": [],
     "pipelines": [],
+    "modular_pipelines": [],
     "quantizers.quantization_config": [],
     "schedulers": [],
     "utils": [
@@ -130,12 +132,26 @@ except OptionalDependencyNotAvailable:
     _import_structure["utils.dummy_pt_objects"] = [name for name in dir(dummy_pt_objects) if not name.startswith("_")]
 
 else:
+    _import_structure["guiders"].extend(
+        [
+            "AdaptiveProjectedGuidance",
+            "AutoGuidance",
+            "ClassifierFreeGuidance",
+            "ClassifierFreeZeroStarGuidance",
+            "SkipLayerGuidance",
+            "SmoothedEnergyGuidance",
+            "TangentialClassifierFreeGuidance",
+        ]
+    )
     _import_structure["hooks"].extend(
         [
             "FasterCacheConfig",
             "HookRegistry",
             "PyramidAttentionBroadcastConfig",
+            "LayerSkipConfig",
+            "SmoothedEnergyGuidanceConfig",
             "apply_faster_cache",
+            "apply_layer_skip",
             "apply_pyramid_attention_broadcast",
         ]
     )
@@ -245,6 +261,13 @@ else:
             "StableDiffusionMixin",
         ]
     )
+    _import_structure["modular_pipelines"].extend(
+        [
+            "ModularLoader",
+            "ComponentSpec",
+            "ComponentsManager",
+        ]
+    )
     _import_structure["quantizers"] = ["DiffusersQuantizer"]
     _import_structure["schedulers"].extend(
         [
@@ -523,6 +546,24 @@ else:
         ]
     )
 
+
+try:
+    if not (is_torch_available() and is_transformers_available()):
+        raise OptionalDependencyNotAvailable()
+except OptionalDependencyNotAvailable:
+    from .utils import dummy_torch_and_transformers_objects  # noqa F403
+
+    _import_structure["utils.dummy_torch_and_transformers_objects"] = [
+        name for name in dir(dummy_torch_and_transformers_objects) if not name.startswith("_")
+    ]
+
+else:
+    _import_structure["modular_pipelines"].extend(
+        [
+            "StableDiffusionXLAutoPipeline",
+            "StableDiffusionXLModularLoader",
+        ]
+    )
 try:
     if not (is_torch_available() and is_transformers_available() and is_opencv_available()):
         raise OptionalDependencyNotAvailable()
@@ -728,10 +769,22 @@ if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT:
     except OptionalDependencyNotAvailable:
         from .utils.dummy_pt_objects import *  # noqa F403
     else:
+        from .guiders import (
+            AdaptiveProjectedGuidance,
+            AutoGuidance,
+            ClassifierFreeGuidance,
+            ClassifierFreeZeroStarGuidance,
+            SkipLayerGuidance,
+            SmoothedEnergyGuidance,
+            TangentialClassifierFreeGuidance,
+        )
         from .hooks import (
             FasterCacheConfig,
             HookRegistry,
+            LayerSkipConfig,
             PyramidAttentionBroadcastConfig,
+            SmoothedEnergyGuidanceConfig,
+            apply_layer_skip,
             apply_faster_cache,
             apply_pyramid_attention_broadcast,
         )
@@ -839,6 +892,11 @@ if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT:
             ScoreSdeVePipeline,
             StableDiffusionMixin,
         )
+        from .modular_pipelines import (
+            ModularLoader,
+            ComponentSpec,
+            ComponentsManager,
+        )
         from .quantizers import DiffusersQuantizer
         from .schedulers import (
             AmusedScheduler,
@@ -1094,7 +1152,16 @@ if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT:
             WuerstchenDecoderPipeline,
             WuerstchenPriorPipeline,
         )
-
+    try:
+        if not (is_torch_available() and is_transformers_available()):
+            raise OptionalDependencyNotAvailable()
+    except OptionalDependencyNotAvailable:
+        from .utils.dummy_torch_and_transformers_objects import *  # noqa F403
+    else:
+        from .modular_pipelines import (
+            StableDiffusionXLAutoPipeline,
+            StableDiffusionXLModularLoader,
+        )
     try:
         if not (is_torch_available() and is_transformers_available() and is_k_diffusion_available()):
             raise OptionalDependencyNotAvailable()

src/diffusers/guiders/__init__.py

@@ -0,0 +1,29 @@
+# Copyright 2024 The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from typing import Union
+
+from ..utils import is_torch_available
+
+
+if is_torch_available():
+    from .adaptive_projected_guidance import AdaptiveProjectedGuidance
+    from .auto_guidance import AutoGuidance
+    from .classifier_free_guidance import ClassifierFreeGuidance
+    from .classifier_free_zero_star_guidance import ClassifierFreeZeroStarGuidance
+    from .skip_layer_guidance import SkipLayerGuidance
+    from .smoothed_energy_guidance import SmoothedEnergyGuidance
+    from .tangential_classifier_free_guidance import TangentialClassifierFreeGuidance
+
+    GuiderType = Union[AdaptiveProjectedGuidance, AutoGuidance, ClassifierFreeGuidance, ClassifierFreeZeroStarGuidance, SkipLayerGuidance, SmoothedEnergyGuidance, TangentialClassifierFreeGuidance]

src/diffusers/guiders/adaptive_projected_guidance.py

@@ -0,0 +1,184 @@
+# Copyright 2024 The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import math
+from typing import Optional, List, TYPE_CHECKING, Dict, Union, Tuple
+
+import torch
+
+from .guider_utils import BaseGuidance, rescale_noise_cfg
+
+if TYPE_CHECKING:
+    from ..modular_pipelines.modular_pipeline import BlockState
+
+
+class AdaptiveProjectedGuidance(BaseGuidance):
+    """
+    Adaptive Projected Guidance (APG): https://huggingface.co/papers/2410.02416
+    
+    Args:
+        guidance_scale (`float`, defaults to `7.5`):
+            The scale parameter for classifier-free guidance. Higher values result in stronger conditioning on the text
+            prompt, while lower values allow for more freedom in generation. Higher values may lead to saturation and
+            deterioration of image quality.
+        adaptive_projected_guidance_momentum (`float`, defaults to `None`):
+            The momentum parameter for the adaptive projected guidance. Disabled if set to `None`.
+        adaptive_projected_guidance_rescale (`float`, defaults to `15.0`):
+            The rescale factor applied to the noise predictions. This is used to improve image quality and fix
+        guidance_rescale (`float`, defaults to `0.0`):
+            The rescale factor applied to the noise predictions. This is used to improve image quality and fix
+            overexposure. Based on Section 3.4 from [Common Diffusion Noise Schedules and Sample Steps are
+            Flawed](https://huggingface.co/papers/2305.08891).
+        use_original_formulation (`bool`, defaults to `False`):
+            Whether to use the original formulation of classifier-free guidance as proposed in the paper. By default,
+            we use the diffusers-native implementation that has been in the codebase for a long time. See
+            [~guiders.classifier_free_guidance.ClassifierFreeGuidance] for more details.
+        start (`float`, defaults to `0.0`):
+            The fraction of the total number of denoising steps after which guidance starts.
+        stop (`float`, defaults to `1.0`):
+            The fraction of the total number of denoising steps after which guidance stops.
+    """
+
+    _input_predictions = ["pred_cond", "pred_uncond"]
+
+    def __init__(
+        self,
+        guidance_scale: float = 7.5,
+        adaptive_projected_guidance_momentum: Optional[float] = None,
+        adaptive_projected_guidance_rescale: float = 15.0,
+        eta: float = 1.0,
+        guidance_rescale: float = 0.0,
+        use_original_formulation: bool = False,
+        start: float = 0.0,
+        stop: float = 1.0,
+    ):
+        super().__init__(start, stop)
+
+        self.guidance_scale = guidance_scale
+        self.adaptive_projected_guidance_momentum = adaptive_projected_guidance_momentum
+        self.adaptive_projected_guidance_rescale = adaptive_projected_guidance_rescale
+        self.eta = eta
+        self.guidance_rescale = guidance_rescale
+        self.use_original_formulation = use_original_formulation
+        self.momentum_buffer = None
+
+    def prepare_inputs(self, data: "BlockState", input_fields: Optional[Dict[str, Union[str, Tuple[str, str]]]] = None) -> List["BlockState"]:
+        
+        if input_fields is None:
+            input_fields = self._input_fields
+        
+        if self._step == 0:
+            if self.adaptive_projected_guidance_momentum is not None:
+                self.momentum_buffer = MomentumBuffer(self.adaptive_projected_guidance_momentum)
+        tuple_indices = [0] if self.num_conditions == 1 else [0, 1]
+        data_batches = []
+        for i in range(self.num_conditions):
+            data_batch = self._prepare_batch(input_fields, data, tuple_indices[i], self._input_predictions[i])
+            data_batches.append(data_batch)
+        return data_batches
+
+    def forward(self, pred_cond: torch.Tensor, pred_uncond: Optional[torch.Tensor] = None) -> torch.Tensor:
+        pred = None
+
+        if not self._is_apg_enabled():
+            pred = pred_cond
+        else:
+            pred = normalized_guidance(
+                pred_cond,
+                pred_uncond,
+                self.guidance_scale,
+                self.momentum_buffer,
+                self.eta,
+                self.adaptive_projected_guidance_rescale,
+                self.use_original_formulation,
+            )
+
+        if self.guidance_rescale > 0.0:
+            pred = rescale_noise_cfg(pred, pred_cond, self.guidance_rescale)
+
+        return pred, {}
+
+    @property
+    def is_conditional(self) -> bool:
+        return self._count_prepared == 1
+
+    @property
+    def num_conditions(self) -> int:
+        num_conditions = 1
+        if self._is_apg_enabled():
+            num_conditions += 1
+        return num_conditions
+
+    def _is_apg_enabled(self) -> bool:
+        if not self._enabled:
+            return False
+        
+        is_within_range = True
+        if self._num_inference_steps is not None:
+            skip_start_step = int(self._start * self._num_inference_steps)
+            skip_stop_step = int(self._stop * self._num_inference_steps)
+            is_within_range = skip_start_step <= self._step < skip_stop_step
+        
+        is_close = False
+        if self.use_original_formulation:
+            is_close = math.isclose(self.guidance_scale, 0.0)
+        else:
+            is_close = math.isclose(self.guidance_scale, 1.0)
+        
+        return is_within_range and not is_close
+
+
+class MomentumBuffer:
+    def __init__(self, momentum: float):
+        self.momentum = momentum
+        self.running_average = 0
+
+    def update(self, update_value: torch.Tensor):
+        new_average = self.momentum * self.running_average
+        self.running_average = update_value + new_average
+
+
+def normalized_guidance(
+    pred_cond: torch.Tensor,
+    pred_uncond: torch.Tensor,
+    guidance_scale: float,
+    momentum_buffer: Optional[MomentumBuffer] = None,
+    eta: float = 1.0,
+    norm_threshold: float = 0.0,
+    use_original_formulation: bool = False,
+):
+    diff = pred_cond - pred_uncond
+    dim = [-i for i in range(1, len(diff.shape))]
+    
+    if momentum_buffer is not None:
+        momentum_buffer.update(diff)
+        diff = momentum_buffer.running_average
+    
+    if norm_threshold > 0:
+        ones = torch.ones_like(diff)
+        diff_norm = diff.norm(p=2, dim=dim, keepdim=True)
+        scale_factor = torch.minimum(ones, norm_threshold / diff_norm)
+        diff = diff * scale_factor
+    
+    v0, v1 = diff.double(), pred_cond.double()
+    v1 = torch.nn.functional.normalize(v1, dim=dim)
+    v0_parallel = (v0 * v1).sum(dim=dim, keepdim=True) * v1
+    v0_orthogonal = v0 - v0_parallel
+    diff_parallel, diff_orthogonal = v0_parallel.type_as(diff), v0_orthogonal.type_as(diff)
+    normalized_update = diff_orthogonal + eta * diff_parallel
+    
+    pred = pred_cond if use_original_formulation else pred_uncond
+    pred = pred + guidance_scale * normalized_update
+    
+    return pred

src/diffusers/guiders/auto_guidance.py

@@ -0,0 +1,177 @@
+# Copyright 2024 The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import math
+from typing import List, Optional, Union, TYPE_CHECKING, Dict, Tuple
+
+import torch
+
+from ..hooks import HookRegistry, LayerSkipConfig
+from ..hooks.layer_skip import _apply_layer_skip_hook
+from .guider_utils import BaseGuidance, rescale_noise_cfg
+
+if TYPE_CHECKING:
+    from ..modular_pipelines.modular_pipeline import BlockState
+
+
+class AutoGuidance(BaseGuidance):
+    """
+    AutoGuidance: https://huggingface.co/papers/2406.02507
+    
+    Args:
+        guidance_scale (`float`, defaults to `7.5`):
+            The scale parameter for classifier-free guidance. Higher values result in stronger conditioning on the text
+            prompt, while lower values allow for more freedom in generation. Higher values may lead to saturation and
+            deterioration of image quality.
+        auto_guidance_layers (`int` or `List[int]`, *optional*):
+            The layer indices to apply skip layer guidance to. Can be a single integer or a list of integers. If not
+            provided, `skip_layer_config` must be provided.
+        auto_guidance_config (`LayerSkipConfig` or `List[LayerSkipConfig]`, *optional*):
+            The configuration for the skip layer guidance. Can be a single `LayerSkipConfig` or a list of
+            `LayerSkipConfig`. If not provided, `skip_layer_guidance_layers` must be provided.
+        dropout (`float`, *optional*):
+            The dropout probability for autoguidance on the enabled skip layers (either with `auto_guidance_layers` or
+            `auto_guidance_config`). If not provided, the dropout probability will be set to 1.0.
+        guidance_rescale (`float`, defaults to `0.0`):
+            The rescale factor applied to the noise predictions. This is used to improve image quality and fix
+            overexposure. Based on Section 3.4 from [Common Diffusion Noise Schedules and Sample Steps are
+            Flawed](https://huggingface.co/papers/2305.08891).
+        use_original_formulation (`bool`, defaults to `False`):
+            Whether to use the original formulation of classifier-free guidance as proposed in the paper. By default,
+            we use the diffusers-native implementation that has been in the codebase for a long time. See
+            [~guiders.classifier_free_guidance.ClassifierFreeGuidance] for more details.
+        start (`float`, defaults to `0.0`):
+            The fraction of the total number of denoising steps after which guidance starts.
+        stop (`float`, defaults to `1.0`):
+            The fraction of the total number of denoising steps after which guidance stops.
+    """
+
+    _input_predictions = ["pred_cond", "pred_uncond"]
+
+    def __init__(
+        self,
+        guidance_scale: float = 7.5,
+        auto_guidance_layers: Optional[Union[int, List[int]]] = None,
+        auto_guidance_config: Union[LayerSkipConfig, List[LayerSkipConfig]] = None,
+        dropout: Optional[float] = None,
+        guidance_rescale: float = 0.0,
+        use_original_formulation: bool = False,
+        start: float = 0.0,
+        stop: float = 1.0,
+    ):
+        super().__init__(start, stop)
+
+        self.guidance_scale = guidance_scale
+        self.auto_guidance_layers = auto_guidance_layers
+        self.auto_guidance_config = auto_guidance_config
+        self.dropout = dropout
+        self.guidance_rescale = guidance_rescale
+        self.use_original_formulation = use_original_formulation
+
+        if auto_guidance_layers is None and auto_guidance_config is None:
+            raise ValueError(
+                "Either `auto_guidance_layers` or `auto_guidance_config` must be provided to enable Skip Layer Guidance."
+            )
+        if auto_guidance_layers is not None and auto_guidance_config is not None:
+            raise ValueError("Only one of `auto_guidance_layers` or `auto_guidance_config` can be provided.")
+        if (dropout is None and auto_guidance_layers is not None) or (dropout is not None and auto_guidance_layers is None):
+            raise ValueError("`dropout` must be provided if `auto_guidance_layers` is provided.")
+
+        if auto_guidance_layers is not None:
+            if isinstance(auto_guidance_layers, int):
+                auto_guidance_layers = [auto_guidance_layers]
+            if not isinstance(auto_guidance_layers, list):
+                raise ValueError(
+                    f"Expected `auto_guidance_layers` to be an int or a list of ints, but got {type(auto_guidance_layers)}."
+                )
+            auto_guidance_config = [LayerSkipConfig(layer, fqn="auto", dropout=dropout) for layer in auto_guidance_layers]
+
+        if isinstance(auto_guidance_config, LayerSkipConfig):
+            auto_guidance_config = [auto_guidance_config]
+
+        if not isinstance(auto_guidance_config, list):
+            raise ValueError(
+                f"Expected `auto_guidance_config` to be a LayerSkipConfig or a list of LayerSkipConfig, but got {type(auto_guidance_config)}."
+            )
+
+        self.auto_guidance_config = auto_guidance_config
+        self._auto_guidance_hook_names = [f"AutoGuidance_{i}" for i in range(len(self.auto_guidance_config))]
+
+    def prepare_models(self, denoiser: torch.nn.Module) -> None:
+        self._count_prepared += 1
+        if self._is_ag_enabled() and self.is_unconditional:
+            for name, config in zip(self._auto_guidance_hook_names, self.auto_guidance_config):
+                _apply_layer_skip_hook(denoiser, config, name=name)
+    
+    def cleanup_models(self, denoiser: torch.nn.Module) -> None:
+        if self._is_ag_enabled() and self.is_unconditional:
+            for name in self._auto_guidance_hook_names:
+                registry = HookRegistry.check_if_exists_or_initialize(denoiser)
+                registry.remove_hook(name, recurse=True)
+    
+    def prepare_inputs(self, data: "BlockState", input_fields: Optional[Dict[str, Union[str, Tuple[str, str]]]] = None) -> List["BlockState"]:
+        
+        if input_fields is None:
+            input_fields = self._input_fields
+        
+        tuple_indices = [0] if self.num_conditions == 1 else [0, 1]
+        data_batches = []
+        for i in range(self.num_conditions):
+            data_batch = self._prepare_batch(input_fields, data, tuple_indices[i], self._input_predictions[i])
+            data_batches.append(data_batch)
+        return data_batches
+
+    def forward(self, pred_cond: torch.Tensor, pred_uncond: Optional[torch.Tensor] = None) -> torch.Tensor:
+        pred = None
+
+        if not self._is_ag_enabled():
+            pred = pred_cond
+        else:
+            shift = pred_cond - pred_uncond
+            pred = pred_cond if self.use_original_formulation else pred_uncond
+            pred = pred + self.guidance_scale * shift
+
+        if self.guidance_rescale > 0.0:
+            pred = rescale_noise_cfg(pred, pred_cond, self.guidance_rescale)
+        
+        return pred, {}
+    
+    @property
+    def is_conditional(self) -> bool:
+        return self._count_prepared == 1
+
+    @property
+    def num_conditions(self) -> int:
+        num_conditions = 1
+        if self._is_ag_enabled():
+            num_conditions += 1
+        return num_conditions
+
+    def _is_ag_enabled(self) -> bool:
+        if not self._enabled:
+            return False
+        
+        is_within_range = True
+        if self._num_inference_steps is not None:
+            skip_start_step = int(self._start * self._num_inference_steps)
+            skip_stop_step = int(self._stop * self._num_inference_steps)
+            is_within_range = skip_start_step <= self._step < skip_stop_step
+        
+        is_close = False
+        if self.use_original_formulation:
+            is_close = math.isclose(self.guidance_scale, 0.0)
+        else:
+            is_close = math.isclose(self.guidance_scale, 1.0)
+        
+        return is_within_range and not is_close

src/diffusers/guiders/classifier_free_guidance.py

@@ -0,0 +1,132 @@
+# Copyright 2024 The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import math
+from typing import Optional, List, TYPE_CHECKING, Dict, Union, Tuple
+
+import torch
+
+from .guider_utils import BaseGuidance, rescale_noise_cfg
+
+if TYPE_CHECKING:
+    from ..modular_pipelines.modular_pipeline import BlockState
+
+
+class ClassifierFreeGuidance(BaseGuidance):
+    """
+    Classifier-free guidance (CFG): https://huggingface.co/papers/2207.12598
+    
+    CFG is a technique used to improve generation quality and condition-following in diffusion models. It works by
+    jointly training a model on both conditional and unconditional data, and using a weighted sum of the two during
+    inference. This allows the model to tradeoff between generation quality and sample diversity.
+    The original paper proposes scaling and shifting the conditional distribution based on the difference between
+    conditional and unconditional predictions. [x_pred = x_cond + scale * (x_cond - x_uncond)]
+    
+    Diffusers implemented the scaling and shifting on the unconditional prediction instead based on the [Imagen
+    paper](https://huggingface.co/papers/2205.11487), which is equivalent to what the original paper proposed in
+    theory. [x_pred = x_uncond + scale * (x_cond - x_uncond)]
+    
+    The intution behind the original formulation can be thought of as moving the conditional distribution estimates
+    further away from the unconditional distribution estimates, while the diffusers-native implementation can be
+    thought of as moving the unconditional distribution towards the conditional distribution estimates to get rid of
+    the unconditional predictions (usually negative features like "bad quality, bad anotomy, watermarks", etc.)
+    
+    The `use_original_formulation` argument can be set to `True` to use the original CFG formulation mentioned in the
+    paper. By default, we use the diffusers-native implementation that has been in the codebase for a long time.
+    
+    Args:
+        guidance_scale (`float`, defaults to `7.5`):
+            The scale parameter for classifier-free guidance. Higher values result in stronger conditioning on the text
+            prompt, while lower values allow for more freedom in generation. Higher values may lead to saturation and
+            deterioration of image quality.
+        guidance_rescale (`float`, defaults to `0.0`):
+            The rescale factor applied to the noise predictions. This is used to improve image quality and fix
+            overexposure. Based on Section 3.4 from [Common Diffusion Noise Schedules and Sample Steps are
+            Flawed](https://huggingface.co/papers/2305.08891).
+        use_original_formulation (`bool`, defaults to `False`):
+            Whether to use the original formulation of classifier-free guidance as proposed in the paper. By default,
+            we use the diffusers-native implementation that has been in the codebase for a long time. See
+            [~guiders.classifier_free_guidance.ClassifierFreeGuidance] for more details.
+        start (`float`, defaults to `0.0`):
+            The fraction of the total number of denoising steps after which guidance starts.
+        stop (`float`, defaults to `1.0`):
+            The fraction of the total number of denoising steps after which guidance stops.
+    """
+
+    _input_predictions = ["pred_cond", "pred_uncond"]
+
+    def __init__(
+        self, guidance_scale: float = 7.5, guidance_rescale: float = 0.0, use_original_formulation: bool = False, start: float = 0.0, stop: float = 1.0
+    ):
+        super().__init__(start, stop)
+
+        self.guidance_scale = guidance_scale
+        self.guidance_rescale = guidance_rescale
+        self.use_original_formulation = use_original_formulation
+    
+    def prepare_inputs(self, data: "BlockState", input_fields: Optional[Dict[str, Union[str, Tuple[str, str]]]] = None) -> List["BlockState"]:
+        
+        if input_fields is None:
+            input_fields = self._input_fields
+        
+        tuple_indices = [0] if self.num_conditions == 1 else [0, 1]
+        data_batches = []
+        for i in range(self.num_conditions):
+            data_batch = self._prepare_batch(input_fields, data, tuple_indices[i], self._input_predictions[i])
+            data_batches.append(data_batch)
+        return data_batches
+
+    def forward(self, pred_cond: torch.Tensor, pred_uncond: Optional[torch.Tensor] = None) -> torch.Tensor:
+        pred = None
+
+        if not self._is_cfg_enabled():
+            pred = pred_cond
+        else:
+            shift = pred_cond - pred_uncond
+            pred = pred_cond if self.use_original_formulation else pred_uncond
+            pred = pred + self.guidance_scale * shift
+
+        if self.guidance_rescale > 0.0:
+            pred = rescale_noise_cfg(pred, pred_cond, self.guidance_rescale)
+
+        return pred, {}
+
+    @property
+    def is_conditional(self) -> bool:
+        return self._count_prepared == 1
+
+    @property
+    def num_conditions(self) -> int:
+        num_conditions = 1
+        if self._is_cfg_enabled():
+            num_conditions += 1
+        return num_conditions
+
+    def _is_cfg_enabled(self) -> bool:
+        if not self._enabled:
+            return False
+        
+        is_within_range = True
+        if self._num_inference_steps is not None:
+            skip_start_step = int(self._start * self._num_inference_steps)
+            skip_stop_step = int(self._stop * self._num_inference_steps)
+            is_within_range = skip_start_step <= self._step < skip_stop_step
+        
+        is_close = False
+        if self.use_original_formulation:
+            is_close = math.isclose(self.guidance_scale, 0.0)
+        else:
+            is_close = math.isclose(self.guidance_scale, 1.0)
+        
+        return is_within_range and not is_close

src/diffusers/guiders/classifier_free_zero_star_guidance.py

@@ -0,0 +1,148 @@
+# Copyright 2024 The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import math
+from typing import Optional, List, TYPE_CHECKING, Dict, Union, Tuple
+
+import torch
+
+from .guider_utils import BaseGuidance, rescale_noise_cfg
+
+if TYPE_CHECKING:
+    from ..modular_pipelines.modular_pipeline import BlockState
+
+
+class ClassifierFreeZeroStarGuidance(BaseGuidance):
+    """
+    Classifier-free Zero* (CFG-Zero*): https://huggingface.co/papers/2503.18886
+    
+    This is an implementation of the Classifier-Free Zero* guidance technique, which is a variant of classifier-free
+    guidance. It proposes zero initialization of the noise predictions for the first few steps of the diffusion
+    process, and also introduces an optimal rescaling factor for the noise predictions, which can help in improving the
+    quality of generated images.
+    
+    The authors of the paper suggest setting zero initialization in the first 4% of the inference steps.
+    
+    Args:
+        guidance_scale (`float`, defaults to `7.5`):
+            The scale parameter for classifier-free guidance. Higher values result in stronger conditioning on the text
+            prompt, while lower values allow for more freedom in generation. Higher values may lead to saturation and
+            deterioration of image quality.
+        zero_init_steps (`int`, defaults to `1`):
+            The number of inference steps for which the noise predictions are zeroed out (see Section 4.2).
+        guidance_rescale (`float`, defaults to `0.0`):
+            The rescale factor applied to the noise predictions. This is used to improve image quality and fix
+            overexposure. Based on Section 3.4 from [Common Diffusion Noise Schedules and Sample Steps are
+            Flawed](https://huggingface.co/papers/2305.08891).
+        use_original_formulation (`bool`, defaults to `False`):
+            Whether to use the original formulation of classifier-free guidance as proposed in the paper. By default,
+            we use the diffusers-native implementation that has been in the codebase for a long time. See
+            [~guiders.classifier_free_guidance.ClassifierFreeGuidance] for more details.
+        start (`float`, defaults to `0.01`):
+            The fraction of the total number of denoising steps after which guidance starts.
+        stop (`float`, defaults to `0.2`):
+            The fraction of the total number of denoising steps after which guidance stops.
+    """
+
+    _input_predictions = ["pred_cond", "pred_uncond"]
+
+    def __init__(
+        self,
+        guidance_scale: float = 7.5,
+        zero_init_steps: int = 1,
+        guidance_rescale: float = 0.0,
+        use_original_formulation: bool = False,
+        start: float = 0.0,
+        stop: float = 1.0,
+    ):
+        super().__init__(start, stop)
+
+        self.guidance_scale = guidance_scale
+        self.zero_init_steps = zero_init_steps
+        self.guidance_rescale = guidance_rescale
+        self.use_original_formulation = use_original_formulation
+    
+    def prepare_inputs(self, data: "BlockState", input_fields: Optional[Dict[str, Union[str, Tuple[str, str]]]] = None) -> List["BlockState"]:
+        
+        if input_fields is None:
+            input_fields = self._input_fields
+        
+        tuple_indices = [0] if self.num_conditions == 1 else [0, 1]
+        data_batches = []
+        for i in range(self.num_conditions):
+            data_batch = self._prepare_batch(input_fields, data, tuple_indices[i], self._input_predictions[i])
+            data_batches.append(data_batch)
+        return data_batches
+
+    def forward(self, pred_cond: torch.Tensor, pred_uncond: Optional[torch.Tensor] = None) -> torch.Tensor:
+        pred = None
+
+        if self._step < self.zero_init_steps:
+            pred = torch.zeros_like(pred_cond)
+        elif not self._is_cfg_enabled():
+            pred = pred_cond
+        else:
+            pred_cond_flat = pred_cond.flatten(1)
+            pred_uncond_flat = pred_uncond.flatten(1)
+            alpha = cfg_zero_star_scale(pred_cond_flat, pred_uncond_flat)
+            alpha = alpha.view(-1, *(1,) * (len(pred_cond.shape) - 1))
+            pred_uncond = pred_uncond * alpha
+            shift = pred_cond - pred_uncond
+            pred = pred_cond if self.use_original_formulation else pred_uncond
+            pred = pred + self.guidance_scale * shift
+
+        if self.guidance_rescale > 0.0:
+            pred = rescale_noise_cfg(pred, pred_cond, self.guidance_rescale)
+
+        return pred, {}
+    
+    @property
+    def is_conditional(self) -> bool:
+        return self._count_prepared == 1
+
+    @property
+    def num_conditions(self) -> int:
+        num_conditions = 1
+        if self._is_cfg_enabled():
+            num_conditions += 1
+        return num_conditions
+
+    def _is_cfg_enabled(self) -> bool:
+        if not self._enabled:
+            return False
+        
+        is_within_range = True
+        if self._num_inference_steps is not None:
+            skip_start_step = int(self._start * self._num_inference_steps)
+            skip_stop_step = int(self._stop * self._num_inference_steps)
+            is_within_range = skip_start_step <= self._step < skip_stop_step
+        
+        is_close = False
+        if self.use_original_formulation:
+            is_close = math.isclose(self.guidance_scale, 0.0)
+        else:
+            is_close = math.isclose(self.guidance_scale, 1.0)
+        
+        return is_within_range and not is_close
+
+
+def cfg_zero_star_scale(cond: torch.Tensor, uncond: torch.Tensor, eps: float = 1e-8) -> torch.Tensor:
+    cond_dtype = cond.dtype
+    cond = cond.float()
+    uncond = uncond.float()
+    dot_product = torch.sum(cond * uncond, dim=1, keepdim=True)
+    squared_norm = torch.sum(uncond**2, dim=1, keepdim=True) + eps
+    # st_star = v_cond^T * v_uncond / ||v_uncond||^2
+    scale = dot_product / squared_norm
+    return scale.to(dtype=cond_dtype)

src/diffusers/guiders/guider_utils.py

@@ -0,0 +1,215 @@
+# Copyright 2024 The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from typing import TYPE_CHECKING, Any, Dict, List, Tuple, Union
+
+import torch
+
+from ..utils import get_logger
+
+
+if TYPE_CHECKING:
+    from ..modular_pipelines.modular_pipeline import BlockState
+
+
+logger = get_logger(__name__)  # pylint: disable=invalid-name
+
+
+class BaseGuidance:
+    r"""Base class providing the skeleton for implementing guidance techniques."""
+
+    _input_predictions = None
+    _identifier_key = "__guidance_identifier__"
+
+    def __init__(self, start: float = 0.0, stop: float = 1.0):
+        self._start = start
+        self._stop = stop
+        self._step: int = None
+        self._num_inference_steps: int = None
+        self._timestep: torch.LongTensor = None
+        self._count_prepared = 0
+        self._input_fields: Dict[str, Union[str, Tuple[str, str]]] = None
+        self._enabled = True
+
+        if not (0.0 <= start < 1.0):
+            raise ValueError(
+                f"Expected `start` to be between 0.0 and 1.0, but got {start}."
+            )
+        if not (start <= stop <= 1.0):
+            raise ValueError(
+                f"Expected `stop` to be between {start} and 1.0, but got {stop}."
+            )
+
+        if self._input_predictions is None or not isinstance(self._input_predictions, list):
+            raise ValueError(
+                "`_input_predictions` must be a list of required prediction names for the guidance technique."
+            )
+
+    def disable(self):
+        self._enabled = False
+    
+    def enable(self):
+        self._enabled = True
+    
+    def set_state(self, step: int, num_inference_steps: int, timestep: torch.LongTensor) -> None:
+        self._step = step
+        self._num_inference_steps = num_inference_steps
+        self._timestep = timestep
+        self._count_prepared = 0
+
+    def set_input_fields(self, **kwargs: Dict[str, Union[str, Tuple[str, str]]]) -> None:
+        """
+        Set the input fields for the guidance technique. The input fields are used to specify the names of the
+        returned attributes containing the prepared data after `prepare_inputs` is called. The prepared data is
+        obtained from the values of the provided keyword arguments to this method.
+
+        Args:
+            **kwargs (`Dict[str, Union[str, Tuple[str, str]]]`):
+                A dictionary where the keys are the names of the fields that will be used to store the data once
+                it is prepared with `prepare_inputs`. The values can be either a string or a tuple of length 2,
+                which is used to look up the required data provided for preparation.
+
+                If a string is provided, it will be used as the conditional data (or unconditional if used with
+                a guidance method that requires it). If a tuple of length 2 is provided, the first element must
+                be the conditional data identifier and the second element must be the unconditional data identifier
+                or None.
+
+                Example:
+                
+                ```
+                data = {"prompt_embeds": <some tensor>, "negative_prompt_embeds": <some tensor>, "latents": <some tensor>}
+
+                BaseGuidance.set_input_fields(
+                    latents="latents",
+                    prompt_embeds=("prompt_embeds", "negative_prompt_embeds"),
+                )
+                ```
+        """
+        for key, value in kwargs.items():
+            is_string = isinstance(value, str)
+            is_tuple_of_str_with_len_2 = isinstance(value, tuple) and len(value) == 2 and all(isinstance(v, str) for v in value)
+            if not (is_string or is_tuple_of_str_with_len_2):
+                raise ValueError(
+                    f"Expected `set_input_fields` to be called with a string or a tuple of string with length 2, but got {type(value)} for key {key}."
+                )
+        self._input_fields = kwargs
+    
+    def prepare_models(self, denoiser: torch.nn.Module) -> None:
+        """
+        Prepares the models for the guidance technique on a given batch of data. This method should be overridden in
+        subclasses to implement specific model preparation logic.
+        """
+        self._count_prepared += 1
+    
+    def cleanup_models(self, denoiser: torch.nn.Module) -> None:
+        """
+        Cleans up the models for the guidance technique after a given batch of data. This method should be overridden in
+        subclasses to implement specific model cleanup logic. It is useful for removing any hooks or other stateful
+        modifications made during `prepare_models`.
+        """
+        pass
+    
+    def prepare_inputs(self, data: "BlockState") -> List["BlockState"]:
+        raise NotImplementedError("BaseGuidance::prepare_inputs must be implemented in subclasses.")
+
+    def __call__(self, data: List["BlockState"]) -> Any:
+        if not all(hasattr(d, "noise_pred") for d in data):
+            raise ValueError("Expected all data to have `noise_pred` attribute.")
+        if len(data) != self.num_conditions:
+            raise ValueError(
+                f"Expected {self.num_conditions} data items, but got {len(data)}. Please check the input data."
+            )
+        forward_inputs = {getattr(d, self._identifier_key): d.noise_pred for d in data}
+        return self.forward(**forward_inputs)
+
+    def forward(self, *args, **kwargs) -> Any:
+        raise NotImplementedError("BaseGuidance::forward must be implemented in subclasses.")
+
+    @property
+    def is_conditional(self) -> bool:
+        raise NotImplementedError("BaseGuidance::is_conditional must be implemented in subclasses.")
+    
+    @property
+    def is_unconditional(self) -> bool:
+        return not self.is_conditional
+    
+    @property
+    def num_conditions(self) -> int:
+        raise NotImplementedError("BaseGuidance::num_conditions must be implemented in subclasses.")
+    
+    @classmethod
+    def _prepare_batch(cls, input_fields: Dict[str, Union[str, Tuple[str, str]]], data: "BlockState", tuple_index: int, identifier: str) -> "BlockState":
+        """
+        Prepares a batch of data for the guidance technique. This method is used in the `prepare_inputs` method of
+        the `BaseGuidance` class. It prepares the batch based on the provided tuple index.
+
+        Args:
+            input_fields (`Dict[str, Union[str, Tuple[str, str]]]`):
+                A dictionary where the keys are the names of the fields that will be used to store the data once
+                it is prepared with `prepare_inputs`. The values can be either a string or a tuple of length 2,
+                which is used to look up the required data provided for preparation.
+                If a string is provided, it will be used as the conditional data (or unconditional if used with
+                a guidance method that requires it). If a tuple of length 2 is provided, the first element must
+                be the conditional data identifier and the second element must be the unconditional data identifier
+                or None.
+            data (`BlockState`):
+                The input data to be prepared.
+            tuple_index (`int`):
+                The index to use when accessing input fields that are tuples.
+        
+        Returns:
+            `BlockState`: The prepared batch of data.
+        """
+        from ..modular_pipelines.modular_pipeline import BlockState
+
+        if input_fields is None:
+            raise ValueError("Input fields cannot be None. Please pass `input_fields` to `prepare_inputs` or call `set_input_fields` before preparing inputs.")
+        data_batch = {}
+        for key, value in input_fields.items():
+            try:
+                if isinstance(value, str):
+                    data_batch[key] = getattr(data, value)
+                elif isinstance(value, tuple):
+                    data_batch[key] = getattr(data, value[tuple_index])
+                else:
+                    # We've already checked that value is a string or a tuple of strings with length 2
+                    pass
+            except AttributeError:
+                logger.debug(f"`data` does not have attribute(s) {value}, skipping.")
+        data_batch[cls._identifier_key] = identifier
+        return BlockState(**data_batch)
+
+
+def rescale_noise_cfg(noise_cfg, noise_pred_text, guidance_rescale=0.0):
+    r"""
+    Rescales `noise_cfg` tensor based on `guidance_rescale` to improve image quality and fix overexposure. Based on
+    Section 3.4 from [Common Diffusion Noise Schedules and Sample Steps are
+    Flawed](https://arxiv.org/pdf/2305.08891.pdf).
+    Args:
+        noise_cfg (`torch.Tensor`):
+            The predicted noise tensor for the guided diffusion process.
+        noise_pred_text (`torch.Tensor`):
+            The predicted noise tensor for the text-guided diffusion process.
+        guidance_rescale (`float`, *optional*, defaults to 0.0):
+            A rescale factor applied to the noise predictions.
+    Returns:
+        noise_cfg (`torch.Tensor`): The rescaled noise prediction tensor.
+    """
+    std_text = noise_pred_text.std(dim=list(range(1, noise_pred_text.ndim)), keepdim=True)
+    std_cfg = noise_cfg.std(dim=list(range(1, noise_cfg.ndim)), keepdim=True)
+    # rescale the results from guidance (fixes overexposure)
+    noise_pred_rescaled = noise_cfg * (std_text / std_cfg)
+    # mix with the original results from guidance by factor guidance_rescale to avoid "plain looking" images
+    noise_cfg = guidance_rescale * noise_pred_rescaled + (1 - guidance_rescale) * noise_cfg
+    return noise_cfg

src/diffusers/guiders/skip_layer_guidance.py

@@ -0,0 +1,251 @@
+# Copyright 2024 The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import math
+from typing import List, Optional, Union, TYPE_CHECKING, Dict, Tuple
+
+import torch
+
+from ..hooks import HookRegistry, LayerSkipConfig
+from ..hooks.layer_skip import _apply_layer_skip_hook
+from .guider_utils import BaseGuidance, rescale_noise_cfg
+
+if TYPE_CHECKING:
+    from ..modular_pipelines.modular_pipeline import BlockState
+
+
+class SkipLayerGuidance(BaseGuidance):
+    """
+    Skip Layer Guidance (SLG): https://github.com/Stability-AI/sd3.5
+    
+    Spatio-Temporal Guidance (STG): https://huggingface.co/papers/2411.18664
+    
+    SLG was introduced by StabilityAI for improving structure and anotomy coherence in generated images. It works by
+    skipping the forward pass of specified transformer blocks during the denoising process on an additional conditional
+    batch of data, apart from the conditional and unconditional batches already used in CFG
+    ([~guiders.classifier_free_guidance.ClassifierFreeGuidance]), and then scaling and shifting the CFG predictions
+    based on the difference between conditional without skipping and conditional with skipping predictions.
+    
+    The intution behind SLG can be thought of as moving the CFG predicted distribution estimates further away from
+    worse versions of the conditional distribution estimates (because skipping layers is equivalent to using a worse
+    version of the model for the conditional prediction).
+    
+    STG is an improvement and follow-up work combining ideas from SLG, PAG and similar techniques for improving
+    generation quality in video diffusion models.
+    
+    Additional reading:
+    - [Guiding a Diffusion Model with a Bad Version of Itself](https://huggingface.co/papers/2406.02507)
+    
+    The values for `skip_layer_guidance_scale`, `skip_layer_guidance_start`, and `skip_layer_guidance_stop` are
+    defaulted to the recommendations by StabilityAI for Stable Diffusion 3.5 Medium.
+    
+    Args:
+        guidance_scale (`float`, defaults to `7.5`):
+            The scale parameter for classifier-free guidance. Higher values result in stronger conditioning on the text
+            prompt, while lower values allow for more freedom in generation. Higher values may lead to saturation and
+            deterioration of image quality.
+        skip_layer_guidance_scale (`float`, defaults to `2.8`):
+            The scale parameter for skip layer guidance. Anatomy and structure coherence may improve with higher
+            values, but it may also lead to overexposure and saturation.
+        skip_layer_guidance_start (`float`, defaults to `0.01`):
+            The fraction of the total number of denoising steps after which skip layer guidance starts.
+        skip_layer_guidance_stop (`float`, defaults to `0.2`):
+            The fraction of the total number of denoising steps after which skip layer guidance stops.
+        skip_layer_guidance_layers (`int` or `List[int]`, *optional*):
+            The layer indices to apply skip layer guidance to. Can be a single integer or a list of integers. If not
+            provided, `skip_layer_config` must be provided. The recommended values are `[7, 8, 9]` for Stable Diffusion
+            3.5 Medium.
+        skip_layer_config (`LayerSkipConfig` or `List[LayerSkipConfig]`, *optional*):
+            The configuration for the skip layer guidance. Can be a single `LayerSkipConfig` or a list of
+            `LayerSkipConfig`. If not provided, `skip_layer_guidance_layers` must be provided.
+        guidance_rescale (`float`, defaults to `0.0`):
+            The rescale factor applied to the noise predictions. This is used to improve image quality and fix
+            overexposure. Based on Section 3.4 from [Common Diffusion Noise Schedules and Sample Steps are
+            Flawed](https://huggingface.co/papers/2305.08891).
+        use_original_formulation (`bool`, defaults to `False`):
+            Whether to use the original formulation of classifier-free guidance as proposed in the paper. By default,
+            we use the diffusers-native implementation that has been in the codebase for a long time. See
+            [~guiders.classifier_free_guidance.ClassifierFreeGuidance] for more details.
+        start (`float`, defaults to `0.01`):
+            The fraction of the total number of denoising steps after which guidance starts.
+        stop (`float`, defaults to `0.2`):
+            The fraction of the total number of denoising steps after which guidance stops.
+    """
+
+    _input_predictions = ["pred_cond", "pred_uncond", "pred_cond_skip"]
+
+    def __init__(
+        self,
+        guidance_scale: float = 7.5,
+        skip_layer_guidance_scale: float = 2.8,
+        skip_layer_guidance_start: float = 0.01,
+        skip_layer_guidance_stop: float = 0.2,
+        skip_layer_guidance_layers: Optional[Union[int, List[int]]] = None,
+        skip_layer_config: Union[LayerSkipConfig, List[LayerSkipConfig]] = None,
+        guidance_rescale: float = 0.0,
+        use_original_formulation: bool = False,
+        start: float = 0.0,
+        stop: float = 1.0,
+    ):
+        super().__init__(start, stop)
+
+        self.guidance_scale = guidance_scale
+        self.skip_layer_guidance_scale = skip_layer_guidance_scale
+        self.skip_layer_guidance_start = skip_layer_guidance_start
+        self.skip_layer_guidance_stop = skip_layer_guidance_stop
+        self.guidance_rescale = guidance_rescale
+        self.use_original_formulation = use_original_formulation
+
+        if not (0.0 <= skip_layer_guidance_start < 1.0):
+            raise ValueError(
+                f"Expected `skip_layer_guidance_start` to be between 0.0 and 1.0, but got {skip_layer_guidance_start}."
+            )
+        if not (skip_layer_guidance_start <= skip_layer_guidance_stop <= 1.0):
+            raise ValueError(
+                f"Expected `skip_layer_guidance_stop` to be between 0.0 and 1.0, but got {skip_layer_guidance_stop}."
+            )
+
+        if skip_layer_guidance_layers is None and skip_layer_config is None:
+            raise ValueError(
+                "Either `skip_layer_guidance_layers` or `skip_layer_config` must be provided to enable Skip Layer Guidance."
+            )
+        if skip_layer_guidance_layers is not None and skip_layer_config is not None:
+            raise ValueError("Only one of `skip_layer_guidance_layers` or `skip_layer_config` can be provided.")
+
+        if skip_layer_guidance_layers is not None:
+            if isinstance(skip_layer_guidance_layers, int):
+                skip_layer_guidance_layers = [skip_layer_guidance_layers]
+            if not isinstance(skip_layer_guidance_layers, list):
+                raise ValueError(
+                    f"Expected `skip_layer_guidance_layers` to be an int or a list of ints, but got {type(skip_layer_guidance_layers)}."
+                )
+            skip_layer_config = [LayerSkipConfig(layer, fqn="auto") for layer in skip_layer_guidance_layers]
+
+        if isinstance(skip_layer_config, LayerSkipConfig):
+            skip_layer_config = [skip_layer_config]
+
+        if not isinstance(skip_layer_config, list):
+            raise ValueError(
+                f"Expected `skip_layer_config` to be a LayerSkipConfig or a list of LayerSkipConfig, but got {type(skip_layer_config)}."
+            )
+
+        self.skip_layer_config = skip_layer_config
+        self._skip_layer_hook_names = [f"SkipLayerGuidance_{i}" for i in range(len(self.skip_layer_config))]
+
+    def prepare_models(self, denoiser: torch.nn.Module) -> None:
+        self._count_prepared += 1
+        if self._is_slg_enabled() and self.is_conditional and self._count_prepared > 1:
+            for name, config in zip(self._skip_layer_hook_names, self.skip_layer_config):
+                _apply_layer_skip_hook(denoiser, config, name=name)
+    
+    def cleanup_models(self, denoiser: torch.nn.Module) -> None:
+        if self._is_slg_enabled() and self.is_conditional and self._count_prepared > 1:
+            registry = HookRegistry.check_if_exists_or_initialize(denoiser)
+            # Remove the hooks after inference
+            for hook_name in self._skip_layer_hook_names:
+                registry.remove_hook(hook_name, recurse=True)
+    
+    def prepare_inputs(self, data: "BlockState", input_fields: Optional[Dict[str, Union[str, Tuple[str, str]]]] = None) -> List["BlockState"]:
+        
+        if input_fields is None:
+            input_fields = self._input_fields
+        
+        if self.num_conditions == 1:
+            tuple_indices = [0]
+            input_predictions = ["pred_cond"]
+        elif self.num_conditions == 2:
+            tuple_indices = [0, 1]
+            input_predictions = ["pred_cond", "pred_uncond"] if self._is_cfg_enabled() else ["pred_cond", "pred_cond_skip"]
+        else:
+            tuple_indices = [0, 1, 0]
+            input_predictions = ["pred_cond", "pred_uncond", "pred_cond_skip"]
+        data_batches = []
+        for i in range(self.num_conditions):
+            data_batch = self._prepare_batch(input_fields, data, tuple_indices[i], input_predictions[i])
+            data_batches.append(data_batch)
+        return data_batches
+
+    def forward(
+        self,
+        pred_cond: torch.Tensor,
+        pred_uncond: Optional[torch.Tensor] = None,
+        pred_cond_skip: Optional[torch.Tensor] = None,
+    ) -> torch.Tensor:
+        pred = None
+
+        if not self._is_cfg_enabled() and not self._is_slg_enabled():
+            pred = pred_cond
+        elif not self._is_cfg_enabled():
+            shift = pred_cond - pred_cond_skip
+            pred = pred_cond if self.use_original_formulation else pred_cond_skip
+            pred = pred + self.skip_layer_guidance_scale * shift
+        elif not self._is_slg_enabled():
+            shift = pred_cond - pred_uncond
+            pred = pred_cond if self.use_original_formulation else pred_uncond
+            pred = pred + self.guidance_scale * shift
+        else:
+            shift = pred_cond - pred_uncond
+            shift_skip = pred_cond - pred_cond_skip
+            pred = pred_cond if self.use_original_formulation else pred_uncond
+            pred = pred + self.guidance_scale * shift + self.skip_layer_guidance_scale * shift_skip
+
+        if self.guidance_rescale > 0.0:
+            pred = rescale_noise_cfg(pred, pred_cond, self.guidance_rescale)
+
+        return pred, {}
+    
+    @property
+    def is_conditional(self) -> bool:
+        return self._count_prepared == 1 or self._count_prepared == 3
+
+    @property
+    def num_conditions(self) -> int:
+        num_conditions = 1
+        if self._is_cfg_enabled():
+            num_conditions += 1
+        if self._is_slg_enabled():
+            num_conditions += 1
+        return num_conditions
+
+    def _is_cfg_enabled(self) -> bool:
+        if not self._enabled:
+            return False
+        
+        is_within_range = True
+        if self._num_inference_steps is not None:
+            skip_start_step = int(self._start * self._num_inference_steps)
+            skip_stop_step = int(self._stop * self._num_inference_steps)
+            is_within_range = skip_start_step <= self._step < skip_stop_step
+        
+        is_close = False
+        if self.use_original_formulation:
+            is_close = math.isclose(self.guidance_scale, 0.0)
+        else:
+            is_close = math.isclose(self.guidance_scale, 1.0)
+        
+        return is_within_range and not is_close
+
+    def _is_slg_enabled(self) -> bool:
+        if not self._enabled:
+            return False
+        
+        is_within_range = True
+        if self._num_inference_steps is not None:
+            skip_start_step = int(self.skip_layer_guidance_start * self._num_inference_steps)
+            skip_stop_step = int(self.skip_layer_guidance_stop * self._num_inference_steps)
+            is_within_range = skip_start_step < self._step < skip_stop_step
+        
+        is_zero = math.isclose(self.skip_layer_guidance_scale, 0.0)
+        
+        return is_within_range and not is_zero

src/diffusers/guiders/smoothed_energy_guidance.py

@@ -0,0 +1,244 @@
+# Copyright 2024 The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import math
+from typing import List, Optional, Union, TYPE_CHECKING, Dict, Tuple
+
+import torch
+
+from ..hooks import HookRegistry
+from ..hooks.smoothed_energy_guidance_utils import SmoothedEnergyGuidanceConfig, _apply_smoothed_energy_guidance_hook
+from .guider_utils import BaseGuidance, rescale_noise_cfg
+
+if TYPE_CHECKING:
+    from ..modular_pipelines.modular_pipeline import BlockState
+
+
+class SmoothedEnergyGuidance(BaseGuidance):
+    """
+    Smoothed Energy Guidance (SEG): https://huggingface.co/papers/2408.00760
+
+    SEG is only supported as an experimental prototype feature for now, so the implementation may be modified
+    in the future without warning or guarantee of reproducibility. This implementation assumes:
+    - Generated images are square (height == width)
+    - The model does not combine different modalities together (e.g., text and image latent streams are
+      not combined together such as Flux)
+    
+    Args:
+        guidance_scale (`float`, defaults to `7.5`):
+            The scale parameter for classifier-free guidance. Higher values result in stronger conditioning on the text
+            prompt, while lower values allow for more freedom in generation. Higher values may lead to saturation and
+            deterioration of image quality.
+        seg_guidance_scale (`float`, defaults to `3.0`):
+            The scale parameter for smoothed energy guidance. Anatomy and structure coherence may improve with higher
+            values, but it may also lead to overexposure and saturation.
+        seg_blur_sigma (`float`, defaults to `9999999.0`):
+            The amount by which we blur the attention weights. Setting this value greater than 9999.0 results in
+            infinite blur, which means uniform queries. Controlling it exponentially is empirically effective.
+        seg_blur_threshold_inf (`float`, defaults to `9999.0`):
+            The threshold above which the blur is considered infinite.
+        seg_guidance_start (`float`, defaults to `0.0`):
+            The fraction of the total number of denoising steps after which smoothed energy guidance starts.
+        seg_guidance_stop (`float`, defaults to `1.0`):
+            The fraction of the total number of denoising steps after which smoothed energy guidance stops.
+        seg_guidance_layers (`int` or `List[int]`, *optional*):
+            The layer indices to apply smoothed energy guidance to. Can be a single integer or a list of integers. If not
+            provided, `seg_guidance_config` must be provided. The recommended values are `[7, 8, 9]` for Stable Diffusion
+            3.5 Medium.
+        seg_guidance_config (`SmoothedEnergyGuidanceConfig` or `List[SmoothedEnergyGuidanceConfig]`, *optional*):
+            The configuration for the smoothed energy layer guidance. Can be a single `SmoothedEnergyGuidanceConfig` or a list of
+            `SmoothedEnergyGuidanceConfig`. If not provided, `seg_guidance_layers` must be provided.
+        guidance_rescale (`float`, defaults to `0.0`):
+            The rescale factor applied to the noise predictions. This is used to improve image quality and fix
+            overexposure. Based on Section 3.4 from [Common Diffusion Noise Schedules and Sample Steps are
+            Flawed](https://huggingface.co/papers/2305.08891).
+        use_original_formulation (`bool`, defaults to `False`):
+            Whether to use the original formulation of classifier-free guidance as proposed in the paper. By default,
+            we use the diffusers-native implementation that has been in the codebase for a long time. See
+            [~guiders.classifier_free_guidance.ClassifierFreeGuidance] for more details.
+        start (`float`, defaults to `0.01`):
+            The fraction of the total number of denoising steps after which guidance starts.
+        stop (`float`, defaults to `0.2`):
+            The fraction of the total number of denoising steps after which guidance stops.
+    """
+
+    _input_predictions = ["pred_cond", "pred_uncond", "pred_cond_seg"]
+
+    def __init__(
+        self,
+        guidance_scale: float = 7.5,
+        seg_guidance_scale: float = 2.8,
+        seg_blur_sigma: float = 9999999.0,
+        seg_blur_threshold_inf: float = 9999.0,
+        seg_guidance_start: float = 0.0,
+        seg_guidance_stop: float = 1.0,
+        seg_guidance_layers: Optional[Union[int, List[int]]] = None,
+        seg_guidance_config: Union[SmoothedEnergyGuidanceConfig, List[SmoothedEnergyGuidanceConfig]] = None,
+        guidance_rescale: float = 0.0,
+        use_original_formulation: bool = False,
+        start: float = 0.0,
+        stop: float = 1.0,
+    ):
+        super().__init__(start, stop)
+
+        self.guidance_scale = guidance_scale
+        self.seg_guidance_scale = seg_guidance_scale
+        self.seg_blur_sigma = seg_blur_sigma
+        self.seg_blur_threshold_inf = seg_blur_threshold_inf
+        self.seg_guidance_start = seg_guidance_start
+        self.seg_guidance_stop = seg_guidance_stop
+        self.guidance_rescale = guidance_rescale
+        self.use_original_formulation = use_original_formulation
+
+        if not (0.0 <= seg_guidance_start < 1.0):
+            raise ValueError(
+                f"Expected `seg_guidance_start` to be between 0.0 and 1.0, but got {seg_guidance_start}."
+            )
+        if not (seg_guidance_start <= seg_guidance_stop <= 1.0):
+            raise ValueError(
+                f"Expected `seg_guidance_stop` to be between 0.0 and 1.0, but got {seg_guidance_stop}."
+            )
+
+        if seg_guidance_layers is None and seg_guidance_config is None:
+            raise ValueError(
+                "Either `seg_guidance_layers` or `seg_guidance_config` must be provided to enable Smoothed Energy Guidance."
+            )
+        if seg_guidance_layers is not None and seg_guidance_config is not None:
+            raise ValueError("Only one of `seg_guidance_layers` or `seg_guidance_config` can be provided.")
+
+        if seg_guidance_layers is not None:
+            if isinstance(seg_guidance_layers, int):
+                seg_guidance_layers = [seg_guidance_layers]
+            if not isinstance(seg_guidance_layers, list):
+                raise ValueError(
+                    f"Expected `seg_guidance_layers` to be an int or a list of ints, but got {type(seg_guidance_layers)}."
+                )
+            seg_guidance_config = [SmoothedEnergyGuidanceConfig(layer, fqn="auto") for layer in seg_guidance_layers]
+
+        if isinstance(seg_guidance_config, SmoothedEnergyGuidanceConfig):
+            seg_guidance_config = [seg_guidance_config]
+
+        if not isinstance(seg_guidance_config, list):
+            raise ValueError(
+                f"Expected `seg_guidance_config` to be a SmoothedEnergyGuidanceConfig or a list of SmoothedEnergyGuidanceConfig, but got {type(seg_guidance_config)}."
+            )
+
+        self.seg_guidance_config = seg_guidance_config
+        self._seg_layer_hook_names = [f"SmoothedEnergyGuidance_{i}" for i in range(len(self.seg_guidance_config))]
+
+    def prepare_models(self, denoiser: torch.nn.Module) -> None:
+        if self._is_seg_enabled() and self.is_conditional and self._count_prepared > 1:
+            for name, config in zip(self._seg_layer_hook_names, self.seg_guidance_config):
+                _apply_smoothed_energy_guidance_hook(denoiser, config, self.seg_blur_sigma, name=name)
+    
+    def cleanup_models(self, denoiser: torch.nn.Module):
+        if self._is_seg_enabled() and self.is_conditional and self._count_prepared > 1:
+            registry = HookRegistry.check_if_exists_or_initialize(denoiser)
+            # Remove the hooks after inference
+            for hook_name in self._seg_layer_hook_names:
+                registry.remove_hook(hook_name, recurse=True)
+    
+    def prepare_inputs(self, data: "BlockState", input_fields: Optional[Dict[str, Union[str, Tuple[str, str]]]] = None) -> List["BlockState"]:
+        
+        if input_fields is None:
+            input_fields = self._input_fields
+        
+        if self.num_conditions == 1:
+            tuple_indices = [0]
+            input_predictions = ["pred_cond"]
+        elif self.num_conditions == 2:
+            tuple_indices = [0, 1]
+            input_predictions = ["pred_cond", "pred_uncond"] if self._is_cfg_enabled() else ["pred_cond", "pred_cond_seg"]
+        else:
+            tuple_indices = [0, 1, 0]
+            input_predictions = ["pred_cond", "pred_uncond", "pred_cond_seg"]
+        data_batches = []
+        for i in range(self.num_conditions):
+            data_batch = self._prepare_batch(input_fields, data, tuple_indices[i], input_predictions[i])
+            data_batches.append(data_batch)
+        return data_batches
+
+    def forward(
+        self,
+        pred_cond: torch.Tensor,
+        pred_uncond: Optional[torch.Tensor] = None,
+        pred_cond_seg: Optional[torch.Tensor] = None,
+    ) -> torch.Tensor:
+        pred = None
+
+        if not self._is_cfg_enabled() and not self._is_seg_enabled():
+            pred = pred_cond
+        elif not self._is_cfg_enabled():
+            shift = pred_cond - pred_cond_seg
+            pred = pred_cond if self.use_original_formulation else pred_cond_seg
+            pred = pred + self.seg_guidance_scale * shift
+        elif not self._is_seg_enabled():
+            shift = pred_cond - pred_uncond
+            pred = pred_cond if self.use_original_formulation else pred_uncond
+            pred = pred + self.guidance_scale * shift
+        else:
+            shift = pred_cond - pred_uncond
+            shift_seg = pred_cond - pred_cond_seg
+            pred = pred_cond if self.use_original_formulation else pred_uncond
+            pred = pred + self.guidance_scale * shift + self.seg_guidance_scale * shift_seg
+
+        if self.guidance_rescale > 0.0:
+            pred = rescale_noise_cfg(pred, pred_cond, self.guidance_rescale)
+
+        return pred, {}
+    
+    @property
+    def is_conditional(self) -> bool:
+        return self._count_prepared == 1 or self._count_prepared == 3
+
+    @property
+    def num_conditions(self) -> int:
+        num_conditions = 1
+        if self._is_cfg_enabled():
+            num_conditions += 1
+        if self._is_seg_enabled():
+            num_conditions += 1
+        return num_conditions
+
+    def _is_cfg_enabled(self) -> bool:
+        if not self._enabled:
+            return False
+        
+        is_within_range = True
+        if self._num_inference_steps is not None:
+            skip_start_step = int(self._start * self._num_inference_steps)
+            skip_stop_step = int(self._stop * self._num_inference_steps)
+            is_within_range = skip_start_step <= self._step < skip_stop_step
+        
+        is_close = False
+        if self.use_original_formulation:
+            is_close = math.isclose(self.guidance_scale, 0.0)
+        else:
+            is_close = math.isclose(self.guidance_scale, 1.0)
+        
+        return is_within_range and not is_close
+
+    def _is_seg_enabled(self) -> bool:
+        if not self._enabled:
+            return False
+        
+        is_within_range = True
+        if self._num_inference_steps is not None:
+            skip_start_step = int(self.seg_guidance_start * self._num_inference_steps)
+            skip_stop_step = int(self.seg_guidance_stop * self._num_inference_steps)
+            is_within_range = skip_start_step < self._step < skip_stop_step
+        
+        is_zero = math.isclose(self.seg_guidance_scale, 0.0)
+        
+        return is_within_range and not is_zero

src/diffusers/guiders/tangential_classifier_free_guidance.py

@@ -0,0 +1,137 @@
+# Copyright 2024 The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import math
+from typing import Optional, List, TYPE_CHECKING, Dict, Union, Tuple
+
+import torch
+
+from .guider_utils import BaseGuidance, rescale_noise_cfg
+
+if TYPE_CHECKING:
+    from ..modular_pipelines.modular_pipeline import BlockState
+
+
+class TangentialClassifierFreeGuidance(BaseGuidance):
+    """
+    Tangential Classifier Free Guidance (TCFG): https://huggingface.co/papers/2503.18137
+    
+    Args:
+        guidance_scale (`float`, defaults to `7.5`):
+            The scale parameter for classifier-free guidance. Higher values result in stronger conditioning on the text
+            prompt, while lower values allow for more freedom in generation. Higher values may lead to saturation and
+            deterioration of image quality.
+        guidance_rescale (`float`, defaults to `0.0`):
+            The rescale factor applied to the noise predictions. This is used to improve image quality and fix
+            overexposure. Based on Section 3.4 from [Common Diffusion Noise Schedules and Sample Steps are
+            Flawed](https://huggingface.co/papers/2305.08891).
+        use_original_formulation (`bool`, defaults to `False`):
+            Whether to use the original formulation of classifier-free guidance as proposed in the paper. By default,
+            we use the diffusers-native implementation that has been in the codebase for a long time. See
+            [~guiders.classifier_free_guidance.ClassifierFreeGuidance] for more details.
+        start (`float`, defaults to `0.0`):
+            The fraction of the total number of denoising steps after which guidance starts.
+        stop (`float`, defaults to `1.0`):
+            The fraction of the total number of denoising steps after which guidance stops.
+    """
+
+    _input_predictions = ["pred_cond", "pred_uncond"]
+
+    def __init__(
+        self,
+        guidance_scale: float = 7.5,
+        guidance_rescale: float = 0.0,
+        use_original_formulation: bool = False,
+        start: float = 0.0,
+        stop: float = 1.0,
+    ):
+        super().__init__(start, stop)
+
+        self.guidance_scale = guidance_scale
+        self.guidance_rescale = guidance_rescale
+        self.use_original_formulation = use_original_formulation
+
+    def prepare_inputs(self, data: "BlockState", input_fields: Optional[Dict[str, Union[str, Tuple[str, str]]]] = None) -> List["BlockState"]:
+        
+        if input_fields is None:
+            input_fields = self._input_fields
+        
+        tuple_indices = [0] if self.num_conditions == 1 else [0, 1]
+        data_batches = []
+        for i in range(self.num_conditions):
+            data_batch = self._prepare_batch(input_fields, data, tuple_indices[i], self._input_predictions[i])
+            data_batches.append(data_batch)
+        return data_batches
+
+    def forward(self, pred_cond: torch.Tensor, pred_uncond: Optional[torch.Tensor] = None) -> torch.Tensor:
+        pred = None
+
+        if not self._is_tcfg_enabled():
+            pred = pred_cond
+        else:
+            pred = normalized_guidance(pred_cond, pred_uncond, self.guidance_scale, self.use_original_formulation)
+
+        if self.guidance_rescale > 0.0:
+            pred = rescale_noise_cfg(pred, pred_cond, self.guidance_rescale)
+
+        return pred, {}
+
+    @property
+    def is_conditional(self) -> bool:
+        return self._num_outputs_prepared == 1
+
+    @property
+    def num_conditions(self) -> int:
+        num_conditions = 1
+        if self._is_tcfg_enabled():
+            num_conditions += 1
+        return num_conditions
+
+    def _is_tcfg_enabled(self) -> bool:
+        if not self._enabled:
+            return False
+        
+        is_within_range = True
+        if self._num_inference_steps is not None:
+            skip_start_step = int(self._start * self._num_inference_steps)
+            skip_stop_step = int(self._stop * self._num_inference_steps)
+            is_within_range = skip_start_step <= self._step < skip_stop_step
+        
+        is_close = False
+        if self.use_original_formulation:
+            is_close = math.isclose(self.guidance_scale, 0.0)
+        else:
+            is_close = math.isclose(self.guidance_scale, 1.0)
+        
+        return is_within_range and not is_close
+
+
+def normalized_guidance(pred_cond: torch.Tensor, pred_uncond: torch.Tensor, guidance_scale: float, use_original_formulation: bool = False) -> torch.Tensor:
+    cond_dtype = pred_cond.dtype    
+    preds = torch.stack([pred_cond, pred_uncond], dim=1).float()
+    preds = preds.flatten(2)
+    U, S, Vh = torch.linalg.svd(preds, full_matrices=False)
+    Vh_modified = Vh.clone()
+    Vh_modified[:, 1] = 0
+
+    uncond_flat = pred_uncond.reshape(pred_uncond.size(0), 1, -1).float()
+    x_Vh = torch.matmul(uncond_flat, Vh.transpose(-2, -1))
+    x_Vh_V = torch.matmul(x_Vh, Vh_modified)
+    pred_uncond = x_Vh_V.reshape(pred_uncond.shape).to(cond_dtype)
+    
+    pred = pred_cond if use_original_formulation else pred_uncond
+    shift = pred_cond - pred_uncond
+    pred = pred + guidance_scale * shift
+    
+    return pred

src/diffusers/hooks/__init__.py

@@ -5,5 +5,7 @@ if is_torch_available():
     from .faster_cache import FasterCacheConfig, apply_faster_cache
     from .group_offloading import apply_group_offloading
     from .hooks import HookRegistry, ModelHook
+    from .layer_skip import LayerSkipConfig, apply_layer_skip
     from .layerwise_casting import apply_layerwise_casting, apply_layerwise_casting_hook
     from .pyramid_attention_broadcast import PyramidAttentionBroadcastConfig, apply_pyramid_attention_broadcast
+    from .smoothed_energy_guidance_utils import SmoothedEnergyGuidanceConfig

src/diffusers/hooks/_common.py

@@ -0,0 +1,43 @@
+# Copyright 2024 The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from typing import Optional
+
+import torch
+
+from ..models.attention import FeedForward, LuminaFeedForward
+from ..models.attention_processor import Attention, MochiAttention
+
+
+_ATTENTION_CLASSES = (Attention, MochiAttention)
+_FEEDFORWARD_CLASSES = (FeedForward, LuminaFeedForward)
+
+_SPATIAL_TRANSFORMER_BLOCK_IDENTIFIERS = ("blocks", "transformer_blocks", "single_transformer_blocks", "layers")
+_TEMPORAL_TRANSFORMER_BLOCK_IDENTIFIERS = ("temporal_transformer_blocks",)
+_CROSS_TRANSFORMER_BLOCK_IDENTIFIERS = ("blocks", "transformer_blocks", "layers")
+
+_ALL_TRANSFORMER_BLOCK_IDENTIFIERS = tuple(
+    {
+        *_SPATIAL_TRANSFORMER_BLOCK_IDENTIFIERS,
+        *_TEMPORAL_TRANSFORMER_BLOCK_IDENTIFIERS,
+        *_CROSS_TRANSFORMER_BLOCK_IDENTIFIERS,
+    }
+)
+
+
+def _get_submodule_from_fqn(module: torch.nn.Module, fqn: str) -> Optional[torch.nn.Module]:
+    for submodule_name, submodule in module.named_modules():
+        if submodule_name == fqn:
+            return submodule
+    return None

src/diffusers/hooks/_helpers.py

@@ -0,0 +1,271 @@
+# Copyright 2024 The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from dataclasses import dataclass
+from typing import Any, Callable, Type
+
+from ..models.attention import BasicTransformerBlock
+from ..models.attention_processor import AttnProcessor2_0
+from ..models.transformers.cogvideox_transformer_3d import CogVideoXBlock
+from ..models.transformers.transformer_cogview4 import CogView4AttnProcessor, CogView4TransformerBlock
+from ..models.transformers.transformer_flux import FluxSingleTransformerBlock, FluxTransformerBlock
+from ..models.transformers.transformer_hunyuan_video import (
+    HunyuanVideoSingleTransformerBlock,
+    HunyuanVideoTokenReplaceSingleTransformerBlock,
+    HunyuanVideoTokenReplaceTransformerBlock,
+    HunyuanVideoTransformerBlock,
+)
+from ..models.transformers.transformer_ltx import LTXVideoTransformerBlock
+from ..models.transformers.transformer_mochi import MochiTransformerBlock
+from ..models.transformers.transformer_wan import WanTransformerBlock
+
+
+@dataclass
+class AttentionProcessorMetadata:
+    skip_processor_output_fn: Callable[[Any], Any]
+
+
+@dataclass
+class TransformerBlockMetadata:
+    skip_block_output_fn: Callable[[Any], Any]
+    return_hidden_states_index: int = None
+    return_encoder_hidden_states_index: int = None
+
+
+class AttentionProcessorRegistry:
+    _registry = {}
+
+    @classmethod
+    def register(cls, model_class: Type, metadata: AttentionProcessorMetadata):
+        cls._registry[model_class] = metadata
+
+    @classmethod
+    def get(cls, model_class: Type) -> AttentionProcessorMetadata:
+        if model_class not in cls._registry:
+            raise ValueError(f"Model class {model_class} not registered.")
+        return cls._registry[model_class]
+
+
+class TransformerBlockRegistry:
+    _registry = {}
+
+    @classmethod
+    def register(cls, model_class: Type, metadata: TransformerBlockMetadata):
+        cls._registry[model_class] = metadata
+
+    @classmethod
+    def get(cls, model_class: Type) -> TransformerBlockMetadata:
+        if model_class not in cls._registry:
+            raise ValueError(f"Model class {model_class} not registered.")
+        return cls._registry[model_class]
+
+
+def _register_attention_processors_metadata():
+    # AttnProcessor2_0
+    AttentionProcessorRegistry.register(
+        model_class=AttnProcessor2_0,
+        metadata=AttentionProcessorMetadata(
+            skip_processor_output_fn=_skip_proc_output_fn_Attention_AttnProcessor2_0,
+        ),
+    )
+
+    # CogView4AttnProcessor
+    AttentionProcessorRegistry.register(
+        model_class=CogView4AttnProcessor,
+        metadata=AttentionProcessorMetadata(
+            skip_processor_output_fn=_skip_proc_output_fn_Attention_CogView4AttnProcessor,
+        ),
+    )
+
+
+def _register_transformer_blocks_metadata():
+    # BasicTransformerBlock
+    TransformerBlockRegistry.register(
+        model_class=BasicTransformerBlock,
+        metadata=TransformerBlockMetadata(
+            skip_block_output_fn=_skip_block_output_fn_BasicTransformerBlock,
+            return_hidden_states_index=0,
+            return_encoder_hidden_states_index=None,
+        ),
+    )
+
+    # CogVideoX
+    TransformerBlockRegistry.register(
+        model_class=CogVideoXBlock,
+        metadata=TransformerBlockMetadata(
+            skip_block_output_fn=_skip_block_output_fn_CogVideoXBlock,
+            return_hidden_states_index=0,
+            return_encoder_hidden_states_index=1,
+        ),
+    )
+
+    # CogView4
+    TransformerBlockRegistry.register(
+        model_class=CogView4TransformerBlock,
+        metadata=TransformerBlockMetadata(
+            skip_block_output_fn=_skip_block_output_fn_CogView4TransformerBlock,
+            return_hidden_states_index=0,
+            return_encoder_hidden_states_index=1,
+        ),
+    )
+
+    # Flux
+    TransformerBlockRegistry.register(
+        model_class=FluxTransformerBlock,
+        metadata=TransformerBlockMetadata(
+            skip_block_output_fn=_skip_block_output_fn_FluxTransformerBlock,
+            return_hidden_states_index=1,
+            return_encoder_hidden_states_index=0,
+        ),
+    )
+    TransformerBlockRegistry.register(
+        model_class=FluxSingleTransformerBlock,
+        metadata=TransformerBlockMetadata(
+            skip_block_output_fn=_skip_block_output_fn_FluxSingleTransformerBlock,
+            return_hidden_states_index=1,
+            return_encoder_hidden_states_index=0,
+        ),
+    )
+
+    # HunyuanVideo
+    TransformerBlockRegistry.register(
+        model_class=HunyuanVideoTransformerBlock,
+        metadata=TransformerBlockMetadata(
+            skip_block_output_fn=_skip_block_output_fn_HunyuanVideoTransformerBlock,
+            return_hidden_states_index=0,
+            return_encoder_hidden_states_index=1,
+        ),
+    )
+    TransformerBlockRegistry.register(
+        model_class=HunyuanVideoSingleTransformerBlock,
+        metadata=TransformerBlockMetadata(
+            skip_block_output_fn=_skip_block_output_fn_HunyuanVideoSingleTransformerBlock,
+            return_hidden_states_index=0,
+            return_encoder_hidden_states_index=1,
+        ),
+    )
+    TransformerBlockRegistry.register(
+        model_class=HunyuanVideoTokenReplaceTransformerBlock,
+        metadata=TransformerBlockMetadata(
+            skip_block_output_fn=_skip_block_output_fn_HunyuanVideoTokenReplaceTransformerBlock,
+            return_hidden_states_index=0,
+            return_encoder_hidden_states_index=1,
+        ),
+    )
+    TransformerBlockRegistry.register(
+        model_class=HunyuanVideoTokenReplaceSingleTransformerBlock,
+        metadata=TransformerBlockMetadata(
+            skip_block_output_fn=_skip_block_output_fn_HunyuanVideoTokenReplaceSingleTransformerBlock,
+            return_hidden_states_index=0,
+            return_encoder_hidden_states_index=1,
+        ),
+    )
+
+    # LTXVideo
+    TransformerBlockRegistry.register(
+        model_class=LTXVideoTransformerBlock,
+        metadata=TransformerBlockMetadata(
+            skip_block_output_fn=_skip_block_output_fn_LTXVideoTransformerBlock,
+            return_hidden_states_index=0,
+            return_encoder_hidden_states_index=None,
+        ),
+    )
+
+    # Mochi
+    TransformerBlockRegistry.register(
+        model_class=MochiTransformerBlock,
+        metadata=TransformerBlockMetadata(
+            skip_block_output_fn=_skip_block_output_fn_MochiTransformerBlock,
+            return_hidden_states_index=0,
+            return_encoder_hidden_states_index=1,
+        ),
+    )
+
+    # Wan
+    TransformerBlockRegistry.register(
+        model_class=WanTransformerBlock,
+        metadata=TransformerBlockMetadata(
+            skip_block_output_fn=_skip_block_output_fn_WanTransformerBlock,
+            return_hidden_states_index=0,
+            return_encoder_hidden_states_index=None,
+        ),
+    )
+
+
+# fmt: off
+def _skip_attention___ret___hidden_states(self, *args, **kwargs):
+    hidden_states = kwargs.get("hidden_states", None)
+    if hidden_states is None and len(args) > 0:
+        hidden_states = args[0]
+    return hidden_states
+
+
+def _skip_attention___ret___hidden_states___encoder_hidden_states(self, *args, **kwargs):
+    hidden_states = kwargs.get("hidden_states", None)
+    encoder_hidden_states = kwargs.get("encoder_hidden_states", None)
+    if hidden_states is None and len(args) > 0:
+        hidden_states = args[0]
+    if encoder_hidden_states is None and len(args) > 1:
+        encoder_hidden_states = args[1]
+    return hidden_states, encoder_hidden_states
+
+
+_skip_proc_output_fn_Attention_AttnProcessor2_0 = _skip_attention___ret___hidden_states
+_skip_proc_output_fn_Attention_CogView4AttnProcessor = _skip_attention___ret___hidden_states___encoder_hidden_states
+
+
+def _skip_block_output_fn___hidden_states_0___ret___hidden_states(self, *args, **kwargs):
+    hidden_states = kwargs.get("hidden_states", None)
+    if hidden_states is None and len(args) > 0:
+        hidden_states = args[0]
+    return hidden_states
+
+
+def _skip_block_output_fn___hidden_states_0___encoder_hidden_states_1___ret___hidden_states___encoder_hidden_states(self, *args, **kwargs):
+    hidden_states = kwargs.get("hidden_states", None)
+    encoder_hidden_states = kwargs.get("encoder_hidden_states", None)
+    if hidden_states is None and len(args) > 0:
+        hidden_states = args[0]
+    if encoder_hidden_states is None and len(args) > 1:
+        encoder_hidden_states = args[1]
+    return hidden_states, encoder_hidden_states
+
+
+def _skip_block_output_fn___hidden_states_0___encoder_hidden_states_1___ret___encoder_hidden_states___hidden_states(self, *args, **kwargs):
+    hidden_states = kwargs.get("hidden_states", None)
+    encoder_hidden_states = kwargs.get("encoder_hidden_states", None)
+    if hidden_states is None and len(args) > 0:
+        hidden_states = args[0]
+    if encoder_hidden_states is None and len(args) > 1:
+        encoder_hidden_states = args[1]
+    return encoder_hidden_states, hidden_states
+
+
+_skip_block_output_fn_BasicTransformerBlock = _skip_block_output_fn___hidden_states_0___ret___hidden_states
+_skip_block_output_fn_CogVideoXBlock = _skip_block_output_fn___hidden_states_0___encoder_hidden_states_1___ret___hidden_states___encoder_hidden_states
+_skip_block_output_fn_CogView4TransformerBlock = _skip_block_output_fn___hidden_states_0___encoder_hidden_states_1___ret___hidden_states___encoder_hidden_states
+_skip_block_output_fn_FluxTransformerBlock = _skip_block_output_fn___hidden_states_0___encoder_hidden_states_1___ret___encoder_hidden_states___hidden_states
+_skip_block_output_fn_FluxSingleTransformerBlock = _skip_block_output_fn___hidden_states_0___encoder_hidden_states_1___ret___encoder_hidden_states___hidden_states
+_skip_block_output_fn_HunyuanVideoTransformerBlock = _skip_block_output_fn___hidden_states_0___encoder_hidden_states_1___ret___hidden_states___encoder_hidden_states
+_skip_block_output_fn_HunyuanVideoSingleTransformerBlock = _skip_block_output_fn___hidden_states_0___encoder_hidden_states_1___ret___hidden_states___encoder_hidden_states
+_skip_block_output_fn_HunyuanVideoTokenReplaceTransformerBlock = _skip_block_output_fn___hidden_states_0___encoder_hidden_states_1___ret___hidden_states___encoder_hidden_states
+_skip_block_output_fn_HunyuanVideoTokenReplaceSingleTransformerBlock = _skip_block_output_fn___hidden_states_0___encoder_hidden_states_1___ret___hidden_states___encoder_hidden_states
+_skip_block_output_fn_LTXVideoTransformerBlock = _skip_block_output_fn___hidden_states_0___ret___hidden_states
+_skip_block_output_fn_MochiTransformerBlock = _skip_block_output_fn___hidden_states_0___encoder_hidden_states_1___ret___hidden_states___encoder_hidden_states
+_skip_block_output_fn_WanTransformerBlock = _skip_block_output_fn___hidden_states_0___ret___hidden_states
+# fmt: on
+
+
+_register_attention_processors_metadata()
+_register_transformer_blocks_metadata()

src/diffusers/hooks/layer_skip.py

@@ -0,0 +1,231 @@
+# Copyright 2024 The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import math
+from dataclasses import dataclass
+from typing import Callable, List, Optional
+
+import torch
+
+from ..utils import get_logger
+from ..utils.torch_utils import unwrap_module
+from ._common import _ALL_TRANSFORMER_BLOCK_IDENTIFIERS, _ATTENTION_CLASSES, _FEEDFORWARD_CLASSES, _get_submodule_from_fqn
+from ._helpers import AttentionProcessorRegistry, TransformerBlockRegistry
+from .hooks import HookRegistry, ModelHook
+
+
+logger = get_logger(__name__)  # pylint: disable=invalid-name
+
+_LAYER_SKIP_HOOK = "layer_skip_hook"
+
+
+# Aryan/YiYi TODO: we need to make guider class a config mixin so I think this is not needed
+# either remove or make it serializable
+@dataclass
+class LayerSkipConfig:
+    r"""
+    Configuration for skipping internal transformer blocks when executing a transformer model.
+    
+    Args:
+        indices (`List[int]`):
+            The indices of the layer to skip. This is typically the first layer in the transformer block.
+        fqn (`str`, defaults to `"auto"`):
+            The fully qualified name identifying the stack of transformer blocks. Typically, this is
+            `transformer_blocks`, `single_transformer_blocks`, `blocks`, `layers`, or `temporal_transformer_blocks`.
+            For automatic detection, set this to `"auto"`.
+            "auto" only works on DiT models. For UNet models, you must provide the correct fqn.
+        skip_attention (`bool`, defaults to `True`):
+            Whether to skip attention blocks.
+        skip_ff (`bool`, defaults to `True`):
+            Whether to skip feed-forward blocks.
+        skip_attention_scores (`bool`, defaults to `False`):
+            Whether to skip attention score computation in the attention blocks. This is equivalent to using `value`
+            projections as the output of scaled dot product attention.
+        dropout (`float`, defaults to `1.0`):
+            The dropout probability for dropping the outputs of the skipped layers. By default, this is set to `1.0`,
+            meaning that the outputs of the skipped layers are completely ignored. If set to `0.0`, the outputs of the
+            skipped layers are fully retained, which is equivalent to not skipping any layers.
+    """
+
+    indices: List[int]
+    fqn: str = "auto"
+    skip_attention: bool = True
+    skip_attention_scores: bool = False
+    skip_ff: bool = True
+    dropout: float = 1.0
+
+    def __post_init__(self):
+        if not (0 <= self.dropout <= 1):
+            raise ValueError(f"Expected `dropout` to be between 0.0 and 1.0, but got {self.dropout}.")
+        if not math.isclose(self.dropout, 1.0) and self.skip_attention_scores:
+            raise ValueError(
+                "Cannot set `skip_attention_scores` to True when `dropout` is not 1.0. Please set `dropout` to 1.0."
+            )
+
+
+class AttentionScoreSkipFunctionMode(torch.overrides.TorchFunctionMode):
+    def __torch_function__(self, func, types, args=(), kwargs=None):
+        if kwargs is None:
+            kwargs = {}
+        if func is torch.nn.functional.scaled_dot_product_attention:
+            value = kwargs.get("value", None)
+            if value is None:
+                value = args[2]
+            return value
+        return func(*args, **kwargs)
+
+
+class AttentionProcessorSkipHook(ModelHook):
+    def __init__(self, skip_processor_output_fn: Callable, skip_attention_scores: bool = False, dropout: float = 1.0):
+        self.skip_processor_output_fn = skip_processor_output_fn
+        self.skip_attention_scores = skip_attention_scores
+        self.dropout = dropout
+
+    def new_forward(self, module: torch.nn.Module, *args, **kwargs):
+        if self.skip_attention_scores:
+            if not math.isclose(self.dropout, 1.0):
+                raise ValueError(
+                    "Cannot set `skip_attention_scores` to True when `dropout` is not 1.0. Please set `dropout` to 1.0."
+                )
+            with AttentionScoreSkipFunctionMode():
+                output = self.fn_ref.original_forward(*args, **kwargs)
+        else:
+            if math.isclose(self.dropout, 1.0):
+                output = self.skip_processor_output_fn(module, *args, **kwargs)
+            else:
+                output = self.fn_ref.original_forward(*args, **kwargs)
+                output = torch.nn.functional.dropout(output, p=self.dropout)
+        return output
+
+
+class FeedForwardSkipHook(ModelHook):
+    def __init__(self, dropout: float):
+        super().__init__()
+        self.dropout = dropout
+
+    def new_forward(self, module: torch.nn.Module, *args, **kwargs):
+        if math.isclose(self.dropout, 1.0):
+            output = kwargs.get("hidden_states", None)
+            if output is None:
+                output = kwargs.get("x", None)
+            if output is None and len(args) > 0:
+                output = args[0]
+        else:
+            output = self.fn_ref.original_forward(*args, **kwargs)
+            output = torch.nn.functional.dropout(output, p=self.dropout)
+        return output
+
+
+class TransformerBlockSkipHook(ModelHook):
+    def __init__(self, dropout: float):
+        super().__init__()
+        self.dropout = dropout
+
+    def initialize_hook(self, module):
+        self._metadata = TransformerBlockRegistry.get(unwrap_module(module).__class__)
+        return module
+
+    def new_forward(self, module: torch.nn.Module, *args, **kwargs):
+        if math.isclose(self.dropout, 1.0):
+            output = self._metadata.skip_block_output_fn(module, *args, **kwargs)
+        else:
+            output = self.fn_ref.original_forward(*args, **kwargs)
+            output = torch.nn.functional.dropout(output, p=self.dropout)
+        return output
+
+def apply_layer_skip(module: torch.nn.Module, config: LayerSkipConfig) -> None:
+    r"""
+    Apply layer skipping to internal layers of a transformer.
+    
+    Args:
+        module (`torch.nn.Module`):
+            The transformer model to which the layer skip hook should be applied.
+        config (`LayerSkipConfig`):
+            The configuration for the layer skip hook.
+    
+    Example:
+    
+    ```python
+    >>> from diffusers import apply_layer_skip_hook, CogVideoXTransformer3DModel, LayerSkipConfig
+    >>> transformer = CogVideoXTransformer3DModel.from_pretrained("THUDM/CogVideoX-5b", torch_dtype=torch.bfloat16)
+    >>> config = LayerSkipConfig(layer_index=[10, 20], fqn="transformer_blocks")
+    >>> apply_layer_skip_hook(transformer, config)
+    ```
+    """
+    _apply_layer_skip_hook(module, config)
+
+
+def _apply_layer_skip_hook(module: torch.nn.Module, config: LayerSkipConfig, name: Optional[str] = None) -> None:
+    name = name or _LAYER_SKIP_HOOK
+
+    if config.skip_attention and config.skip_attention_scores:
+        raise ValueError("Cannot set both `skip_attention` and `skip_attention_scores` to True. Please choose one.")
+    if not math.isclose(config.dropout, 1.0) and config.skip_attention_scores:
+        raise ValueError("Cannot set `skip_attention_scores` to True when `dropout` is not 1.0. Please set `dropout` to 1.0.")
+
+    if config.fqn == "auto":
+        for identifier in _ALL_TRANSFORMER_BLOCK_IDENTIFIERS:
+            if hasattr(module, identifier):
+                config.fqn = identifier
+                break
+        else:
+            raise ValueError(
+                "Could not find a suitable identifier for the transformer blocks automatically. Please provide a valid "
+                "`fqn` (fully qualified name) that identifies a stack of transformer blocks."
+            )
+
+    transformer_blocks = _get_submodule_from_fqn(module, config.fqn)
+    if transformer_blocks is None or not isinstance(transformer_blocks, torch.nn.ModuleList):
+        raise ValueError(
+            f"Could not find {config.fqn} in the provided module, or configured `fqn` (fully qualified name) does not identify "
+            f"a `torch.nn.ModuleList`. Please provide a valid `fqn` that identifies a stack of transformer blocks."
+        )
+    if len(config.indices) == 0:
+        raise ValueError("Layer index list is empty. Please provide a non-empty list of layer indices to skip.")
+
+    blocks_found = False
+    for i, block in enumerate(transformer_blocks):
+        if i not in config.indices:
+            continue
+        
+        blocks_found = True
+        
+        if config.skip_attention and config.skip_ff:
+            logger.debug(f"Applying TransformerBlockSkipHook to '{config.fqn}.{i}'")
+            registry = HookRegistry.check_if_exists_or_initialize(block)
+            hook = TransformerBlockSkipHook(config.dropout)
+            registry.register_hook(hook, name)
+        
+        elif config.skip_attention or config.skip_attention_scores:
+            for submodule_name, submodule in block.named_modules():
+                if isinstance(submodule, _ATTENTION_CLASSES) and not submodule.is_cross_attention:
+                    logger.debug(f"Applying AttentionProcessorSkipHook to '{config.fqn}.{i}.{submodule_name}'")
+                    output_fn = AttentionProcessorRegistry.get(submodule.processor.__class__).skip_processor_output_fn
+                    registry = HookRegistry.check_if_exists_or_initialize(submodule)
+                    hook = AttentionProcessorSkipHook(output_fn, config.skip_attention_scores, config.dropout)
+                    registry.register_hook(hook, name)
+        
+        if config.skip_ff:
+            for submodule_name, submodule in block.named_modules():
+                if isinstance(submodule, _FEEDFORWARD_CLASSES):
+                    logger.debug(f"Applying FeedForwardSkipHook to '{config.fqn}.{i}.{submodule_name}'")
+                    registry = HookRegistry.check_if_exists_or_initialize(submodule)
+                    hook = FeedForwardSkipHook(config.dropout)
+                    registry.register_hook(hook, name)
+
+    if not blocks_found:
+        raise ValueError(
+            f"Could not find any transformer blocks matching the provided indices {config.indices} and "
+            f"fully qualified name '{config.fqn}'. Please check the indices and fqn for correctness."
+        )

src/diffusers/hooks/smoothed_energy_guidance_utils.py

@@ -0,0 +1,158 @@
+# Copyright 2024 The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import math
+from dataclasses import dataclass
+from typing import Any, Dict, List, Optional, Tuple
+
+import torch
+import torch.nn.functional as F
+
+from ..utils import get_logger
+from ._common import _ATTENTION_CLASSES, _get_submodule_from_fqn
+from .hooks import HookRegistry, ModelHook
+
+
+logger = get_logger(__name__)  # pylint: disable=invalid-name
+
+_SMOOTHED_ENERGY_GUIDANCE_HOOK = "smoothed_energy_guidance_hook"
+
+
+@dataclass
+class SmoothedEnergyGuidanceConfig:
+    r"""
+    Configuration for skipping internal transformer blocks when executing a transformer model.
+    
+    Args:
+        indices (`List[int]`):
+            The indices of the layer to skip. This is typically the first layer in the transformer block.
+        fqn (`str`, defaults to `"auto"`):
+            The fully qualified name identifying the stack of transformer blocks. Typically, this is
+            `transformer_blocks`, `single_transformer_blocks`, `blocks`, `layers`, or `temporal_transformer_blocks`.
+            For automatic detection, set this to `"auto"`.
+            "auto" only works on DiT models. For UNet models, you must provide the correct fqn.
+        _query_proj_identifiers (`List[str]`, defaults to `None`):
+            The identifiers for the query projection layers. Typically, these are `to_q`, `query`, or `q_proj`.
+            If `None`, `to_q` is used by default.
+    """
+
+    indices: List[int]
+    fqn: str = "auto"
+    _query_proj_identifiers: List[str] = None
+
+
+class SmoothedEnergyGuidanceHook(ModelHook):
+    def __init__(self, blur_sigma: float = 1.0, blur_threshold_inf: float = 9999.9) -> None:
+        super().__init__()
+        self.blur_sigma = blur_sigma
+        self.blur_threshold_inf = blur_threshold_inf
+
+    def post_forward(self, module: torch.nn.Module, output: torch.Tensor) -> torch.Tensor:
+        # Copied from https://github.com/SusungHong/SEG-SDXL/blob/cf8256d640d5373541cfea3b3b6caf93272cf986/pipeline_seg.py#L172C31-L172C102
+        kernel_size = math.ceil(6 * self.blur_sigma) + 1 - math.ceil(6 * self.blur_sigma) % 2
+        smoothed_output = _gaussian_blur_2d(output, kernel_size, self.blur_sigma, self.blur_threshold_inf)
+        return smoothed_output
+
+
+def _apply_smoothed_energy_guidance_hook(module: torch.nn.Module, config: SmoothedEnergyGuidanceConfig, blur_sigma: float, name: Optional[str] = None) -> None:
+    name = name or _SMOOTHED_ENERGY_GUIDANCE_HOOK
+    
+    if config.fqn == "auto":
+        for identifier in _ALL_TRANSFORMER_BLOCK_IDENTIFIERS:
+            if hasattr(module, identifier):
+                config.fqn = identifier
+                break
+        else:
+            raise ValueError(
+                "Could not find a suitable identifier for the transformer blocks automatically. Please provide a valid "
+                "`fqn` (fully qualified name) that identifies a stack of transformer blocks."
+            )
+    
+    if config._query_proj_identifiers is None:
+        config._query_proj_identifiers = ["to_q"]
+    
+    transformer_blocks = _get_submodule_from_fqn(module, config.fqn)
+    blocks_found = False
+    for i, block in enumerate(transformer_blocks):
+        if i not in config.indices:
+            continue
+        
+        blocks_found = True
+        
+        for submodule_name, submodule in block.named_modules():
+            if not isinstance(submodule, _ATTENTION_CLASSES) or submodule.is_cross_attention:
+                continue
+            for identifier in config._query_proj_identifiers:
+                query_proj = getattr(submodule, identifier, None)
+                if query_proj is None or not isinstance(query_proj, torch.nn.Linear):
+                    continue
+                logger.debug(
+                    f"Registering smoothed energy guidance hook on {config.fqn}.{i}.{submodule_name}.{identifier}"
+                )
+                registry = HookRegistry.check_if_exists_or_initialize(query_proj)
+                hook = SmoothedEnergyGuidanceHook(blur_sigma)
+                registry.register_hook(hook, name)
+    
+    if not blocks_found:
+        raise ValueError(
+            f"Could not find any transformer blocks matching the provided indices {config.indices} and "
+            f"fully qualified name '{config.fqn}'. Please check the indices and fqn for correctness."
+        )
+
+
+# Modified from https://github.com/SusungHong/SEG-SDXL/blob/cf8256d640d5373541cfea3b3b6caf93272cf986/pipeline_seg.py#L71
+def _gaussian_blur_2d(query: torch.Tensor, kernel_size: int, sigma: float, sigma_threshold_inf: float) -> torch.Tensor:
+    """
+    This implementation assumes that the input query is for visual (image/videos) tokens to apply the 2D gaussian
+    blur. However, some models use joint text-visual token attention for which this may not be suitable. Additionally,
+    this implementation also assumes that the visual tokens come from a square image/video. In practice, despite
+    these assumptions, applying the 2D square gaussian blur on the query projections generates reasonable results
+    for Smoothed Energy Guidance.
+
+    SEG is only supported as an experimental prototype feature for now, so the implementation may be modified
+    in the future without warning or guarantee of reproducibility.
+    """
+    assert query.ndim == 3
+    
+    is_inf = sigma > sigma_threshold_inf
+    batch_size, seq_len, embed_dim = query.shape
+
+    seq_len_sqrt = int(math.sqrt(seq_len))
+    num_square_tokens = seq_len_sqrt * seq_len_sqrt
+    query_slice = query[:, :num_square_tokens, :]
+    query_slice = query_slice.permute(0, 2, 1)
+    query_slice = query_slice.reshape(batch_size, embed_dim, seq_len_sqrt, seq_len_sqrt)
+    
+    if is_inf:
+        kernel_size = min(kernel_size, seq_len_sqrt - (seq_len_sqrt % 2 - 1))
+        kernel_size_half = (kernel_size - 1) / 2
+
+        x = torch.linspace(-kernel_size_half, kernel_size_half, steps=kernel_size)
+        pdf = torch.exp(-0.5 * (x / sigma).pow(2))
+        kernel1d = pdf / pdf.sum()
+        kernel1d = kernel1d.to(query)
+        kernel2d = torch.matmul(kernel1d[:, None], kernel1d[None, :])
+        kernel2d = kernel2d.expand(embed_dim, 1, kernel2d.shape[0], kernel2d.shape[1])
+
+        padding = [kernel_size // 2, kernel_size // 2, kernel_size // 2, kernel_size // 2]
+        query_slice = F.pad(query_slice, padding, mode="reflect")
+        query_slice = F.conv2d(query_slice, kernel2d, groups=embed_dim)
+    else:
+        query_slice[:] = query_slice.mean(dim=(-2, -1), keepdim=True)
+
+    query_slice = query_slice.reshape(batch_size, embed_dim, num_square_tokens)
+    query_slice = query_slice.permute(0, 2, 1)
+    query[:, :num_square_tokens, :] = query_slice.clone()
+    
+    return query

src/diffusers/loaders/__init__.py

@@ -84,6 +84,7 @@ if is_torch_available():
             "IPAdapterMixin",
             "FluxIPAdapterMixin",
             "SD3IPAdapterMixin",
+            "ModularIPAdapterMixin",
         ]
 
 _import_structure["peft"] = ["PeftAdapterMixin"]
@@ -102,6 +103,7 @@ if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT:
                 FluxIPAdapterMixin,
                 IPAdapterMixin,
                 SD3IPAdapterMixin,
+                ModularIPAdapterMixin,
             )
             from .lora_pipeline import (
                 AmusedLoraLoaderMixin,

src/diffusers/loaders/ip_adapter.py

@@ -356,6 +356,265 @@ class IPAdapterMixin:
             )
         self.unet.set_attn_processor(attn_procs)
 
+class ModularIPAdapterMixin:
+    """Mixin for handling IP Adapters."""
+
+    @validate_hf_hub_args
+    def load_ip_adapter(
+        self,
+        pretrained_model_name_or_path_or_dict: Union[str, List[str], Dict[str, torch.Tensor]],
+        subfolder: Union[str, List[str]],
+        weight_name: Union[str, List[str]],
+        **kwargs,
+    ):
+        """
+        Parameters:
+            pretrained_model_name_or_path_or_dict (`str` or `List[str]` or `os.PathLike` or `List[os.PathLike]` or `dict` or `List[dict]`):
+                Can be either:
+
+                    - A string, the *model id* (for example `google/ddpm-celebahq-256`) of a pretrained model hosted on
+                      the Hub.
+                    - A path to a *directory* (for example `./my_model_directory`) containing the model weights saved
+                      with [`ModelMixin.save_pretrained`].
+                    - A [torch state
+                      dict](https://pytorch.org/tutorials/beginner/saving_loading_models.html#what-is-a-state-dict).
+            subfolder (`str` or `List[str]`):
+                The subfolder location of a model file within a larger model repository on the Hub or locally. If a
+                list is passed, it should have the same length as `weight_name`.
+            weight_name (`str` or `List[str]`):
+                The name of the weight file to load. If a list is passed, it should have the same length as
+                `subfolder`.
+            cache_dir (`Union[str, os.PathLike]`, *optional*):
+                Path to a directory where a downloaded pretrained model configuration is cached if the standard cache
+                is not used.
+            force_download (`bool`, *optional*, defaults to `False`):
+                Whether or not to force the (re-)download of the model weights and configuration files, overriding the
+                cached versions if they exist.
+
+            proxies (`Dict[str, str]`, *optional*):
+                A dictionary of proxy servers to use by protocol or endpoint, for example, `{'http': 'foo.bar:3128',
+                'http://hostname': 'foo.bar:4012'}`. The proxies are used on each request.
+            local_files_only (`bool`, *optional*, defaults to `False`):
+                Whether to only load local model weights and configuration files or not. If set to `True`, the model
+                won't be downloaded from the Hub.
+            token (`str` or *bool*, *optional*):
+                The token to use as HTTP bearer authorization for remote files. If `True`, the token generated from
+                `diffusers-cli login` (stored in `~/.huggingface`) is used.
+            revision (`str`, *optional*, defaults to `"main"`):
+                The specific model version to use. It can be a branch name, a tag name, a commit id, or any identifier
+                allowed by Git.
+            low_cpu_mem_usage (`bool`, *optional*, defaults to `True` if torch version >= 1.9.0 else `False`):
+                Speed up model loading only loading the pretrained weights and not initializing the weights. This also
+                tries to not use more than 1x model size in CPU memory (including peak memory) while loading the model.
+                Only supported for PyTorch >= 1.9.0. If you are using an older version of PyTorch, setting this
+                argument to `True` will raise an error.
+        """
+
+        # handle the list inputs for multiple IP Adapters
+        if not isinstance(weight_name, list):
+            weight_name = [weight_name]
+
+        if not isinstance(pretrained_model_name_or_path_or_dict, list):
+            pretrained_model_name_or_path_or_dict = [pretrained_model_name_or_path_or_dict]
+        if len(pretrained_model_name_or_path_or_dict) == 1:
+            pretrained_model_name_or_path_or_dict = pretrained_model_name_or_path_or_dict * len(weight_name)
+
+        if not isinstance(subfolder, list):
+            subfolder = [subfolder]
+        if len(subfolder) == 1:
+            subfolder = subfolder * len(weight_name)
+
+        if len(weight_name) != len(pretrained_model_name_or_path_or_dict):
+            raise ValueError("`weight_name` and `pretrained_model_name_or_path_or_dict` must have the same length.")
+
+        if len(weight_name) != len(subfolder):
+            raise ValueError("`weight_name` and `subfolder` must have the same length.")
+
+        # Load the main state dict first.
+        cache_dir = kwargs.pop("cache_dir", None)
+        force_download = kwargs.pop("force_download", False)
+        proxies = kwargs.pop("proxies", None)
+        local_files_only = kwargs.pop("local_files_only", None)
+        token = kwargs.pop("token", None)
+        revision = kwargs.pop("revision", None)
+        low_cpu_mem_usage = kwargs.pop("low_cpu_mem_usage", _LOW_CPU_MEM_USAGE_DEFAULT)
+
+        if low_cpu_mem_usage and not is_accelerate_available():
+            low_cpu_mem_usage = False
+            logger.warning(
+                "Cannot initialize model with low cpu memory usage because `accelerate` was not found in the"
+                " environment. Defaulting to `low_cpu_mem_usage=False`. It is strongly recommended to install"
+                " `accelerate` for faster and less memory-intense model loading. You can do so with: \n```\npip"
+                " install accelerate\n```\n."
+            )
+
+        if low_cpu_mem_usage is True and not is_torch_version(">=", "1.9.0"):
+            raise NotImplementedError(
+                "Low memory initialization requires torch >= 1.9.0. Please either update your PyTorch version or set"
+                " `low_cpu_mem_usage=False`."
+            )
+
+        user_agent = {
+            "file_type": "attn_procs_weights",
+            "framework": "pytorch",
+        }
+        state_dicts = []
+        for pretrained_model_name_or_path_or_dict, weight_name, subfolder in zip(
+            pretrained_model_name_or_path_or_dict, weight_name, subfolder
+        ):
+            if not isinstance(pretrained_model_name_or_path_or_dict, dict):
+                model_file = _get_model_file(
+                    pretrained_model_name_or_path_or_dict,
+                    weights_name=weight_name,
+                    cache_dir=cache_dir,
+                    force_download=force_download,
+                    proxies=proxies,
+                    local_files_only=local_files_only,
+                    token=token,
+                    revision=revision,
+                    subfolder=subfolder,
+                    user_agent=user_agent,
+                )
+                if weight_name.endswith(".safetensors"):
+                    state_dict = {"image_proj": {}, "ip_adapter": {}}
+                    with safe_open(model_file, framework="pt", device="cpu") as f:
+                        for key in f.keys():
+                            if key.startswith("image_proj."):
+                                state_dict["image_proj"][key.replace("image_proj.", "")] = f.get_tensor(key)
+                            elif key.startswith("ip_adapter."):
+                                state_dict["ip_adapter"][key.replace("ip_adapter.", "")] = f.get_tensor(key)
+                else:
+                    state_dict = load_state_dict(model_file)
+            else:
+                state_dict = pretrained_model_name_or_path_or_dict
+
+            keys = list(state_dict.keys())
+            if "image_proj" not in keys and "ip_adapter" not in keys:
+                raise ValueError("Required keys are (`image_proj` and `ip_adapter`) missing from the state dict.")
+
+            state_dicts.append(state_dict)
+
+            # create feature extractor if it has not been registered to the pipeline yet
+            if hasattr(self, "feature_extractor") and getattr(self, "feature_extractor", None) is None:
+                # FaceID IP adapters don't need the image encoder so it's not present, in this case we default to 224
+                default_clip_size = 224
+                clip_image_size = (
+                    self.image_encoder.config.image_size if self.image_encoder is not None else default_clip_size
+                )
+                feature_extractor = CLIPImageProcessor(size=clip_image_size, crop_size=clip_image_size)
+
+        unet_name = getattr(self, "unet_name", "unet")
+        unet = getattr(self, unet_name)
+        unet._load_ip_adapter_weights(state_dicts, low_cpu_mem_usage=low_cpu_mem_usage)
+
+        extra_loras = unet._load_ip_adapter_loras(state_dicts)
+        if extra_loras != {}:
+            if not USE_PEFT_BACKEND:
+                logger.warning("PEFT backend is required to load these weights.")
+            else:
+                # apply the IP Adapter Face ID LoRA weights
+                peft_config = getattr(unet, "peft_config", {})
+                for k, lora in extra_loras.items():
+                    if f"faceid_{k}" not in peft_config:
+                        self.load_lora_weights(lora, adapter_name=f"faceid_{k}")
+                        self.set_adapters([f"faceid_{k}"], adapter_weights=[1.0])
+
+    def set_ip_adapter_scale(self, scale):
+        """
+        Set IP-Adapter scales per-transformer block. Input `scale` could be a single config or a list of configs for
+        granular control over each IP-Adapter behavior. A config can be a float or a dictionary.
+
+        Example:
+
+        ```py
+        # To use original IP-Adapter
+        scale = 1.0
+        pipeline.set_ip_adapter_scale(scale)
+
+        # To use style block only
+        scale = {
+            "up": {"block_0": [0.0, 1.0, 0.0]},
+        }
+        pipeline.set_ip_adapter_scale(scale)
+
+        # To use style+layout blocks
+        scale = {
+            "down": {"block_2": [0.0, 1.0]},
+            "up": {"block_0": [0.0, 1.0, 0.0]},
+        }
+        pipeline.set_ip_adapter_scale(scale)
+
+        # To use style and layout from 2 reference images
+        scales = [{"down": {"block_2": [0.0, 1.0]}}, {"up": {"block_0": [0.0, 1.0, 0.0]}}]
+        pipeline.set_ip_adapter_scale(scales)
+        ```
+        """
+        unet_name = getattr(self, "unet_name", "unet")
+        unet = getattr(self, unet_name)
+        if not isinstance(scale, list):
+            scale = [scale]
+        scale_configs = _maybe_expand_lora_scales(unet, scale, default_scale=0.0)
+
+        for attn_name, attn_processor in unet.attn_processors.items():
+            if isinstance(
+                attn_processor, (IPAdapterAttnProcessor, IPAdapterAttnProcessor2_0, IPAdapterXFormersAttnProcessor)
+            ):
+                if len(scale_configs) != len(attn_processor.scale):
+                    raise ValueError(
+                        f"Cannot assign {len(scale_configs)} scale_configs to "
+                        f"{len(attn_processor.scale)} IP-Adapter."
+                    )
+                elif len(scale_configs) == 1:
+                    scale_configs = scale_configs * len(attn_processor.scale)
+                for i, scale_config in enumerate(scale_configs):
+                    if isinstance(scale_config, dict):
+                        for k, s in scale_config.items():
+                            if attn_name.startswith(k):
+                                attn_processor.scale[i] = s
+                    else:
+                        attn_processor.scale[i] = scale_config
+
+    def unload_ip_adapter(self):
+        """
+        Unloads the IP Adapter weights
+
+        Examples:
+
+        ```python
+        >>> # Assuming `pipeline` is already loaded with the IP Adapter weights.
+        >>> pipeline.unload_ip_adapter()
+        >>> ...
+        ```
+        """
+
+        # remove hidden encoder
+        if self.unet is None:
+            return
+
+        self.unet.encoder_hid_proj = None
+        self.unet.config.encoder_hid_dim_type = None
+
+        # Kolors: restore `encoder_hid_proj` with `text_encoder_hid_proj`
+        if hasattr(self.unet, "text_encoder_hid_proj") and self.unet.text_encoder_hid_proj is not None:
+            self.unet.encoder_hid_proj = self.unet.text_encoder_hid_proj
+            self.unet.text_encoder_hid_proj = None
+            self.unet.config.encoder_hid_dim_type = "text_proj"
+
+        # restore original Unet attention processors layers
+        attn_procs = {}
+        for name, value in self.unet.attn_processors.items():
+            attn_processor_class = (
+                AttnProcessor2_0() if hasattr(F, "scaled_dot_product_attention") else AttnProcessor()
+            )
+            attn_procs[name] = (
+                attn_processor_class
+                if isinstance(
+                    value, (IPAdapterAttnProcessor, IPAdapterAttnProcessor2_0, IPAdapterXFormersAttnProcessor)
+                )
+                else value.__class__()
+            )
+        self.unet.set_attn_processor(attn_procs)
+
 
 class FluxIPAdapterMixin:
     """Mixin for handling Flux IP Adapters."""

src/diffusers/loaders/lora_base.py

@@ -441,7 +441,7 @@ def _func_optionally_disable_offloading(_pipeline):
     is_model_cpu_offload = False
     is_sequential_cpu_offload = False
 
-    if _pipeline is not None and _pipeline.hf_device_map is None:
+    if _pipeline is not None and hasattr(_pipeline, "hf_device_map") and _pipeline.hf_device_map is None:
         for _, component in _pipeline.components.items():
             if isinstance(component, nn.Module) and hasattr(component, "_hf_hook"):
                 if not is_model_cpu_offload:
@@ -491,6 +491,7 @@ class LoraBaseMixin:
             tuple:
                 A tuple indicating if `is_model_cpu_offload` or `is_sequential_cpu_offload` is True.
         """
+
         return _func_optionally_disable_offloading(_pipeline=_pipeline)
 
     @classmethod
@@ -713,8 +714,10 @@ class LoraBaseMixin:
         # Decompose weights into weights for denoiser and text encoders.
         _component_adapter_weights = {}
         for component in self._lora_loadable_modules:
-            model = getattr(self, component)
-
+            model = getattr(self, component, None)
+            if model is None:
+                logger.warning(f"Model {component} not found in pipeline.")
+                continue
             for adapter_name, weights in zip(adapter_names, adapter_weights):
                 if isinstance(weights, dict):
                     component_adapter_weights = weights.pop(component, None)

src/diffusers/loaders/lora_pipeline.py

@@ -636,7 +636,7 @@ class StableDiffusionXLLoraLoaderMixin(LoraBaseMixin):
         # First, ensure that the checkpoint is a compatible one and can be successfully loaded.
         state_dict, network_alphas = self.lora_state_dict(
             pretrained_model_name_or_path_or_dict,
-            unet_config=self.unet.config,
+            unet_config=self.unet.config if hasattr(self, "unet") else None,
             **kwargs,
         )
 
@@ -644,37 +644,40 @@ class StableDiffusionXLLoraLoaderMixin(LoraBaseMixin):
         if not is_correct_format:
             raise ValueError("Invalid LoRA checkpoint.")
 
-        self.load_lora_into_unet(
-            state_dict,
-            network_alphas=network_alphas,
-            unet=self.unet,
-            adapter_name=adapter_name,
-            _pipeline=self,
-            low_cpu_mem_usage=low_cpu_mem_usage,
-            hotswap=hotswap,
-        )
-        self.load_lora_into_text_encoder(
-            state_dict,
-            network_alphas=network_alphas,
-            text_encoder=self.text_encoder,
-            prefix=self.text_encoder_name,
-            lora_scale=self.lora_scale,
-            adapter_name=adapter_name,
-            _pipeline=self,
-            low_cpu_mem_usage=low_cpu_mem_usage,
-            hotswap=hotswap,
-        )
-        self.load_lora_into_text_encoder(
-            state_dict,
-            network_alphas=network_alphas,
-            text_encoder=self.text_encoder_2,
-            prefix=f"{self.text_encoder_name}_2",
-            lora_scale=self.lora_scale,
-            adapter_name=adapter_name,
-            _pipeline=self,
-            low_cpu_mem_usage=low_cpu_mem_usage,
-            hotswap=hotswap,
-        )
+        if hasattr(self, "unet"):
+            self.load_lora_into_unet(
+                state_dict,
+                network_alphas=network_alphas,
+                unet=self.unet,
+                adapter_name=adapter_name,
+                _pipeline=self,
+                low_cpu_mem_usage=low_cpu_mem_usage,
+                hotswap=hotswap,
+            )
+        if hasattr(self, "text_encoder"):
+            self.load_lora_into_text_encoder(
+                state_dict,
+                network_alphas=network_alphas,
+                text_encoder=self.text_encoder,
+                prefix=self.text_encoder_name,
+                lora_scale=self.lora_scale,
+                adapter_name=adapter_name,
+                _pipeline=self,
+                low_cpu_mem_usage=low_cpu_mem_usage,
+                hotswap=hotswap,
+            )
+        if hasattr(self, "text_encoder_2"):
+            self.load_lora_into_text_encoder(
+                state_dict,
+                network_alphas=network_alphas,
+                text_encoder=self.text_encoder_2,
+                prefix=f"{self.text_encoder_name}_2",
+                lora_scale=self.lora_scale,
+                adapter_name=adapter_name,
+                _pipeline=self,
+                low_cpu_mem_usage=low_cpu_mem_usage,
+                hotswap=hotswap,
+            )
 
     @classmethod
     @validate_hf_hub_args

src/diffusers/loaders/unet.py

@@ -408,6 +408,7 @@ class UNet2DConditionLoadersMixin:
             tuple:
                 A tuple indicating if `is_model_cpu_offload` or `is_sequential_cpu_offload` is True.
         """
+
         return _func_optionally_disable_offloading(_pipeline=_pipeline)
 
     def save_attn_procs(

src/diffusers/modular_pipelines/__init__.py

@@ -0,0 +1,82 @@
+from typing import TYPE_CHECKING
+
+from ..utils import (
+    DIFFUSERS_SLOW_IMPORT,
+    OptionalDependencyNotAvailable,
+    _LazyModule,
+    get_objects_from_module,
+    is_torch_available,
+    is_transformers_available,
+)
+
+
+# These modules contain pipelines from multiple libraries/frameworks
+_dummy_objects = {}
+_import_structure = {}
+
+try:
+    if not is_torch_available():
+        raise OptionalDependencyNotAvailable()
+except OptionalDependencyNotAvailable:
+    from ..utils import dummy_pt_objects  # noqa F403
+
+    _dummy_objects.update(get_objects_from_module(dummy_pt_objects))
+else:
+    _import_structure["modular_pipeline"] = [
+        "ModularPipelineMixin",
+        "PipelineBlock",
+        "AutoPipelineBlocks",
+        "SequentialPipelineBlocks",
+        "LoopSequentialPipelineBlocks",
+        "ModularLoader",
+        "PipelineState",
+        "BlockState",
+    ]
+    _import_structure["modular_pipeline_utils"] = [
+        "ComponentSpec",
+        "ConfigSpec",
+        "InputParam",
+        "OutputParam",
+    ]
+    _import_structure["stable_diffusion_xl"] = ["StableDiffusionXLAutoPipeline", "StableDiffusionXLModularLoader"]
+    _import_structure["components_manager"] = ["ComponentsManager"]
+
+if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT:
+    try:
+        if not is_torch_available():
+            raise OptionalDependencyNotAvailable()
+    except OptionalDependencyNotAvailable:
+        from ..utils.dummy_pt_objects import *  # noqa F403
+    else:
+        from .modular_pipeline import (
+            AutoPipelineBlocks,
+            BlockState,
+            LoopSequentialPipelineBlocks,
+            ModularLoader,
+            ModularPipelineMixin,
+            PipelineBlock,
+            PipelineState,
+            SequentialPipelineBlocks,
+        )
+        from .modular_pipeline_utils import (
+            ComponentSpec,
+            ConfigSpec,
+            InputParam,
+            OutputParam,
+        )
+        from .stable_diffusion_xl import (
+            StableDiffusionXLAutoPipeline,
+            StableDiffusionXLModularLoader,
+        )
+        from .components_manager import ComponentsManager
+else:
+    import sys
+
+    sys.modules[__name__] = _LazyModule(
+        __name__,
+        globals()["__file__"],
+        _import_structure,
+        module_spec=__spec__,
+    )
+    for name, value in _dummy_objects.items():
+        setattr(sys.modules[__name__], name, value)

src/diffusers/modular_pipelines/components_manager.py

@@ -0,0 +1,934 @@
+# Copyright 2024 The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from collections import OrderedDict
+from itertools import combinations
+from typing import List, Optional, Union, Dict, Any
+import copy
+
+import torch
+import time
+from dataclasses import dataclass
+
+from ..utils import (
+    is_accelerate_available,
+    logging,
+)
+from ..models.modeling_utils import ModelMixin
+from .modular_pipeline_utils import ComponentSpec
+
+
+import uuid
+
+
+if is_accelerate_available():
+    from accelerate.hooks import ModelHook, add_hook_to_module, remove_hook_from_module
+    from accelerate.state import PartialState
+    from accelerate.utils import send_to_device
+    from accelerate.utils.memory import clear_device_cache
+    from accelerate.utils.modeling import convert_file_size_to_int
+
+logger = logging.get_logger(__name__)  # pylint: disable=invalid-name
+
+
+# YiYi Notes: copied from modeling_utils.py (decide later where to put this)
+def get_memory_footprint(self, return_buffers=True):
+    r"""
+    Get the memory footprint of a model. This will return the memory footprint of the current model in bytes. Useful to
+    benchmark the memory footprint of the current model and design some tests. Solution inspired from the PyTorch
+    discussions: https://discuss.pytorch.org/t/gpu-memory-that-model-uses/56822/2
+
+    Arguments:
+        return_buffers (`bool`, *optional*, defaults to `True`):
+            Whether to return the size of the buffer tensors in the computation of the memory footprint. Buffers are
+            tensors that do not require gradients and not registered as parameters. E.g. mean and std in batch norm
+            layers. Please see: https://discuss.pytorch.org/t/what-pytorch-means-by-buffers/120266/2
+    """
+    mem = sum([param.nelement() * param.element_size() for param in self.parameters()])
+    if return_buffers:
+        mem_bufs = sum([buf.nelement() * buf.element_size() for buf in self.buffers()])
+        mem = mem + mem_bufs
+    return mem
+
+
+class CustomOffloadHook(ModelHook):
+    """
+    A hook that offloads a model on the CPU until its forward pass is called. It ensures the model and its inputs are
+    on the given device. Optionally offloads other models to the CPU before the forward pass is called.
+
+    Args:
+        execution_device(`str`, `int` or `torch.device`, *optional*):
+            The device on which the model should be executed. Will default to the MPS device if it's available, then
+            GPU 0 if there is a GPU, and finally to the CPU.
+    """
+
+    def __init__(
+        self,
+        execution_device: Optional[Union[str, int, torch.device]] = None,
+        other_hooks: Optional[List["UserCustomOffloadHook"]] = None,
+        offload_strategy: Optional["AutoOffloadStrategy"] = None,
+    ):
+        self.execution_device = execution_device if execution_device is not None else PartialState().default_device
+        self.other_hooks = other_hooks
+        self.offload_strategy = offload_strategy
+        self.model_id = None
+
+    def set_strategy(self, offload_strategy: "AutoOffloadStrategy"):
+        self.offload_strategy = offload_strategy
+
+    def add_other_hook(self, hook: "UserCustomOffloadHook"):
+        """
+        Add a hook to the list of hooks to consider for offloading.
+        """
+        if self.other_hooks is None:
+            self.other_hooks = []
+        self.other_hooks.append(hook)
+
+    def init_hook(self, module):
+        return module.to("cpu")
+
+    def pre_forward(self, module, *args, **kwargs):
+        if module.device != self.execution_device:
+            if self.other_hooks is not None:
+                hooks_to_offload = [hook for hook in self.other_hooks if hook.model.device == self.execution_device]
+                # offload all other hooks
+                start_time = time.perf_counter()
+                if self.offload_strategy is not None:
+                    hooks_to_offload = self.offload_strategy(
+                        hooks=hooks_to_offload,
+                        model_id=self.model_id,
+                        model=module,
+                        execution_device=self.execution_device,
+                    )
+                end_time = time.perf_counter()
+                logger.info(
+                    f" time taken to apply offload strategy for {self.model_id}: {(end_time - start_time):.2f} seconds"
+                )
+
+                for hook in hooks_to_offload:
+                    logger.info(
+                        f"moving {self.model_id} to {self.execution_device}, offloading {hook.model_id} to cpu"
+                    )
+                    hook.offload()
+
+                if hooks_to_offload:
+                    clear_device_cache()
+            module.to(self.execution_device)
+        return send_to_device(args, self.execution_device), send_to_device(kwargs, self.execution_device)
+
+
+class UserCustomOffloadHook:
+    """
+    A simple hook grouping a model and a `CustomOffloadHook`, which provides easy APIs for to call the init method of
+    the hook or remove it entirely.
+    """
+
+    def __init__(self, model_id, model, hook):
+        self.model_id = model_id
+        self.model = model
+        self.hook = hook
+
+    def offload(self):
+        self.hook.init_hook(self.model)
+
+    def attach(self):
+        add_hook_to_module(self.model, self.hook)
+        self.hook.model_id = self.model_id
+
+    def remove(self):
+        remove_hook_from_module(self.model)
+        self.hook.model_id = None
+
+    def add_other_hook(self, hook: "UserCustomOffloadHook"):
+        self.hook.add_other_hook(hook)
+
+
+def custom_offload_with_hook(
+    model_id: str,
+    model: torch.nn.Module,
+    execution_device: Union[str, int, torch.device] = None,
+    offload_strategy: Optional["AutoOffloadStrategy"] = None,
+):
+    hook = CustomOffloadHook(execution_device=execution_device, offload_strategy=offload_strategy)
+    user_hook = UserCustomOffloadHook(model_id=model_id, model=model, hook=hook)
+    user_hook.attach()
+    return user_hook
+
+
+class AutoOffloadStrategy:
+    """
+    Offload strategy that should be used with `CustomOffloadHook` to automatically offload models to the CPU based on
+    the available memory on the device.
+    """
+
+    def __init__(self, memory_reserve_margin="3GB"):
+        self.memory_reserve_margin = convert_file_size_to_int(memory_reserve_margin)
+
+    def __call__(self, hooks, model_id, model, execution_device):
+        if len(hooks) == 0:
+            return []
+
+        current_module_size = get_memory_footprint(model)
+
+        mem_on_device = torch.cuda.mem_get_info(execution_device.index)[0]
+        mem_on_device = mem_on_device - self.memory_reserve_margin
+        if current_module_size < mem_on_device:
+            return []
+
+        min_memory_offload = current_module_size - mem_on_device
+        logger.info(f" search for models to offload in order to free up {min_memory_offload / 1024**3:.2f} GB memory")
+
+        # exlucde models that's not currently loaded on the device
+        module_sizes = dict(
+            sorted(
+                {hook.model_id: get_memory_footprint(hook.model) for hook in hooks}.items(),
+                key=lambda x: x[1],
+                reverse=True,
+            )
+        )
+
+        def search_best_candidate(module_sizes, min_memory_offload):
+            """
+            search the optimal combination of models to offload to cpu, given a dictionary of module sizes and a
+            minimum memory offload size. the combination of models should add up to the smallest modulesize that is
+            larger than `min_memory_offload`
+            """
+            model_ids = list(module_sizes.keys())
+            best_candidate = None
+            best_size = float("inf")
+            for r in range(1, len(model_ids) + 1):
+                for candidate_model_ids in combinations(model_ids, r):
+                    candidate_size = sum(
+                        module_sizes[candidate_model_id] for candidate_model_id in candidate_model_ids
+                    )
+                    if candidate_size < min_memory_offload:
+                        continue
+                    else:
+                        if best_candidate is None or candidate_size < best_size:
+                            best_candidate = candidate_model_ids
+                            best_size = candidate_size
+
+            return best_candidate
+
+        best_offload_model_ids = search_best_candidate(module_sizes, min_memory_offload)
+
+        if best_offload_model_ids is None:
+            # if no combination is found, meaning that we cannot meet the memory requirement, offload all models
+            logger.warning("no combination of models to offload to cpu is found, offloading all models")
+            hooks_to_offload = hooks
+        else:
+            hooks_to_offload = [hook for hook in hooks if hook.model_id in best_offload_model_ids]
+
+        return hooks_to_offload
+
+
+
+class ComponentsManager:
+    def __init__(self):
+        self.components = OrderedDict()
+        self.added_time = OrderedDict()  # Store when components were added 
+        self.collections = OrderedDict() # collection_name -> set of component_names
+        self.model_hooks = None
+        self._auto_offload_enabled = False
+
+    
+    def _lookup_ids(self, name=None, collection=None, load_id=None, components: OrderedDict = None):
+        """
+        Lookup component_ids by name, collection, or load_id.
+        """
+        if components is None:
+            components = self.components
+
+        if name:
+            ids_by_name = set() 
+            for component_id, component in components.items():
+                comp_name = self._id_to_name(component_id)
+                if comp_name == name:
+                    ids_by_name.add(component_id)
+        else:
+            ids_by_name = set(components.keys())
+        if collection:
+            ids_by_collection = set()
+            for component_id, component in components.items():
+                if component_id in self.collections[collection]:
+                    ids_by_collection.add(component_id)
+        else:
+            ids_by_collection = set(components.keys())
+        if load_id:
+            ids_by_load_id = set()
+            for name, component in components.items():
+                if hasattr(component, "_diffusers_load_id") and component._diffusers_load_id == load_id:
+                    ids_by_load_id.add(name)
+        else:
+            ids_by_load_id = set(components.keys())
+        
+        ids = ids_by_name.intersection(ids_by_collection).intersection(ids_by_load_id)
+        return ids
+    
+    @staticmethod
+    def _id_to_name(component_id: str):
+        return "_".join(component_id.split("_")[:-1])
+    
+    def add(self, name, component, collection: Optional[str] = None):
+        
+        component_id = f"{name}_{uuid.uuid4()}"
+
+        # check for duplicated components
+        for comp_id, comp in self.components.items():
+            if comp == component:
+                comp_name = self._id_to_name(comp_id)
+                if comp_name == name:
+                    logger.warning(
+                        f"component '{name}' already exists as '{comp_id}'"
+                    )
+                    component_id = comp_id
+                    break
+                else:
+                    logger.warning(
+                        f"Adding component '{name}' as '{component_id}', but it is duplicate of '{comp_id}'"
+                        f"To remove a duplicate, call `components_manager.remove('<component_id>')`."
+                        )
+
+
+        # check for duplicated load_id and warn (we do not delete for you)
+        if hasattr(component, "_diffusers_load_id") and component._diffusers_load_id != "null":
+            components_with_same_load_id = self._lookup_ids(load_id=component._diffusers_load_id)
+            components_with_same_load_id = [id for id in components_with_same_load_id if id != component_id]
+            
+            if components_with_same_load_id:
+                existing = ", ".join(components_with_same_load_id)
+                logger.warning(
+                    f"Adding component '{component_id}', but it has duplicate load_id '{component._diffusers_load_id}' with existing components: {existing}. "
+                    f"To remove a duplicate, call `components_manager.remove('<component_id>')`."
+                )
+
+        # add component to components manager
+        self.components[component_id] = component
+        self.added_time[component_id] = time.time()
+
+        if collection:
+            if collection not in self.collections:
+                self.collections[collection] = set()
+            if not component_id in self.collections[collection]:
+                comp_ids_in_collection = self._lookup_ids(name=name, collection=collection)
+                for comp_id in comp_ids_in_collection:
+                    logger.info(f"Removing existing {name} from collection '{collection}': {comp_id}")
+                    self.remove(comp_id)
+                self.collections[collection].add(component_id)
+                logger.info(f"Added component '{name}' in collection '{collection}': {component_id}")
+        else:
+            logger.info(f"Added component '{name}' as '{component_id}'")
+
+        if self._auto_offload_enabled:
+            self.enable_auto_cpu_offload(self._auto_offload_device)   
+        
+        return component_id
+
+
+    def remove(self, component_id: str = None):
+
+        if component_id not in self.components:
+            logger.warning(f"Component '{component_id}' not found in ComponentsManager")
+            return
+      
+        component = self.components.pop(component_id)
+        self.added_time.pop(component_id)
+
+        for collection in self.collections:
+            if component_id in self.collections[collection]:
+                self.collections[collection].remove(component_id)
+        
+        if self._auto_offload_enabled:
+            self.enable_auto_cpu_offload(self._auto_offload_device)
+        else:
+            if isinstance(component, torch.nn.Module):
+                component.to("cpu")
+            del component
+            import gc
+            gc.collect()
+            if torch.cuda.is_available():
+                torch.cuda.empty_cache()
+
+    def get(self, names: Union[str, List[str]] = None, collection: Optional[str] = None, load_id: Optional[str] = None,
+             as_name_component_tuples: bool = False):
+        """
+        Select components by name with simple pattern matching.
+        
+        Args:
+            names: Component name(s) or pattern(s)
+                Patterns:
+                - "unet" : match any component with base name "unet" (e.g., unet_123abc)
+                - "!unet" : everything except components with base name "unet"
+                - "unet*" : anything with base name starting with "unet"
+                - "!unet*" : anything with base name NOT starting with "unet"
+                - "*unet*" : anything with base name containing "unet"
+                - "!*unet*" : anything with base name NOT containing "unet"
+                - "refiner|vae|unet" : anything with base name exactly matching "refiner", "vae", or "unet"
+                - "!refiner|vae|unet" : anything with base name NOT exactly matching "refiner", "vae", or "unet"
+                - "unet*|vae*" : anything with base name starting with "unet" OR starting with "vae"
+            collection: Optional collection to filter by
+            load_id: Optional load_id to filter by
+            as_name_component_tuples: If True, returns a list of (name, component) tuples using base names
+                                     instead of a dictionary with component IDs as keys
+        
+        Returns:
+            Dictionary mapping component IDs to components, 
+            or list of (base_name, component) tuples if as_name_component_tuples=True
+        """
+        
+        selected_ids = self._lookup_ids(collection=collection, load_id=load_id)
+        components = {k: self.components[k] for k in selected_ids}
+
+        # Helper to extract base name from component_id
+        def get_base_name(component_id):
+            parts = component_id.split('_')
+            # If the last part looks like a UUID, remove it
+            if len(parts) > 1 and len(parts[-1]) >= 8 and '-' in parts[-1]:
+                return '_'.join(parts[:-1])
+            return component_id
+            
+        if names is None:
+            if as_name_component_tuples:
+                return [(get_base_name(comp_id), comp) for comp_id, comp in components.items()]
+            else:
+                return components
+        
+        # Create mapping from component_id to base_name for all components
+        base_names = {comp_id: get_base_name(comp_id) for comp_id in components.keys()}
+        
+        def matches_pattern(component_id, pattern, exact_match=False):
+            """
+            Helper function to check if a component matches a pattern based on its base name.
+            
+            Args:
+                component_id: The component ID to check
+                pattern: The pattern to match against
+                exact_match: If True, only exact matches to base_name are considered
+            """
+            base_name = base_names[component_id]
+            
+            # Exact match with base name
+            if exact_match:
+                return pattern == base_name
+                
+            # Prefix match (ends with *)
+            elif pattern.endswith('*'):
+                prefix = pattern[:-1]
+                return base_name.startswith(prefix)
+                
+            # Contains match (starts with *)
+            elif pattern.startswith('*'):
+                search = pattern[1:-1] if pattern.endswith('*') else pattern[1:]
+                return search in base_name
+                
+            # Exact match (no wildcards)
+            else:
+                return pattern == base_name
+        
+        if isinstance(names, str):
+            # Check if this is a "not" pattern
+            is_not_pattern = names.startswith('!')
+            if is_not_pattern:
+                names = names[1:]  # Remove the ! prefix
+            
+            # Handle OR patterns (containing |)
+            if '|' in names:
+                terms = names.split('|')
+                matches = {}
+                
+                for comp_id, comp in components.items():
+                    # For OR patterns with exact names (no wildcards), we do exact matching on base names
+                    exact_match = all(not (term.startswith('*') or term.endswith('*')) for term in terms)
+                    
+                    # Check if any of the terms match this component
+                    should_include = any(matches_pattern(comp_id, term, exact_match) for term in terms)
+                    
+                    # Flip the decision if this is a NOT pattern
+                    if is_not_pattern:
+                        should_include = not should_include
+                        
+                    if should_include:
+                        matches[comp_id] = comp
+                
+                log_msg = "NOT " if is_not_pattern else ""
+                match_type = "exactly matching" if exact_match else "matching any of patterns"
+                logger.info(f"Getting components {log_msg}{match_type} {terms}: {list(matches.keys())}")
+            
+            # Try exact match with a base name
+            elif any(names == base_name for base_name in base_names.values()):
+                # Find all components with this base name
+                matches = {
+                    comp_id: comp for comp_id, comp in components.items() 
+                    if (base_names[comp_id] == names) != is_not_pattern
+                }
+                
+                if is_not_pattern:
+                    logger.info(f"Getting all components except those with base name '{names}': {list(matches.keys())}")
+                else:
+                    logger.info(f"Getting components with base name '{names}': {list(matches.keys())}")
+            
+            # Prefix match (ends with *)
+            elif names.endswith('*'):
+                prefix = names[:-1]
+                matches = {
+                    comp_id: comp for comp_id, comp in components.items() 
+                    if base_names[comp_id].startswith(prefix) != is_not_pattern
+                }
+                if is_not_pattern:
+                    logger.info(f"Getting components NOT starting with '{prefix}': {list(matches.keys())}")
+                else:
+                    logger.info(f"Getting components starting with '{prefix}': {list(matches.keys())}")
+            
+            # Contains match (starts with *)
+            elif names.startswith('*'):
+                search = names[1:-1] if names.endswith('*') else names[1:]
+                matches = {
+                    comp_id: comp for comp_id, comp in components.items() 
+                    if (search in base_names[comp_id]) != is_not_pattern
+                }
+                if is_not_pattern:
+                    logger.info(f"Getting components NOT containing '{search}': {list(matches.keys())}")
+                else:
+                    logger.info(f"Getting components containing '{search}': {list(matches.keys())}")
+            
+            # Substring match (no wildcards, but not an exact component name)
+            elif any(names in base_name for base_name in base_names.values()):
+                matches = {
+                    comp_id: comp for comp_id, comp in components.items() 
+                    if (names in base_names[comp_id]) != is_not_pattern
+                }
+                if is_not_pattern:
+                    logger.info(f"Getting components NOT containing '{names}': {list(matches.keys())}")
+                else:
+                    logger.info(f"Getting components containing '{names}': {list(matches.keys())}")
+            
+            else:
+                raise ValueError(f"Component or pattern '{names}' not found in ComponentsManager")
+            
+            if not matches:
+                raise ValueError(f"No components found matching pattern '{names}'")
+            
+            if as_name_component_tuples:
+                return [(base_names[comp_id], comp) for comp_id, comp in matches.items()]
+            else:
+                return matches
+        
+        elif isinstance(names, list):
+            results = {}
+            for name in names:
+                result = self.get(name, collection, load_id, as_name_component_tuples=False)
+                results.update(result)
+                
+            if as_name_component_tuples:
+                return [(base_names[comp_id], comp) for comp_id, comp in results.items()]
+            else:
+                return results
+        
+        else:
+            raise ValueError(f"Invalid type for names: {type(names)}")
+
+    def enable_auto_cpu_offload(self, device: Union[str, int, torch.device]="cuda", memory_reserve_margin="3GB"):
+        for name, component in self.components.items():
+            if isinstance(component, torch.nn.Module) and hasattr(component, "_hf_hook"):
+                remove_hook_from_module(component, recurse=True)
+
+        self.disable_auto_cpu_offload()
+        offload_strategy = AutoOffloadStrategy(memory_reserve_margin=memory_reserve_margin)
+        device = torch.device(device)
+        if device.index is None:
+            device = torch.device(f"{device.type}:{0}")
+        all_hooks = []
+        for name, component in self.components.items():
+            if isinstance(component, torch.nn.Module):
+                hook = custom_offload_with_hook(name, component, device, offload_strategy=offload_strategy)
+                all_hooks.append(hook)
+
+        for hook in all_hooks:
+            other_hooks = [h for h in all_hooks if h is not hook]
+            for other_hook in other_hooks:
+                if other_hook.hook.execution_device == hook.hook.execution_device:
+                    hook.add_other_hook(other_hook)
+
+        self.model_hooks = all_hooks
+        self._auto_offload_enabled = True
+        self._auto_offload_device = device
+
+    def disable_auto_cpu_offload(self):
+        if self.model_hooks is None:
+            self._auto_offload_enabled = False
+            return
+
+        for hook in self.model_hooks:
+            hook.offload()
+            hook.remove()
+        if self.model_hooks:
+            clear_device_cache()
+        self.model_hooks = None
+        self._auto_offload_enabled = False
+
+    # YiYi TODO: add quantization info
+    def get_model_info(self, component_id: str, fields: Optional[Union[str, List[str]]] = None) -> Optional[Dict[str, Any]]:
+        """Get comprehensive information about a component.
+        
+        Args:
+            component_id: Name of the component to get info for
+            fields: Optional field(s) to return. Can be a string for single field or list of fields.
+                   If None, returns all fields.
+                   
+        Returns:
+            Dictionary containing requested component metadata.
+            If fields is specified, returns only those fields.
+            If a single field is requested as string, returns just that field's value.
+        """
+        if component_id not in self.components:
+            raise ValueError(f"Component '{component_id}' not found in ComponentsManager")
+
+        component = self.components[component_id]
+        
+        # Build complete info dict first
+        info = {
+            "model_id": component_id,
+            "added_time": self.added_time[component_id],
+            "collection": ", ".join([coll for coll, comps in self.collections.items() if component_id in comps]) or None,
+        }
+        
+        # Additional info for torch.nn.Module components
+        if isinstance(component, torch.nn.Module):
+            # Check for hook information
+            has_hook = hasattr(component, "_hf_hook")
+            execution_device = None
+            if has_hook and hasattr(component._hf_hook, "execution_device"):
+                execution_device = component._hf_hook.execution_device
+                
+            info.update({
+                "class_name": component.__class__.__name__,
+                "size_gb": get_memory_footprint(component) / (1024**3),
+                "adapters": None,  # Default to None
+                "has_hook": has_hook,
+                "execution_device": execution_device,
+            })
+
+            # Get adapters if applicable
+            if hasattr(component, "peft_config"):
+                info["adapters"] = list(component.peft_config.keys())
+
+            # Check for IP-Adapter scales
+            if hasattr(component, "_load_ip_adapter_weights") and hasattr(component, "attn_processors"):
+                processors = copy.deepcopy(component.attn_processors)
+                # First check if any processor is an IP-Adapter
+                processor_types = [v.__class__.__name__ for v in processors.values()]
+                if any("IPAdapter" in ptype for ptype in processor_types):
+                    # Then get scales only from IP-Adapter processors
+                    scales = {
+                        k: v.scale 
+                        for k, v in processors.items() 
+                        if hasattr(v, "scale") and "IPAdapter" in v.__class__.__name__
+                    }
+                    if scales:
+                        info["ip_adapter"] = summarize_dict_by_value_and_parts(scales)
+
+        # If fields specified, filter info
+        if fields is not None:
+            if isinstance(fields, str):
+                # Single field requested, return just that value
+                return {fields: info.get(fields)}
+            else:
+                # List of fields requested, return dict with just those fields
+                return {k: v for k, v in info.items() if k in fields}
+            
+        return info
+
+    def __repr__(self):
+        # Helper to get simple name without UUID
+        def get_simple_name(name):
+            # Extract the base name by splitting on underscore and taking first part
+            # This assumes names are in format "name_uuid"
+            parts = name.split('_')
+            # If we have at least 2 parts and the last part looks like a UUID, remove it
+            if len(parts) > 1 and len(parts[-1]) >= 8 and '-' in parts[-1]:
+                return '_'.join(parts[:-1])
+            return name
+            
+        # Extract load_id if available
+        def get_load_id(component):
+            if hasattr(component, "_diffusers_load_id"):
+                return component._diffusers_load_id
+            return "N/A"
+            
+        # Format device info compactly
+        def format_device(component, info):
+            if not info["has_hook"]:
+                return str(getattr(component, 'device', 'N/A'))
+            else:
+                device = str(getattr(component, 'device', 'N/A'))
+                exec_device = str(info['execution_device'] or 'N/A')
+                return f"{device}({exec_device})"
+                
+        # Get all simple names to calculate width
+        simple_names = [get_simple_name(id) for id in self.components.keys()]
+        
+        # Get max length of load_ids for models
+        load_ids = [
+            get_load_id(component) 
+            for component in self.components.values() 
+            if isinstance(component, torch.nn.Module) and hasattr(component, "_diffusers_load_id")
+        ]
+        max_load_id_len = max([15] + [len(str(lid)) for lid in load_ids]) if load_ids else 15
+        
+        # Get all collections for each component
+        component_collections = {}
+        for name in self.components.keys():
+            component_collections[name] = []
+            for coll, comps in self.collections.items():
+                if name in comps:
+                    component_collections[name].append(coll)
+            if not component_collections[name]:
+                component_collections[name] = ["N/A"]
+        
+        # Find the maximum collection name length
+        all_collections = [coll for colls in component_collections.values() for coll in colls]
+        max_collection_len = max(10, max(len(str(c)) for c in all_collections)) if all_collections else 10
+        
+        col_widths = {
+            "name": max(15, max(len(name) for name in simple_names)),
+            "class": max(25, max(len(component.__class__.__name__) for component in self.components.values())),
+            "device": 15,  # Reduced since using more compact format
+            "dtype": 15,
+            "size": 10,
+            "load_id": max_load_id_len,
+            "collection": max_collection_len
+        }
+
+        # Create the header lines
+        sep_line = "=" * (sum(col_widths.values()) + len(col_widths) * 3 - 1) + "\n"
+        dash_line = "-" * (sum(col_widths.values()) + len(col_widths) * 3 - 1) + "\n"
+
+        output = "Components:\n" + sep_line
+
+        # Separate components into models and others
+        models = {k: v for k, v in self.components.items() if isinstance(v, torch.nn.Module)}
+        others = {k: v for k, v in self.components.items() if not isinstance(v, torch.nn.Module)}
+
+        # Models section
+        if models:
+            output += "Models:\n" + dash_line
+            # Column headers
+            output += f"{'Name':<{col_widths['name']}} | {'Class':<{col_widths['class']}} | "
+            output += f"{'Device':<{col_widths['device']}} | {'Dtype':<{col_widths['dtype']}} | "
+            output += f"{'Size (GB)':<{col_widths['size']}} | {'Load ID':<{col_widths['load_id']}} | Collection\n"
+            output += dash_line
+
+            # Model entries
+            for name, component in models.items():
+                info = self.get_model_info(name)
+                simple_name = get_simple_name(name)
+                device_str = format_device(component, info)
+                dtype = str(component.dtype) if hasattr(component, "dtype") else "N/A"
+                load_id = get_load_id(component)
+                
+                # Print first collection on the main line
+                first_collection = component_collections[name][0] if component_collections[name] else "N/A"
+                
+                output += f"{simple_name:<{col_widths['name']}} | {info['class_name']:<{col_widths['class']}} | "
+                output += f"{device_str:<{col_widths['device']}} | {dtype:<{col_widths['dtype']}} | "
+                output += f"{info['size_gb']:<{col_widths['size']}.2f} | {load_id:<{col_widths['load_id']}} | {first_collection}\n"
+                
+                # Print additional collections on separate lines if they exist
+                for i in range(1, len(component_collections[name])):
+                    collection = component_collections[name][i]
+                    output += f"{'':<{col_widths['name']}} | {'':<{col_widths['class']}} | "
+                    output += f"{'':<{col_widths['device']}} | {'':<{col_widths['dtype']}} | "
+                    output += f"{'':<{col_widths['size']}} | {'':<{col_widths['load_id']}} | {collection}\n"
+            
+            output += dash_line
+
+        # Other components section
+        if others:
+            if models:  # Add extra newline if we had models section
+                output += "\n"
+            output += "Other Components:\n" + dash_line
+            # Column headers for other components
+            output += f"{'Name':<{col_widths['name']}} | {'Class':<{col_widths['class']}} | Collection\n"
+            output += dash_line
+
+            # Other component entries
+            for name, component in others.items():
+                info = self.get_model_info(name)
+                simple_name = get_simple_name(name)
+                
+                # Print first collection on the main line
+                first_collection = component_collections[name][0] if component_collections[name] else "N/A"
+                
+                output += f"{simple_name:<{col_widths['name']}} | {component.__class__.__name__:<{col_widths['class']}} | {first_collection}\n"
+                
+                # Print additional collections on separate lines if they exist
+                for i in range(1, len(component_collections[name])):
+                    collection = component_collections[name][i]
+                    output += f"{'':<{col_widths['name']}} | {'':<{col_widths['class']}} | {collection}\n"
+            
+            output += dash_line
+
+        # Add additional component info
+        output += "\nAdditional Component Info:\n" + "=" * 50 + "\n"
+        for name in self.components:
+            info = self.get_model_info(name)
+            if info is not None and (info.get("adapters") is not None or info.get("ip_adapter")):
+                simple_name = get_simple_name(name)
+                output += f"\n{simple_name}:\n"
+                if info.get("adapters") is not None:
+                    output += f"  Adapters: {info['adapters']}\n"
+                if info.get("ip_adapter"):
+                    output += f"  IP-Adapter: Enabled\n"
+                output += f"  Added Time: {time.strftime('%Y-%m-%d %H:%M:%S', time.localtime(info['added_time']))}\n"
+        
+        return output
+
+    def from_pretrained(self, pretrained_model_name_or_path, prefix: Optional[str] = None, **kwargs):
+        """
+        Load components from a pretrained model and add them to the manager.
+        
+        Args:
+            pretrained_model_name_or_path (str): The path or identifier of the pretrained model
+            prefix (str, optional): Prefix to add to all component names loaded from this model.
+                                  If provided, components will be named as "{prefix}_{component_name}"
+            **kwargs: Additional arguments to pass to DiffusionPipeline.from_pretrained()
+        """
+        subfolder = kwargs.pop("subfolder", None)
+        # YiYi TODO: extend AutoModel to support non-diffusers models
+        if subfolder:
+            from ..models import AutoModel
+            component = AutoModel.from_pretrained(pretrained_model_name_or_path, subfolder=subfolder, **kwargs)
+            component_name = f"{prefix}_{subfolder}" if prefix else subfolder
+            if component_name not in self.components:
+                self.add(component_name, component)
+            else:
+                logger.warning(
+                    f"Component '{component_name}' already exists in ComponentsManager and will not be added. To add it, either:\n"
+                    f"1. remove the existing component with remove('{component_name}')\n"
+                    f"2. Use a different prefix: add_from_pretrained(..., prefix='{prefix}_2')"
+                )
+        else:
+            from ..pipelines.pipeline_utils import DiffusionPipeline
+            pipe = DiffusionPipeline.from_pretrained(pretrained_model_name_or_path, **kwargs)
+            for name, component in pipe.components.items():
+                    
+                if component is None:
+                    continue
+                
+                # Add prefix if specified
+                component_name = f"{prefix}_{name}" if prefix else name
+                
+                if component_name not in self.components:
+                    self.add(component_name, component)
+                else:
+                    logger.warning(
+                        f"Component '{component_name}' already exists in ComponentsManager and will not be added. To add it, either:\n"
+                        f"1. remove the existing component with remove('{component_name}')\n"
+                        f"2. Use a different prefix: add_from_pretrained(..., prefix='{prefix}_2')"
+                    )
+
+    def get_one(self, component_id: Optional[str] = None, name: Optional[str] = None, collection: Optional[str] = None, load_id: Optional[str] = None) -> Any:
+        """
+        Get a single component by name. Raises an error if multiple components match or none are found.
+        
+        Args:
+            name: Component name or pattern
+            collection: Optional collection to filter by
+            load_id: Optional load_id to filter by
+            
+        Returns:
+            A single component
+            
+        Raises:
+            ValueError: If no components match or multiple components match
+        """
+
+        # if component_id is provided, return the component
+        if component_id is not None and (name is not None or collection is not None or load_id is not None):
+            raise ValueError(" if component_id is provided, name, collection, and load_id must be None")
+        elif component_id is not None:
+            if component_id not in self.components:
+                raise ValueError(f"Component '{component_id}' not found in ComponentsManager")
+            return self.components[component_id]
+        
+        results = self.get(name, collection, load_id)
+        
+        if not results:
+            raise ValueError(f"No components found matching '{name}'")
+            
+        if len(results) > 1:
+            raise ValueError(f"Multiple components found matching '{name}': {list(results.keys())}")
+            
+        return next(iter(results.values()))
+
+def summarize_dict_by_value_and_parts(d: Dict[str, Any]) -> Dict[str, Any]:
+    """Summarizes a dictionary by finding common prefixes that share the same value.
+    
+    For a dictionary with dot-separated keys like:
+    {
+        'down_blocks.1.attentions.1.transformer_blocks.0.attn2.processor': [0.6],
+        'down_blocks.1.attentions.1.transformer_blocks.1.attn2.processor': [0.6],
+        'up_blocks.1.attentions.0.transformer_blocks.0.attn2.processor': [0.3],
+    }
+    
+    Returns a dictionary where keys are the shortest common prefixes and values are their shared values:
+    {
+        'down_blocks': [0.6],
+        'up_blocks': [0.3]
+    }
+    """
+    # First group by values - convert lists to tuples to make them hashable
+    value_to_keys = {}
+    for key, value in d.items():
+        value_tuple = tuple(value) if isinstance(value, list) else value
+        if value_tuple not in value_to_keys:
+            value_to_keys[value_tuple] = []
+        value_to_keys[value_tuple].append(key)
+    
+    def find_common_prefix(keys: List[str]) -> str:
+        """Find the shortest common prefix among a list of dot-separated keys."""
+        if not keys:
+            return ""
+        if len(keys) == 1:
+            return keys[0]
+            
+        # Split all keys into parts
+        key_parts = [k.split('.') for k in keys]
+        
+        # Find how many initial parts are common
+        common_length = 0
+        for parts in zip(*key_parts):
+            if len(set(parts)) == 1:  # All parts at this position are the same
+                common_length += 1
+            else:
+                break
+                
+        if common_length == 0:
+            return ""
+            
+        # Return the common prefix
+        return '.'.join(key_parts[0][:common_length])
+
+    # Create summary by finding common prefixes for each value group
+    summary = {}
+    for value_tuple, keys in value_to_keys.items():
+        prefix = find_common_prefix(keys)
+        if prefix:  # Only add if we found a common prefix
+            # Convert tuple back to list if it was originally a list
+            value = list(value_tuple) if isinstance(d[keys[0]], list) else value_tuple
+            summary[prefix] = value
+        else:
+            summary[""] = value  # Use empty string if no common prefix
+            
+    return summary

src/diffusers/modular_pipelines/modular_pipeline_utils.py

@@ -0,0 +1,595 @@
+# Copyright 2023 The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import re
+import inspect
+from dataclasses import dataclass, asdict, field, fields
+from typing import Any, Dict, List, Optional, Tuple, Type, Union, Literal
+
+from ..utils.import_utils import is_torch_available
+from ..configuration_utils import FrozenDict, ConfigMixin
+
+if is_torch_available():
+    import torch
+
+
+# YiYi TODO:
+# 1. validate the dataclass fields
+# 2. add a validator for create_* methods, make sure they are valid inputs to pass to from_pretrained()
+@dataclass
+class ComponentSpec:
+    """Specification for a pipeline component.
+    
+    A component can be created in two ways:
+    1. From scratch using __init__ with a config dict
+    2. using `from_pretrained`
+    
+    Attributes:
+        name: Name of the component
+        type_hint: Type of the component (e.g. UNet2DConditionModel)
+        description: Optional description of the component
+        config: Optional config dict for __init__ creation
+        repo: Optional repo path for from_pretrained creation
+        subfolder: Optional subfolder in repo
+        variant: Optional variant in repo
+        revision: Optional revision in repo
+        default_creation_method: Preferred creation method - "from_config" or "from_pretrained"
+    """
+    name: Optional[str] = None
+    type_hint: Optional[Type] = None
+    description: Optional[str] = None
+    config: Optional[FrozenDict[str, Any]] = None
+    # YiYi Notes: should we change it to pretrained_model_name_or_path for consistency? a bit long for a field name
+    repo: Optional[Union[str, List[str]]] = field(default=None, metadata={"loading": True})
+    subfolder: Optional[str] = field(default=None, metadata={"loading": True})
+    variant: Optional[str] = field(default=None, metadata={"loading": True})
+    revision: Optional[str] = field(default=None, metadata={"loading": True})
+    default_creation_method: Literal["from_config", "from_pretrained"] = "from_pretrained"
+    
+    
+    def __hash__(self):
+        """Make ComponentSpec hashable, using load_id as the hash value."""
+        return hash((self.name, self.load_id, self.default_creation_method))
+    
+    def __eq__(self, other):
+        """Compare ComponentSpec objects based on name and load_id."""
+        if not isinstance(other, ComponentSpec):
+            return False
+        return (self.name == other.name and 
+                self.load_id == other.load_id and 
+                self.default_creation_method == other.default_creation_method)
+    
+    @classmethod
+    def from_component(cls, name: str, component: Any) -> Any:
+        """Create a ComponentSpec from a Component created by `create` or `load` method."""
+        
+        if not hasattr(component, "_diffusers_load_id"):
+            raise ValueError("Component is not created by `create` or `load` method")
+        # throw a error if component is created with `create` method but not a subclass of ConfigMixin
+        # YiYi TODO: remove this check if we remove support for non configmixin in `create()` method
+        if component._diffusers_load_id == "null" and not isinstance(component, ConfigMixin):
+            raise ValueError(
+                "We currently only support creating ComponentSpec from a component with "
+                "created with `ComponentSpec.load` method"
+                "or created with `ComponentSpec.create` and a subclass of ConfigMixin"
+            )
+        
+        type_hint = component.__class__
+        default_creation_method = "from_config" if component._diffusers_load_id == "null" else "from_pretrained"
+        
+        if isinstance(component, ConfigMixin):
+            config = component.config
+        else:
+            config = None
+     
+        load_spec = cls.decode_load_id(component._diffusers_load_id)
+    
+        return cls(name=name, type_hint=type_hint, config=config, default_creation_method=default_creation_method, **load_spec)
+        
+    @classmethod
+    def loading_fields(cls) -> List[str]:
+        """
+        Return the names of all loading‐related fields
+        (i.e. those whose field.metadata["loading"] is True).
+        """
+        return [f.name for f in fields(cls) if f.metadata.get("loading", False)]
+    
+    
+    @property
+    def load_id(self) -> str:
+        """
+        Unique identifier for this spec's pretrained load,
+        composed of repo|subfolder|variant|revision (no empty segments).
+        """
+        parts = [getattr(self, k) for k in self.loading_fields()]
+        parts = ["null" if p is None else p for p in parts]
+        return "|".join(p for p in parts if p)
+    
+    @classmethod
+    def decode_load_id(cls, load_id: str) -> Dict[str, Optional[str]]:
+        """
+        Decode a load_id string back into a dictionary of loading fields and values.
+        
+        Args:
+            load_id: The load_id string to decode, format: "repo|subfolder|variant|revision"
+                     where None values are represented as "null"
+        
+        Returns:
+            Dict mapping loading field names to their values. e.g.
+            {
+                "repo": "path/to/repo",
+                "subfolder": "subfolder",
+                "variant": "variant",
+                "revision": "revision"
+            }
+            If a segment value is "null", it's replaced with None.
+            Returns None if load_id is "null" (indicating component not created with `load` method).
+        """
+            
+        # Get all loading fields in order
+        loading_fields = cls.loading_fields()
+        result = {f: None for f in loading_fields}
+
+        if load_id == "null":
+            return result
+        
+        # Split the load_id
+        parts = load_id.split("|")
+        
+        # Map parts to loading fields by position
+        for i, part in enumerate(parts):
+            if i < len(loading_fields):
+                # Convert "null" string back to None
+                result[loading_fields[i]] = None if part == "null" else part
+        
+        return result
+    
+    
+    # YiYi TODO: I think we should only support ConfigMixin for this method (after we make guider and image_processors config mixin)
+    # otherwise we cannot do spec -> spec.create() -> component -> ComponentSpec.from_component(component)
+    # the config info is lost in the process
+    # remove error check in from_component spec and ModularLoader.update() if we remove support for non configmixin in `create()` method
+    def create(self, config: Optional[Union[FrozenDict, Dict[str, Any]]] = None, **kwargs) -> Any:
+        """Create component using from_config with config."""
+
+        if self.type_hint is None or not isinstance(self.type_hint, type):
+            raise ValueError(
+                f"`type_hint` is required when using from_config creation method."
+            )
+    
+        config = config or self.config or {}
+    
+        if issubclass(self.type_hint, ConfigMixin):
+            component = self.type_hint.from_config(config, **kwargs)
+        else:
+            signature_params = inspect.signature(self.type_hint.__init__).parameters
+            init_kwargs = {}
+            for k, v in config.items():
+                if k in signature_params:
+                    init_kwargs[k] = v
+            for k, v in kwargs.items():
+                if k in signature_params:
+                    init_kwargs[k] = v
+            component = self.type_hint(**init_kwargs)
+        
+        component._diffusers_load_id = "null"
+        if hasattr(component, "config"):
+            self.config = component.config
+        
+        return component
+    
+    # YiYi TODO: add guard for type of model, if it is supported by from_pretrained
+    def load(self, **kwargs) -> Any:
+        """Load component using from_pretrained."""
+        
+        # select loading fields from kwargs passed from user: e.g. repo, subfolder, variant, revision, note the list could change
+        passed_loading_kwargs = {key: kwargs.pop(key) for key in self.loading_fields() if key in kwargs}
+        # merge loading field value in the spec with user passed values to create load_kwargs
+        load_kwargs = {key: passed_loading_kwargs.get(key, getattr(self, key)) for key in self.loading_fields()}
+        # repo is a required argument for from_pretrained, a.k.a. pretrained_model_name_or_path
+        repo = load_kwargs.pop("repo", None)
+        if repo is None:
+            raise ValueError(f"`repo` info is required when using `load` method (you can directly set it in `repo` field of the ComponentSpec or pass it as an argument)")
+        
+        if self.type_hint is None:
+            try:
+                from diffusers import AutoModel
+                component = AutoModel.from_pretrained(repo, **load_kwargs, **kwargs)
+            except Exception as e:
+                raise ValueError(f"Unable to load {self.name} without `type_hint`: {e}")
+            # update type_hint if AutoModel load successfully
+            self.type_hint = component.__class__
+        else:
+            try:
+                component = self.type_hint.from_pretrained(repo, **load_kwargs, **kwargs)
+            except Exception as e:
+                raise ValueError(f"Unable to load {self.name} using load method: {e}")
+        
+        self.repo = repo
+        for k, v in load_kwargs.items():
+            setattr(self, k, v)
+        component._diffusers_load_id = self.load_id
+            
+        return component
+    
+
+
+@dataclass 
+class ConfigSpec:
+    """Specification for a pipeline configuration parameter."""
+    name: str
+    default: Any
+    description: Optional[str] = None
+
+
+# YiYi Notes: both inputs and intermediates_inputs are InputParam objects
+# however some fields are not relevant for intermediates_inputs
+# e.g. unlike inputs, required only used in docstring for intermediate_inputs, we do not check if a required intermediate inputs is passed
+# default is not used for intermediates_inputs, we only use default from inputs, so it is ignored if it is set for intermediates_inputs
+# -> should we use different class for inputs and intermediates_inputs?
+@dataclass
+class InputParam:
+    """Specification for an input parameter."""
+    name: str = None
+    type_hint: Any = None
+    default: Any = None
+    required: bool = False
+    description: str = ""
+    kwargs_type: str = None # YiYi Notes: experimenting with this, not sure if we should keep it
+
+    def __repr__(self):
+        return f"<{self.name}: {'required' if self.required else 'optional'}, default={self.default}>"
+
+
+@dataclass 
+class OutputParam:
+    """Specification for an output parameter."""
+    name: str
+    type_hint: Any = None
+    description: str = ""
+    kwargs_type: str = None
+
+    def __repr__(self):
+        return f"<{self.name}: {self.type_hint.__name__ if hasattr(self.type_hint, '__name__') else str(self.type_hint)}>"
+
+
+def format_inputs_short(inputs):
+    """
+    Format input parameters into a string representation, with required params first followed by optional ones.
+    
+    Args:
+        inputs: List of input parameters with 'required' and 'name' attributes, and 'default' for optional params
+        
+    Returns:
+        str: Formatted string of input parameters
+    
+    Example:
+        >>> inputs = [
+        ...     InputParam(name="prompt", required=True),
+        ...     InputParam(name="image", required=True),
+        ...     InputParam(name="guidance_scale", required=False, default=7.5),
+        ...     InputParam(name="num_inference_steps", required=False, default=50)
+        ... ]
+        >>> format_inputs_short(inputs)
+        'prompt, image, guidance_scale=7.5, num_inference_steps=50'
+    """
+    required_inputs = [param for param in inputs if param.required]
+    optional_inputs = [param for param in inputs if not param.required]
+    
+    required_str = ", ".join(param.name for param in required_inputs)
+    optional_str = ", ".join(f"{param.name}={param.default}" for param in optional_inputs)
+    
+    inputs_str = required_str
+    if optional_str:
+        inputs_str = f"{inputs_str}, {optional_str}" if required_str else optional_str
+        
+    return inputs_str
+
+
+def format_intermediates_short(intermediates_inputs, required_intermediates_inputs, intermediates_outputs):
+    """
+    Formats intermediate inputs and outputs of a block into a string representation.
+    
+    Args:
+        intermediates_inputs: List of intermediate input parameters
+        required_intermediates_inputs: List of required intermediate input names
+        intermediates_outputs: List of intermediate output parameters
+    
+    Returns:
+        str: Formatted string like:
+            Intermediates:
+                - inputs: Required(latents), dtype
+                - modified: latents  # variables that appear in both inputs and outputs
+                - outputs: images    # new outputs only
+    """
+    # Handle inputs
+    input_parts = []
+    for inp in intermediates_inputs:
+        if inp.name in required_intermediates_inputs:
+            input_parts.append(f"Required({inp.name})")
+        else:
+            if inp.name is None and inp.kwargs_type is not None:
+                inp_name = "*_" + inp.kwargs_type
+            else:
+                inp_name = inp.name
+            input_parts.append(inp_name)
+    
+    # Handle modified variables (appear in both inputs and outputs)
+    inputs_set = {inp.name for inp in intermediates_inputs}
+    modified_parts = []
+    new_output_parts = []
+    
+    for out in intermediates_outputs:
+        if out.name in inputs_set:
+            modified_parts.append(out.name)
+        else:
+            new_output_parts.append(out.name)
+    
+    result = []
+    if input_parts:
+        result.append(f"    - inputs: {', '.join(input_parts)}")
+    if modified_parts:
+        result.append(f"    - modified: {', '.join(modified_parts)}")
+    if new_output_parts:
+        result.append(f"    - outputs: {', '.join(new_output_parts)}")
+        
+    return "\n".join(result) if result else "    (none)"
+
+
+def format_params(params, header="Args", indent_level=4, max_line_length=115):
+    """Format a list of InputParam or OutputParam objects into a readable string representation.
+
+    Args:
+        params: List of InputParam or OutputParam objects to format
+        header: Header text to use (e.g. "Args" or "Returns")
+        indent_level: Number of spaces to indent each parameter line (default: 4)
+        max_line_length: Maximum length for each line before wrapping (default: 115)
+
+    Returns:
+        A formatted string representing all parameters
+    """
+    if not params:
+        return ""
+        
+    base_indent = " " * indent_level
+    param_indent = " " * (indent_level + 4)
+    desc_indent = " " * (indent_level + 8)
+    formatted_params = []
+    
+    def get_type_str(type_hint):
+        if hasattr(type_hint, "__origin__") and type_hint.__origin__ is Union:
+            types = [t.__name__ if hasattr(t, "__name__") else str(t) for t in type_hint.__args__]
+            return f"Union[{', '.join(types)}]"
+        return type_hint.__name__ if hasattr(type_hint, "__name__") else str(type_hint)
+    
+    def wrap_text(text, indent, max_length):
+        """Wrap text while preserving markdown links and maintaining indentation."""
+        words = text.split()
+        lines = []
+        current_line = []
+        current_length = 0
+
+        for word in words:
+            word_length = len(word) + (1 if current_line else 0)
+            
+            if current_line and current_length + word_length > max_length:
+                lines.append(" ".join(current_line))
+                current_line = [word]
+                current_length = len(word)
+            else:
+                current_line.append(word)
+                current_length += word_length
+        
+        if current_line:
+            lines.append(" ".join(current_line))
+            
+        return f"\n{indent}".join(lines)
+    
+    # Add the header
+    formatted_params.append(f"{base_indent}{header}:")
+    
+    for param in params:
+        # Format parameter name and type
+        type_str = get_type_str(param.type_hint) if param.type_hint != Any else ""
+        param_str = f"{param_indent}{param.name} (`{type_str}`"
+        
+        # Add optional tag and default value if parameter is an InputParam and optional
+        if hasattr(param, "required"):
+            if not param.required:
+                param_str += ", *optional*"
+                if param.default is not None:
+                    param_str += f", defaults to {param.default}"
+        param_str += "):"
+            
+        # Add description on a new line with additional indentation and wrapping
+        if param.description:
+            desc = re.sub(
+                r'\[(.*?)\]\((https?://[^\s\)]+)\)',
+                r'[\1](\2)',
+                param.description
+            )
+            wrapped_desc = wrap_text(desc, desc_indent, max_line_length)
+            param_str += f"\n{desc_indent}{wrapped_desc}"
+            
+        formatted_params.append(param_str)
+    
+    return "\n\n".join(formatted_params)
+
+
+def format_input_params(input_params, indent_level=4, max_line_length=115):
+    """Format a list of InputParam objects into a readable string representation.
+
+    Args:
+        input_params: List of InputParam objects to format
+        indent_level: Number of spaces to indent each parameter line (default: 4)
+        max_line_length: Maximum length for each line before wrapping (default: 115)
+
+    Returns:
+        A formatted string representing all input parameters
+    """
+    return format_params(input_params, "Inputs", indent_level, max_line_length)
+
+
+def format_output_params(output_params, indent_level=4, max_line_length=115):
+    """Format a list of OutputParam objects into a readable string representation.
+
+    Args:
+        output_params: List of OutputParam objects to format
+        indent_level: Number of spaces to indent each parameter line (default: 4)
+        max_line_length: Maximum length for each line before wrapping (default: 115)
+
+    Returns:
+        A formatted string representing all output parameters
+    """
+    return format_params(output_params, "Outputs", indent_level, max_line_length)
+
+
+def format_components(components, indent_level=4, max_line_length=115, add_empty_lines=True):
+    """Format a list of ComponentSpec objects into a readable string representation.
+
+    Args:
+        components: List of ComponentSpec objects to format
+        indent_level: Number of spaces to indent each component line (default: 4)
+        max_line_length: Maximum length for each line before wrapping (default: 115)
+        add_empty_lines: Whether to add empty lines between components (default: True)
+
+    Returns:
+        A formatted string representing all components
+    """
+    if not components:
+        return ""
+        
+    base_indent = " " * indent_level
+    component_indent = " " * (indent_level + 4)
+    formatted_components = []
+    
+    # Add the header
+    formatted_components.append(f"{base_indent}Components:")
+    if add_empty_lines:
+        formatted_components.append("")
+    
+    # Add each component with optional empty lines between them
+    for i, component in enumerate(components):
+        # Get type name, handling special cases
+        type_name = component.type_hint.__name__ if hasattr(component.type_hint, "__name__") else str(component.type_hint)
+        
+        component_desc = f"{component_indent}{component.name} (`{type_name}`)"
+        if component.description:
+            component_desc += f": {component.description}"
+        
+        # Get the loading fields dynamically
+        loading_field_values = []
+        for field_name in component.loading_fields():
+            field_value = getattr(component, field_name)
+            if field_value is not None:
+                loading_field_values.append(f"{field_name}={field_value}")
+        
+        # Add loading field information if available
+        if loading_field_values:
+            component_desc += f" [{', '.join(loading_field_values)}]"
+            
+        formatted_components.append(component_desc)
+        
+        # Add an empty line after each component except the last one
+        if add_empty_lines and i < len(components) - 1:
+            formatted_components.append("")
+    
+    return "\n".join(formatted_components)
+
+
+def format_configs(configs, indent_level=4, max_line_length=115, add_empty_lines=True):
+    """Format a list of ConfigSpec objects into a readable string representation.
+
+    Args:
+        configs: List of ConfigSpec objects to format
+        indent_level: Number of spaces to indent each config line (default: 4)
+        max_line_length: Maximum length for each line before wrapping (default: 115)
+        add_empty_lines: Whether to add empty lines between configs (default: True)
+
+    Returns:
+        A formatted string representing all configs
+    """
+    if not configs:
+        return ""
+        
+    base_indent = " " * indent_level
+    config_indent = " " * (indent_level + 4)
+    formatted_configs = []
+    
+    # Add the header
+    formatted_configs.append(f"{base_indent}Configs:")
+    if add_empty_lines:
+        formatted_configs.append("")
+    
+    # Add each config with optional empty lines between them
+    for i, config in enumerate(configs):
+        config_desc = f"{config_indent}{config.name} (default: {config.default})"
+        if config.description:
+            config_desc += f": {config.description}"
+        formatted_configs.append(config_desc)
+        
+        # Add an empty line after each config except the last one
+        if add_empty_lines and i < len(configs) - 1:
+            formatted_configs.append("")
+    
+    return "\n".join(formatted_configs)
+
+
+def make_doc_string(inputs, intermediates_inputs, outputs, description="", class_name=None, expected_components=None, expected_configs=None):
+    """
+    Generates a formatted documentation string describing the pipeline block's parameters and structure.
+    
+    Args:
+        inputs: List of input parameters
+        intermediates_inputs: List of intermediate input parameters
+        outputs: List of output parameters
+        description (str, *optional*): Description of the block
+        class_name (str, *optional*): Name of the class to include in the documentation
+        expected_components (List[ComponentSpec], *optional*): List of expected components
+        expected_configs (List[ConfigSpec], *optional*): List of expected configurations
+    
+    Returns:
+        str: A formatted string containing information about components, configs, call parameters, 
+            intermediate inputs/outputs, and final outputs.
+    """
+    output = ""
+
+    # Add class name if provided
+    if class_name:
+        output += f"class {class_name}\n\n"
+
+    # Add description
+    if description:
+        desc_lines = description.strip().split('\n')
+        aligned_desc = '\n'.join('  ' + line for line in desc_lines)
+        output += aligned_desc + "\n\n"
+
+    # Add components section if provided
+    if expected_components and len(expected_components) > 0:
+        components_str = format_components(expected_components, indent_level=2)
+        output += components_str + "\n\n"
+
+    # Add configs section if provided
+    if expected_configs and len(expected_configs) > 0:
+        configs_str = format_configs(expected_configs, indent_level=2)
+        output += configs_str + "\n\n"
+
+    # Add inputs section
+    output += format_input_params(inputs + intermediates_inputs, indent_level=2)
+    
+    # Add outputs section
+    output += "\n\n"
+    output += format_output_params(outputs, indent_level=2)
+
+    return output 

src/diffusers/modular_pipelines/stable_diffusion_xl/__init__.py

@@ -0,0 +1,51 @@
+from typing import TYPE_CHECKING
+
+from ...utils import (
+    DIFFUSERS_SLOW_IMPORT,
+    OptionalDependencyNotAvailable,
+    _LazyModule,
+    get_objects_from_module,
+    is_torch_available,
+    is_transformers_available,
+)
+
+
+_dummy_objects = {}
+_import_structure = {}
+
+try:
+    if not (is_transformers_available() and is_torch_available()):
+        raise OptionalDependencyNotAvailable()
+except OptionalDependencyNotAvailable:
+    from ...utils import dummy_torch_and_transformers_objects  # noqa F403
+
+    _dummy_objects.update(get_objects_from_module(dummy_torch_and_transformers_objects))
+else:
+    _import_structure["modular_pipeline_presets"] = ["StableDiffusionXLAutoPipeline"]
+    _import_structure["modular_loader"] = ["StableDiffusionXLModularLoader"]
+    _import_structure["encoders"] = ["StableDiffusionXLAutoIPAdapterStep", "StableDiffusionXLTextEncoderStep", "StableDiffusionXLAutoVaeEncoderStep"]
+    _import_structure["decoders"] = ["StableDiffusionXLAutoDecodeStep"]
+
+if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT:
+    try:
+        if not (is_transformers_available() and is_torch_available()):
+            raise OptionalDependencyNotAvailable()
+    except OptionalDependencyNotAvailable:
+        from ...utils.dummy_torch_and_transformers_objects import *  # noqa F403
+    else:
+        from .modular_pipeline_presets import StableDiffusionXLAutoPipeline
+        from .modular_loader import StableDiffusionXLModularLoader
+        from .encoders import StableDiffusionXLAutoIPAdapterStep, StableDiffusionXLTextEncoderStep, StableDiffusionXLAutoVaeEncoderStep
+        from .decoders import StableDiffusionXLAutoDecodeStep
+else:
+    import sys
+
+    sys.modules[__name__] = _LazyModule(
+        __name__,
+        globals()["__file__"],
+        _import_structure,
+        module_spec=__spec__,
+    )
+
+    for name, value in _dummy_objects.items():
+        setattr(sys.modules[__name__], name, value)

src/diffusers/modular_pipelines/stable_diffusion_xl/decoders.py

@@ -0,0 +1,215 @@
+# Copyright 2024 The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import inspect
+from typing import Any, List, Optional, Tuple, Union, Dict
+
+import PIL
+import torch
+import numpy as np
+from collections import OrderedDict
+
+from ...image_processor import VaeImageProcessor, PipelineImageInput
+from ...models import AutoencoderKL
+from ...models.attention_processor import AttnProcessor2_0, XFormersAttnProcessor
+from ...utils import logging
+
+from ...pipelines.stable_diffusion_xl.pipeline_output import StableDiffusionXLPipelineOutput
+from ...configuration_utils import FrozenDict
+
+from ..modular_pipeline_utils import ComponentSpec, ConfigSpec, InputParam, OutputParam
+from ..modular_pipeline import (
+    AutoPipelineBlocks,
+    PipelineBlock,
+    PipelineState,
+    SequentialPipelineBlocks,
+)
+
+logger = logging.get_logger(__name__)  # pylint: disable=invalid-name
+
+
+
+
+class StableDiffusionXLDecodeStep(PipelineBlock):
+
+    model_name = "stable-diffusion-xl"
+    
+    @property
+    def expected_components(self) -> List[ComponentSpec]:
+        return [
+            ComponentSpec("vae", AutoencoderKL),
+            ComponentSpec(
+                "image_processor", 
+                VaeImageProcessor, 
+                config=FrozenDict({"vae_scale_factor": 8}), 
+                default_creation_method="from_config"),
+        ]
+
+    @property
+    def description(self) -> str:
+        return "Step that decodes the denoised latents into images"
+
+    @property
+    def inputs(self) -> List[Tuple[str, Any]]:
+        return [
+            InputParam("output_type", default="pil"),
+        ]
+
+    @property
+    def intermediates_inputs(self) -> List[str]:
+        return [InputParam("latents", required=True, type_hint=torch.Tensor, description="The denoised latents from the denoising step")]
+
+    @property
+    def intermediates_outputs(self) -> List[str]:
+        return [OutputParam("images", type_hint=Union[List[PIL.Image.Image], List[torch.Tensor], List[np.array]], description="The generated images, can be a PIL.Image.Image, torch.Tensor or a numpy array")]
+
+    # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion_upscale.StableDiffusionUpscalePipeline.upcast_vae with self -> components
+    @staticmethod
+    def upcast_vae(components):
+        dtype = components.vae.dtype
+        components.vae.to(dtype=torch.float32)
+        use_torch_2_0_or_xformers = isinstance(
+            components.vae.decoder.mid_block.attentions[0].processor,
+            (
+                AttnProcessor2_0,
+                XFormersAttnProcessor,
+            ),
+        )
+        # if xformers or torch_2_0 is used attention block does not need
+        # to be in float32 which can save lots of memory
+        if use_torch_2_0_or_xformers:
+            components.vae.post_quant_conv.to(dtype)
+            components.vae.decoder.conv_in.to(dtype)
+            components.vae.decoder.mid_block.to(dtype)
+
+    @torch.no_grad()
+    def __call__(self, components, state: PipelineState) -> PipelineState:
+        block_state = self.get_block_state(state)
+
+        if not block_state.output_type == "latent":
+            latents = block_state.latents
+            # make sure the VAE is in float32 mode, as it overflows in float16
+            block_state.needs_upcasting = components.vae.dtype == torch.float16 and components.vae.config.force_upcast
+
+            if block_state.needs_upcasting:
+                self.upcast_vae(components)
+                latents = latents.to(next(iter(components.vae.post_quant_conv.parameters())).dtype)
+            elif latents.dtype != components.vae.dtype:
+                if torch.backends.mps.is_available():
+                    # some platforms (eg. apple mps) misbehave due to a pytorch bug: https://github.com/pytorch/pytorch/pull/99272
+                    components.vae = components.vae.to(latents.dtype)
+
+            # unscale/denormalize the latents
+            # denormalize with the mean and std if available and not None
+            block_state.has_latents_mean = (
+                hasattr(components.vae.config, "latents_mean") and components.vae.config.latents_mean is not None
+            )
+            block_state.has_latents_std = (
+                hasattr(components.vae.config, "latents_std") and components.vae.config.latents_std is not None
+            )
+            if block_state.has_latents_mean and block_state.has_latents_std:
+                block_state.latents_mean = (
+                    torch.tensor(components.vae.config.latents_mean).view(1, 4, 1, 1).to(latents.device, latents.dtype)
+                )
+                block_state.latents_std = (
+                    torch.tensor(components.vae.config.latents_std).view(1, 4, 1, 1).to(latents.device, latents.dtype)
+                )
+                latents = latents * block_state.latents_std / components.vae.config.scaling_factor + block_state.latents_mean
+            else:
+                latents = latents / components.vae.config.scaling_factor
+
+            block_state.images = components.vae.decode(latents, return_dict=False)[0]
+
+            # cast back to fp16 if needed
+            if block_state.needs_upcasting:
+                components.vae.to(dtype=torch.float16)
+        else:
+            block_state.images = block_state.latents
+
+        # apply watermark if available
+        if hasattr(components, "watermark") and components.watermark is not None:
+            block_state.images = components.watermark.apply_watermark(block_state.images)
+
+        block_state.images = components.image_processor.postprocess(block_state.images, output_type=block_state.output_type)
+
+        self.add_block_state(state, block_state)
+
+        return components, state
+
+
+class StableDiffusionXLInpaintOverlayMaskStep(PipelineBlock):
+    model_name = "stable-diffusion-xl"
+
+    @property
+    def description(self) -> str:
+        return "A post-processing step that overlays the mask on the image (inpainting task only).\n" + \
+               "only needed when you are using the `padding_mask_crop` option when pre-processing the image and mask"
+
+    @property
+    def inputs(self) -> List[Tuple[str, Any]]:
+        return [
+            InputParam("image", required=True),
+            InputParam("mask_image", required=True), 
+            InputParam("padding_mask_crop"),
+        ]
+    
+    @property
+    def intermediates_inputs(self) -> List[str]:
+        return [
+            InputParam("images", required=True, type_hint=Union[List[PIL.Image.Image], List[torch.Tensor], List[np.array]], description="The generated images from the decode step"),
+            InputParam("crops_coords", required=True, type_hint=Tuple[int, int], description="The crop coordinates to use for preprocess/postprocess the image and mask, for inpainting task only. Can be generated in vae_encode step.")
+        ]
+
+    @property
+    def intermediates_outputs(self) -> List[str]:
+        return [OutputParam("images", type_hint=Union[List[PIL.Image.Image], List[torch.Tensor], List[np.array]], description="The generated images with the mask overlayed")]
+
+    @torch.no_grad()
+    def __call__(self, components, state: PipelineState) -> PipelineState:
+        block_state = self.get_block_state(state)
+
+        if block_state.padding_mask_crop is not None and block_state.crops_coords is not None:
+            block_state.images = [components.image_processor.apply_overlay(block_state.mask_image, block_state.image, i, block_state.crops_coords) for i in block_state.images]
+
+        self.add_block_state(state, block_state)
+
+        return components, state
+
+
+
+class StableDiffusionXLInpaintDecodeStep(SequentialPipelineBlocks):
+    block_classes = [StableDiffusionXLDecodeStep, StableDiffusionXLInpaintOverlayMaskStep]
+    block_names = ["decode", "mask_overlay"]
+
+    @property
+    def description(self):
+        return "Inpaint decode step that decode the denoised latents into images outputs.\n" + \
+               "This is a sequential pipeline blocks:\n" + \
+               " - `StableDiffusionXLDecodeStep` is used to decode the denoised latents into images\n" + \
+               " - `StableDiffusionXLInpaintOverlayMaskStep` is used to overlay the mask on the image"
+
+
+class StableDiffusionXLAutoDecodeStep(AutoPipelineBlocks):
+    block_classes = [StableDiffusionXLInpaintDecodeStep, StableDiffusionXLDecodeStep]
+    block_names = ["inpaint", "non-inpaint"]
+    block_trigger_inputs = ["padding_mask_crop", None]
+
+    @property
+    def description(self):
+        return "Decode step that decode the denoised latents into images outputs.\n" + \
+               "This is an auto pipeline block that works for inpainting and non-inpainting tasks.\n" + \
+               " - `StableDiffusionXLInpaintDecodeStep` (inpaint) is used when `padding_mask_crop` is provided.\n" + \
+               " - `StableDiffusionXLDecodeStep` (non-inpaint) is used when `padding_mask_crop` is not provided."
+
+

src/diffusers/modular_pipelines/stable_diffusion_xl/encoders.py

@@ -0,0 +1,858 @@
+# Copyright 2024 The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import inspect
+from typing import Any, List, Optional, Tuple, Union, Dict
+
+import PIL
+import torch
+from collections import OrderedDict
+
+from ...image_processor import VaeImageProcessor, PipelineImageInput
+from ...loaders import StableDiffusionXLLoraLoaderMixin, TextualInversionLoaderMixin, ModularIPAdapterMixin
+from ...models import ControlNetModel, ImageProjection, UNet2DConditionModel, AutoencoderKL, ControlNetUnionModel
+from ...models.attention_processor import AttnProcessor2_0, XFormersAttnProcessor
+from ...models.lora import adjust_lora_scale_text_encoder
+from ...utils import (
+    USE_PEFT_BACKEND,
+    logging,
+    scale_lora_layers,
+    unscale_lora_layers,
+)
+from ...utils.torch_utils import randn_tensor, unwrap_module
+from ...pipelines.controlnet.multicontrolnet import MultiControlNetModel
+from ...configuration_utils import FrozenDict
+
+from transformers import (
+    CLIPTextModel,
+    CLIPImageProcessor,
+    CLIPTextModelWithProjection,
+    CLIPTokenizer,
+    CLIPVisionModelWithProjection,
+)
+
+from ...schedulers import EulerDiscreteScheduler
+from ...guiders import ClassifierFreeGuidance
+
+from .modular_loader import StableDiffusionXLModularLoader
+from ..modular_pipeline import PipelineBlock, PipelineState, AutoPipelineBlocks, SequentialPipelineBlocks
+from ..modular_pipeline_utils import ComponentSpec, InputParam, OutputParam, ConfigSpec
+
+import numpy as np
+
+logger = logging.get_logger(__name__)  # pylint: disable=invalid-name
+
+
+# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion_img2img.retrieve_latents
+def retrieve_latents(
+    encoder_output: torch.Tensor, generator: Optional[torch.Generator] = None, sample_mode: str = "sample"
+):
+    if hasattr(encoder_output, "latent_dist") and sample_mode == "sample":
+        return encoder_output.latent_dist.sample(generator)
+    elif hasattr(encoder_output, "latent_dist") and sample_mode == "argmax":
+        return encoder_output.latent_dist.mode()
+    elif hasattr(encoder_output, "latents"):
+        return encoder_output.latents
+    else:
+        raise AttributeError("Could not access latents of provided encoder_output")
+
+
+class StableDiffusionXLIPAdapterStep(PipelineBlock):
+    model_name = "stable-diffusion-xl"
+
+    
+    @property
+    def description(self) -> str:
+        return (
+            "IP Adapter step that handles all the ip adapter related tasks: Load/unload ip adapter weights into unet, prepare ip adapter image embeddings, etc"
+            " See [ModularIPAdapterMixin](https://huggingface.co/docs/diffusers/api/loaders/ip_adapter#diffusers.loaders.ModularIPAdapterMixin)"
+            " for more details"
+        )
+    
+    @property
+    def expected_components(self) -> List[ComponentSpec]:
+        return [
+            ComponentSpec("image_encoder", CLIPVisionModelWithProjection),
+            ComponentSpec("feature_extractor", CLIPImageProcessor, config=FrozenDict({"size": 224, "crop_size": 224}), default_creation_method="from_config"),
+            ComponentSpec("unet", UNet2DConditionModel),
+            ComponentSpec(
+                "guider", 
+                ClassifierFreeGuidance, 
+                config=FrozenDict({"guidance_scale": 7.5}),
+                default_creation_method="from_config"),
+        ]
+
+    @property
+    def inputs(self) -> List[InputParam]:
+        return [
+            InputParam(
+                "ip_adapter_image", 
+                PipelineImageInput, 
+                required=True,
+                description="The image(s) to be used as ip adapter"
+            )
+        ]
+
+
+    @property
+    def intermediates_outputs(self) -> List[OutputParam]:
+        return [
+            OutputParam("ip_adapter_embeds", type_hint=torch.Tensor, description="IP adapter image embeddings"),
+            OutputParam("negative_ip_adapter_embeds", type_hint=torch.Tensor, description="Negative IP adapter image embeddings")
+        ]
+    
+    # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.encode_image with self -> components
+    @staticmethod
+    def encode_image(components, image, device, num_images_per_prompt, output_hidden_states=None):
+        dtype = next(components.image_encoder.parameters()).dtype
+
+        if not isinstance(image, torch.Tensor):
+            image = components.feature_extractor(image, return_tensors="pt").pixel_values
+
+        image = image.to(device=device, dtype=dtype)
+        if output_hidden_states:
+            image_enc_hidden_states = components.image_encoder(image, output_hidden_states=True).hidden_states[-2]
+            image_enc_hidden_states = image_enc_hidden_states.repeat_interleave(num_images_per_prompt, dim=0)
+            uncond_image_enc_hidden_states = components.image_encoder(
+                torch.zeros_like(image), output_hidden_states=True
+            ).hidden_states[-2]
+            uncond_image_enc_hidden_states = uncond_image_enc_hidden_states.repeat_interleave(
+                num_images_per_prompt, dim=0
+            )
+            return image_enc_hidden_states, uncond_image_enc_hidden_states
+        else:
+            image_embeds = components.image_encoder(image).image_embeds
+            image_embeds = image_embeds.repeat_interleave(num_images_per_prompt, dim=0)
+            uncond_image_embeds = torch.zeros_like(image_embeds)
+
+            return image_embeds, uncond_image_embeds
+    
+    # modified from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.prepare_ip_adapter_image_embeds
+    def prepare_ip_adapter_image_embeds(
+        self, components, ip_adapter_image, ip_adapter_image_embeds, device, num_images_per_prompt, prepare_unconditional_embeds
+    ):
+        image_embeds = []
+        if prepare_unconditional_embeds:
+            negative_image_embeds = []
+        if ip_adapter_image_embeds is None:
+            if not isinstance(ip_adapter_image, list):
+                ip_adapter_image = [ip_adapter_image]
+
+            if len(ip_adapter_image) != len(components.unet.encoder_hid_proj.image_projection_layers):
+                raise ValueError(
+                    f"`ip_adapter_image` must have same length as the number of IP Adapters. Got {len(ip_adapter_image)} images and {len(components.unet.encoder_hid_proj.image_projection_layers)} IP Adapters."
+                )
+
+            for single_ip_adapter_image, image_proj_layer in zip(
+                ip_adapter_image, components.unet.encoder_hid_proj.image_projection_layers
+            ):
+                output_hidden_state = not isinstance(image_proj_layer, ImageProjection)
+                single_image_embeds, single_negative_image_embeds = self.encode_image(
+                    components, single_ip_adapter_image, device, 1, output_hidden_state
+                )
+
+                image_embeds.append(single_image_embeds[None, :])
+                if prepare_unconditional_embeds:
+                    negative_image_embeds.append(single_negative_image_embeds[None, :])
+        else:
+            for single_image_embeds in ip_adapter_image_embeds:
+                if prepare_unconditional_embeds:
+                    single_negative_image_embeds, single_image_embeds = single_image_embeds.chunk(2)
+                    negative_image_embeds.append(single_negative_image_embeds)
+                image_embeds.append(single_image_embeds)
+
+        ip_adapter_image_embeds = []
+        for i, single_image_embeds in enumerate(image_embeds):
+            single_image_embeds = torch.cat([single_image_embeds] * num_images_per_prompt, dim=0)
+            if prepare_unconditional_embeds:
+                single_negative_image_embeds = torch.cat([negative_image_embeds[i]] * num_images_per_prompt, dim=0)
+                single_image_embeds = torch.cat([single_negative_image_embeds, single_image_embeds], dim=0)
+
+            single_image_embeds = single_image_embeds.to(device=device)
+            ip_adapter_image_embeds.append(single_image_embeds)
+
+        return ip_adapter_image_embeds
+
+    @torch.no_grad()
+    def __call__(self, components: StableDiffusionXLModularLoader, state: PipelineState) -> PipelineState:
+        block_state = self.get_block_state(state)
+
+        block_state.prepare_unconditional_embeds = components.guider.num_conditions > 1
+        block_state.device = components._execution_device
+
+        block_state.ip_adapter_embeds = self.prepare_ip_adapter_image_embeds(
+            components,
+            ip_adapter_image=block_state.ip_adapter_image,
+            ip_adapter_image_embeds=None,
+            device=block_state.device,
+            num_images_per_prompt=1,
+            prepare_unconditional_embeds=block_state.prepare_unconditional_embeds,
+        )
+        if block_state.prepare_unconditional_embeds:
+            block_state.negative_ip_adapter_embeds = []
+            for i, image_embeds in enumerate(block_state.ip_adapter_embeds):
+                negative_image_embeds, image_embeds = image_embeds.chunk(2)
+                block_state.negative_ip_adapter_embeds.append(negative_image_embeds)
+                block_state.ip_adapter_embeds[i] = image_embeds
+
+        self.add_block_state(state, block_state)
+        return components, state
+
+
+class StableDiffusionXLTextEncoderStep(PipelineBlock):
+
+    model_name = "stable-diffusion-xl"
+
+    @property
+    def description(self) -> str:
+        return(
+            "Text Encoder step that generate text_embeddings to guide the image generation"
+        )
+    
+    @property
+    def expected_components(self) -> List[ComponentSpec]:
+        return [
+            ComponentSpec("text_encoder", CLIPTextModel),
+            ComponentSpec("text_encoder_2", CLIPTextModelWithProjection),
+            ComponentSpec("tokenizer", CLIPTokenizer),
+            ComponentSpec("tokenizer_2", CLIPTokenizer),
+            ComponentSpec(
+                "guider", 
+                ClassifierFreeGuidance, 
+                config=FrozenDict({"guidance_scale": 7.5}), 
+                default_creation_method="from_config"),
+        ]
+
+    @property
+    def expected_configs(self) -> List[ConfigSpec]:
+        return [ConfigSpec("force_zeros_for_empty_prompt", True)]
+
+    @property
+    def inputs(self) -> List[InputParam]:
+        return [
+            InputParam("prompt"),
+            InputParam("prompt_2"),
+            InputParam("negative_prompt"),
+            InputParam("negative_prompt_2"),
+            InputParam("cross_attention_kwargs"),
+            InputParam("clip_skip"),
+        ]
+
+
+    @property
+    def intermediates_outputs(self) -> List[OutputParam]:
+        return [
+            OutputParam("prompt_embeds", type_hint=torch.Tensor, kwargs_type="guider_input_fields",description="text embeddings used to guide the image generation"),
+            OutputParam("negative_prompt_embeds", type_hint=torch.Tensor, kwargs_type="guider_input_fields", description="negative text embeddings used to guide the image generation"),
+            OutputParam("pooled_prompt_embeds", type_hint=torch.Tensor, kwargs_type="guider_input_fields", description="pooled text embeddings used to guide the image generation"),
+            OutputParam("negative_pooled_prompt_embeds", type_hint=torch.Tensor, kwargs_type="guider_input_fields", description="negative pooled text embeddings used to guide the image generation"),
+        ]
+
+    @staticmethod
+    def check_inputs(block_state):
+
+        if block_state.prompt is not None and (not isinstance(block_state.prompt, str) and not isinstance(block_state.prompt, list)):
+            raise ValueError(f"`prompt` has to be of type `str` or `list` but is {type(block_state.prompt)}")
+        elif block_state.prompt_2 is not None and (not isinstance(block_state.prompt_2, str) and not isinstance(block_state.prompt_2, list)):
+            raise ValueError(f"`prompt_2` has to be of type `str` or `list` but is {type(block_state.prompt_2)}")
+
+    @staticmethod
+    def encode_prompt(
+        components,
+        prompt: str,
+        prompt_2: Optional[str] = None,
+        device: Optional[torch.device] = None,
+        num_images_per_prompt: int = 1,
+        prepare_unconditional_embeds: bool = True,
+        negative_prompt: Optional[str] = None,
+        negative_prompt_2: Optional[str] = None,
+        prompt_embeds: Optional[torch.Tensor] = None,
+        negative_prompt_embeds: Optional[torch.Tensor] = None,
+        pooled_prompt_embeds: Optional[torch.Tensor] = None,
+        negative_pooled_prompt_embeds: Optional[torch.Tensor] = None,
+        lora_scale: Optional[float] = None,
+        clip_skip: Optional[int] = None,
+    ):
+        r"""
+        Encodes the prompt into text encoder hidden states.
+
+        Args:
+            prompt (`str` or `List[str]`, *optional*):
+                prompt to be encoded
+            prompt_2 (`str` or `List[str]`, *optional*):
+                The prompt or prompts to be sent to the `tokenizer_2` and `text_encoder_2`. If not defined, `prompt` is
+                used in both text-encoders
+            device: (`torch.device`):
+                torch device
+            num_images_per_prompt (`int`):
+                number of images that should be generated per prompt
+            prepare_unconditional_embeds (`bool`):
+                whether to use prepare unconditional embeddings or not
+            negative_prompt (`str` or `List[str]`, *optional*):
+                The prompt or prompts not to guide the image generation. If not defined, one has to pass
+                `negative_prompt_embeds` instead. Ignored when not using guidance (i.e., ignored if `guidance_scale` is
+                less than `1`).
+            negative_prompt_2 (`str` or `List[str]`, *optional*):
+                The prompt or prompts not to guide the image generation to be sent to `tokenizer_2` and
+                `text_encoder_2`. If not defined, `negative_prompt` is used in both text-encoders
+            prompt_embeds (`torch.Tensor`, *optional*):
+                Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not
+                provided, text embeddings will be generated from `prompt` input argument.
+            negative_prompt_embeds (`torch.Tensor`, *optional*):
+                Pre-generated negative text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt
+                weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt` input
+                argument.
+            pooled_prompt_embeds (`torch.Tensor`, *optional*):
+                Pre-generated pooled text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting.
+                If not provided, pooled text embeddings will be generated from `prompt` input argument.
+            negative_pooled_prompt_embeds (`torch.Tensor`, *optional*):
+                Pre-generated negative pooled text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt
+                weighting. If not provided, pooled negative_prompt_embeds will be generated from `negative_prompt`
+                input argument.
+            lora_scale (`float`, *optional*):
+                A lora scale that will be applied to all LoRA layers of the text encoder if LoRA layers are loaded.
+            clip_skip (`int`, *optional*):
+                Number of layers to be skipped from CLIP while computing the prompt embeddings. A value of 1 means that
+                the output of the pre-final layer will be used for computing the prompt embeddings.
+        """
+        device = device or components._execution_device
+
+        # set lora scale so that monkey patched LoRA
+        # function of text encoder can correctly access it
+        if lora_scale is not None and isinstance(components, StableDiffusionXLLoraLoaderMixin):
+            components._lora_scale = lora_scale
+
+            # dynamically adjust the LoRA scale
+            if components.text_encoder is not None:
+                if not USE_PEFT_BACKEND:
+                    adjust_lora_scale_text_encoder(components.text_encoder, lora_scale)
+                else:
+                    scale_lora_layers(components.text_encoder, lora_scale)
+
+            if components.text_encoder_2 is not None:
+                if not USE_PEFT_BACKEND:
+                    adjust_lora_scale_text_encoder(components.text_encoder_2, lora_scale)
+                else:
+                    scale_lora_layers(components.text_encoder_2, lora_scale)
+
+        prompt = [prompt] if isinstance(prompt, str) else prompt
+
+        if prompt is not None:
+            batch_size = len(prompt)
+        else:
+            batch_size = prompt_embeds.shape[0]
+
+        # Define tokenizers and text encoders
+        tokenizers = [components.tokenizer, components.tokenizer_2] if components.tokenizer is not None else [components.tokenizer_2]
+        text_encoders = (
+            [components.text_encoder, components.text_encoder_2] if components.text_encoder is not None else [components.text_encoder_2]
+        )
+
+        if prompt_embeds is None:
+            prompt_2 = prompt_2 or prompt
+            prompt_2 = [prompt_2] if isinstance(prompt_2, str) else prompt_2
+
+            # textual inversion: process multi-vector tokens if necessary
+            prompt_embeds_list = []
+            prompts = [prompt, prompt_2]
+            for prompt, tokenizer, text_encoder in zip(prompts, tokenizers, text_encoders):
+                if isinstance(components, TextualInversionLoaderMixin):
+                    prompt = components.maybe_convert_prompt(prompt, tokenizer)
+
+                text_inputs = tokenizer(
+                    prompt,
+                    padding="max_length",
+                    max_length=tokenizer.model_max_length,
+                    truncation=True,
+                    return_tensors="pt",
+                )
+
+                text_input_ids = text_inputs.input_ids
+                untruncated_ids = tokenizer(prompt, padding="longest", return_tensors="pt").input_ids
+
+                if untruncated_ids.shape[-1] >= text_input_ids.shape[-1] and not torch.equal(
+                    text_input_ids, untruncated_ids
+                ):
+                    removed_text = tokenizer.batch_decode(untruncated_ids[:, tokenizer.model_max_length - 1 : -1])
+                    logger.warning(
+                        "The following part of your input was truncated because CLIP can only handle sequences up to"
+                        f" {tokenizer.model_max_length} tokens: {removed_text}"
+                    )
+
+                prompt_embeds = text_encoder(text_input_ids.to(device), output_hidden_states=True)
+
+                # We are only ALWAYS interested in the pooled output of the final text encoder
+                pooled_prompt_embeds = prompt_embeds[0]
+                if clip_skip is None:
+                    prompt_embeds = prompt_embeds.hidden_states[-2]
+                else:
+                    # "2" because SDXL always indexes from the penultimate layer.
+                    prompt_embeds = prompt_embeds.hidden_states[-(clip_skip + 2)]
+
+                prompt_embeds_list.append(prompt_embeds)
+
+            prompt_embeds = torch.concat(prompt_embeds_list, dim=-1)
+
+        # get unconditional embeddings for classifier free guidance
+        zero_out_negative_prompt = negative_prompt is None and components.config.force_zeros_for_empty_prompt
+        if prepare_unconditional_embeds and negative_prompt_embeds is None and zero_out_negative_prompt:
+            negative_prompt_embeds = torch.zeros_like(prompt_embeds)
+            negative_pooled_prompt_embeds = torch.zeros_like(pooled_prompt_embeds)
+        elif prepare_unconditional_embeds and negative_prompt_embeds is None:
+            negative_prompt = negative_prompt or ""
+            negative_prompt_2 = negative_prompt_2 or negative_prompt
+
+            # normalize str to list
+            negative_prompt = batch_size * [negative_prompt] if isinstance(negative_prompt, str) else negative_prompt
+            negative_prompt_2 = (
+                batch_size * [negative_prompt_2] if isinstance(negative_prompt_2, str) else negative_prompt_2
+            )
+
+            uncond_tokens: List[str]
+            if prompt is not None and type(prompt) is not type(negative_prompt):
+                raise TypeError(
+                    f"`negative_prompt` should be the same type to `prompt`, but got {type(negative_prompt)} !="
+                    f" {type(prompt)}."
+                )
+            elif batch_size != len(negative_prompt):
+                raise ValueError(
+                    f"`negative_prompt`: {negative_prompt} has batch size {len(negative_prompt)}, but `prompt`:"
+                    f" {prompt} has batch size {batch_size}. Please make sure that passed `negative_prompt` matches"
+                    " the batch size of `prompt`."
+                )
+            else:
+                uncond_tokens = [negative_prompt, negative_prompt_2]
+
+            negative_prompt_embeds_list = []
+            for negative_prompt, tokenizer, text_encoder in zip(uncond_tokens, tokenizers, text_encoders):
+                if isinstance(components, TextualInversionLoaderMixin):
+                    negative_prompt = components.maybe_convert_prompt(negative_prompt, tokenizer)
+
+                max_length = prompt_embeds.shape[1]
+                uncond_input = tokenizer(
+                    negative_prompt,
+                    padding="max_length",
+                    max_length=max_length,
+                    truncation=True,
+                    return_tensors="pt",
+                )
+
+                negative_prompt_embeds = text_encoder(
+                    uncond_input.input_ids.to(device),
+                    output_hidden_states=True,
+                )
+                # We are only ALWAYS interested in the pooled output of the final text encoder
+                negative_pooled_prompt_embeds = negative_prompt_embeds[0]
+                negative_prompt_embeds = negative_prompt_embeds.hidden_states[-2]
+
+                negative_prompt_embeds_list.append(negative_prompt_embeds)
+
+            negative_prompt_embeds = torch.concat(negative_prompt_embeds_list, dim=-1)
+
+        if components.text_encoder_2 is not None:
+            prompt_embeds = prompt_embeds.to(dtype=components.text_encoder_2.dtype, device=device)
+        else:
+            prompt_embeds = prompt_embeds.to(dtype=components.unet.dtype, device=device)
+
+        bs_embed, seq_len, _ = prompt_embeds.shape
+        # duplicate text embeddings for each generation per prompt, using mps friendly method
+        prompt_embeds = prompt_embeds.repeat(1, num_images_per_prompt, 1)
+        prompt_embeds = prompt_embeds.view(bs_embed * num_images_per_prompt, seq_len, -1)
+
+        if prepare_unconditional_embeds:
+            # duplicate unconditional embeddings for each generation per prompt, using mps friendly method
+            seq_len = negative_prompt_embeds.shape[1]
+
+            if components.text_encoder_2 is not None:
+                negative_prompt_embeds = negative_prompt_embeds.to(dtype=components.text_encoder_2.dtype, device=device)
+            else:
+                negative_prompt_embeds = negative_prompt_embeds.to(dtype=components.unet.dtype, device=device)
+
+            negative_prompt_embeds = negative_prompt_embeds.repeat(1, num_images_per_prompt, 1)
+            negative_prompt_embeds = negative_prompt_embeds.view(batch_size * num_images_per_prompt, seq_len, -1)
+
+        pooled_prompt_embeds = pooled_prompt_embeds.repeat(1, num_images_per_prompt).view(
+            bs_embed * num_images_per_prompt, -1
+        )
+        if prepare_unconditional_embeds:
+            negative_pooled_prompt_embeds = negative_pooled_prompt_embeds.repeat(1, num_images_per_prompt).view(
+                bs_embed * num_images_per_prompt, -1
+            )
+
+        if components.text_encoder is not None:
+            if isinstance(components, StableDiffusionXLLoraLoaderMixin) and USE_PEFT_BACKEND:
+                # Retrieve the original scale by scaling back the LoRA layers
+                unscale_lora_layers(components.text_encoder, lora_scale)
+
+        if components.text_encoder_2 is not None:
+            if isinstance(components, StableDiffusionXLLoraLoaderMixin) and USE_PEFT_BACKEND:
+                # Retrieve the original scale by scaling back the LoRA layers
+                unscale_lora_layers(components.text_encoder_2, lora_scale)
+
+        return prompt_embeds, negative_prompt_embeds, pooled_prompt_embeds, negative_pooled_prompt_embeds
+
+
+    @torch.no_grad()
+    def __call__(self, components: StableDiffusionXLModularLoader, state: PipelineState) -> PipelineState:
+        # Get inputs and intermediates
+        block_state = self.get_block_state(state)
+        self.check_inputs(block_state)
+
+        block_state.prepare_unconditional_embeds = components.guider.num_conditions > 1
+        block_state.device = components._execution_device
+
+        # Encode input prompt
+        block_state.text_encoder_lora_scale = (
+            block_state.cross_attention_kwargs.get("scale", None) if block_state.cross_attention_kwargs is not None else None
+        )
+        (
+            block_state.prompt_embeds,
+            block_state.negative_prompt_embeds,
+            block_state.pooled_prompt_embeds,
+            block_state.negative_pooled_prompt_embeds,
+        ) = self.encode_prompt(
+            components,
+            block_state.prompt,
+            block_state.prompt_2,
+            block_state.device,
+            1,
+            block_state.prepare_unconditional_embeds,
+            block_state.negative_prompt,
+            block_state.negative_prompt_2,
+            prompt_embeds=None,
+            negative_prompt_embeds=None,
+            pooled_prompt_embeds=None,
+            negative_pooled_prompt_embeds=None,
+            lora_scale=block_state.text_encoder_lora_scale,
+            clip_skip=block_state.clip_skip,
+        )
+        # Add outputs
+        self.add_block_state(state, block_state)
+        return components, state
+
+
+class StableDiffusionXLVaeEncoderStep(PipelineBlock):
+
+    model_name = "stable-diffusion-xl"
+
+    
+    @property
+    def description(self) -> str:
+        return (
+            "Vae Encoder step that encode the input image into a latent representation"
+        )
+
+    @property
+    def expected_components(self) -> List[ComponentSpec]:
+        return [
+            ComponentSpec("vae", AutoencoderKL),
+            ComponentSpec(
+                "image_processor", 
+                VaeImageProcessor, 
+                config=FrozenDict({"vae_scale_factor": 8}), 
+                default_creation_method="from_config"),
+        ]
+
+    @property
+    def inputs(self) -> List[InputParam]:
+        return [
+            InputParam("image", required=True),
+            InputParam("height"),
+            InputParam("width"),
+        ]
+
+    @property
+    def intermediates_inputs(self) -> List[InputParam]:
+        return [
+            InputParam("generator"),
+            InputParam("dtype", type_hint=torch.dtype, description="Data type of model tensor inputs"), 
+            InputParam("preprocess_kwargs", type_hint=Optional[dict], description="A kwargs dictionary that if specified is passed along to the `ImageProcessor` as defined under `self.image_processor` in [diffusers.image_processor.VaeImageProcessor]")]
+
+    @property
+    def intermediates_outputs(self) -> List[OutputParam]:
+        return [OutputParam("image_latents", type_hint=torch.Tensor, description="The latents representing the reference image for image-to-image/inpainting generation")]
+
+    # Modified from diffusers.pipelines.stable_diffusion_xl.pipeline_stable_diffusion_xl_inpaint.StableDiffusionXLInpaintPipeline._encode_vae_image with self -> components
+    # YiYi TODO: update the _encode_vae_image so that we can use #Coped from
+    def _encode_vae_image(self, components, image: torch.Tensor, generator: torch.Generator):
+        
+        latents_mean = latents_std = None
+        if hasattr(components.vae.config, "latents_mean") and components.vae.config.latents_mean is not None:
+            latents_mean = torch.tensor(components.vae.config.latents_mean).view(1, 4, 1, 1)
+        if hasattr(components.vae.config, "latents_std") and components.vae.config.latents_std is not None:
+            latents_std = torch.tensor(components.vae.config.latents_std).view(1, 4, 1, 1)
+        
+        dtype = image.dtype
+        if components.vae.config.force_upcast:
+            image = image.float()
+            components.vae.to(dtype=torch.float32)
+
+        if isinstance(generator, list):
+            image_latents = [
+                retrieve_latents(components.vae.encode(image[i : i + 1]), generator=generator[i])
+                for i in range(image.shape[0])
+            ]
+            image_latents = torch.cat(image_latents, dim=0)
+        else:
+            image_latents = retrieve_latents(components.vae.encode(image), generator=generator)
+
+        if components.vae.config.force_upcast:
+            components.vae.to(dtype)
+
+        image_latents = image_latents.to(dtype)
+        if latents_mean is not None and latents_std is not None:
+            latents_mean = latents_mean.to(device=image_latents.device, dtype=dtype)
+            latents_std = latents_std.to(device=image_latents.device, dtype=dtype)
+            image_latents = (image_latents - latents_mean) * components.vae.config.scaling_factor / latents_std
+        else:
+            image_latents = components.vae.config.scaling_factor * image_latents
+
+        return image_latents 
+    
+
+
+    @torch.no_grad()
+    def __call__(self, components: StableDiffusionXLModularLoader, state: PipelineState) -> PipelineState:
+        block_state = self.get_block_state(state)
+        block_state.preprocess_kwargs = block_state.preprocess_kwargs or {}
+        block_state.device = components._execution_device
+        block_state.dtype = block_state.dtype if block_state.dtype is not None else components.vae.dtype
+        
+        block_state.image = components.image_processor.preprocess(block_state.image, height=block_state.height, width=block_state.width, **block_state.preprocess_kwargs)
+        block_state.image = block_state.image.to(device=block_state.device, dtype=block_state.dtype)
+
+        block_state.batch_size = block_state.image.shape[0]
+
+        # if generator is a list, make sure the length of it matches the length of images (both should be batch_size)
+        if isinstance(block_state.generator, list) and len(block_state.generator) != block_state.batch_size:
+            raise ValueError(
+                f"You have passed a list of generators of length {len(block_state.generator)}, but requested an effective batch"
+                f" size of {block_state.batch_size}. Make sure the batch size matches the length of the generators."
+            )
+
+
+        block_state.image_latents = self._encode_vae_image(components, image=block_state.image, generator=block_state.generator)
+
+        self.add_block_state(state, block_state)
+
+        return components, state
+
+
+class StableDiffusionXLInpaintVaeEncoderStep(PipelineBlock):
+    model_name = "stable-diffusion-xl"
+    
+    @property
+    def expected_components(self) -> List[ComponentSpec]:
+        return [
+            ComponentSpec("vae", AutoencoderKL),
+            ComponentSpec(
+                "image_processor", 
+                VaeImageProcessor, 
+                config=FrozenDict({"vae_scale_factor": 8}), 
+                default_creation_method="from_config"),
+            ComponentSpec(
+                "mask_processor", 
+                VaeImageProcessor, 
+                config=FrozenDict({"do_normalize": False, "vae_scale_factor": 8, "do_binarize": True, "do_convert_grayscale": True}),
+                default_creation_method="from_config"),
+        ]
+        
+
+    @property
+    def description(self) -> str:
+        return (
+            "Vae encoder step that prepares the image and mask for the inpainting process"
+        )
+
+    @property
+    def inputs(self) -> List[InputParam]:
+        return [
+            InputParam("height"),
+            InputParam("width"),
+            InputParam("image", required=True),
+            InputParam("mask_image", required=True),
+            InputParam("padding_mask_crop"),
+        ]
+
+    @property
+    def intermediates_inputs(self) -> List[InputParam]:
+        return [
+            InputParam("dtype", type_hint=torch.dtype, description="The dtype of the model inputs"),
+            InputParam("generator"),
+        ]
+
+    @property
+    def intermediates_outputs(self) -> List[OutputParam]:
+        return [OutputParam("image_latents", type_hint=torch.Tensor, description="The latents representation of the input image"), 
+                OutputParam("mask", type_hint=torch.Tensor, description="The mask to use for the inpainting process"), 
+                OutputParam("masked_image_latents", type_hint=torch.Tensor, description="The masked image latents to use for the inpainting process (only for inpainting-specifid unet)"), 
+                OutputParam("crops_coords", type_hint=Optional[Tuple[int, int]], description="The crop coordinates to use for the preprocess/postprocess of the image and mask")]
+
+    # Modified from diffusers.pipelines.stable_diffusion_xl.pipeline_stable_diffusion_xl_inpaint.StableDiffusionXLInpaintPipeline._encode_vae_image with self -> components
+    # YiYi TODO: update the _encode_vae_image so that we can use #Coped from
+    def _encode_vae_image(self, components, image: torch.Tensor, generator: torch.Generator):
+        
+        latents_mean = latents_std = None
+        if hasattr(components.vae.config, "latents_mean") and components.vae.config.latents_mean is not None:
+            latents_mean = torch.tensor(components.vae.config.latents_mean).view(1, 4, 1, 1)
+        if hasattr(components.vae.config, "latents_std") and components.vae.config.latents_std is not None:
+            latents_std = torch.tensor(components.vae.config.latents_std).view(1, 4, 1, 1)
+        
+        dtype = image.dtype
+        if components.vae.config.force_upcast:
+            image = image.float()
+            components.vae.to(dtype=torch.float32)
+
+        if isinstance(generator, list):
+            image_latents = [
+                retrieve_latents(components.vae.encode(image[i : i + 1]), generator=generator[i])
+                for i in range(image.shape[0])
+            ]
+            image_latents = torch.cat(image_latents, dim=0)
+        else:
+            image_latents = retrieve_latents(components.vae.encode(image), generator=generator)
+
+        if components.vae.config.force_upcast:
+            components.vae.to(dtype)
+
+        image_latents = image_latents.to(dtype)
+        if latents_mean is not None and latents_std is not None:
+            latents_mean = latents_mean.to(device=image_latents.device, dtype=dtype)
+            latents_std = latents_std.to(device=image_latents.device, dtype=dtype)
+            image_latents = (image_latents - latents_mean) * self.vae.config.scaling_factor / latents_std
+        else:
+            image_latents = components.vae.config.scaling_factor * image_latents
+
+        return image_latents 
+
+    # modified from diffusers.pipelines.stable_diffusion_xl.pipeline_stable_diffusion_xl_inpaint.StableDiffusionXLInpaintPipeline.prepare_mask_latents
+    # do not accept do_classifier_free_guidance
+    def prepare_mask_latents(
+        self, components, mask, masked_image, batch_size, height, width, dtype, device, generator
+    ):
+        # resize the mask to latents shape as we concatenate the mask to the latents
+        # we do that before converting to dtype to avoid breaking in case we're using cpu_offload
+        # and half precision
+        mask = torch.nn.functional.interpolate(
+            mask, size=(height // components.vae_scale_factor, width // components.vae_scale_factor)
+        )
+        mask = mask.to(device=device, dtype=dtype)
+
+        # duplicate mask and masked_image_latents for each generation per prompt, using mps friendly method
+        if mask.shape[0] < batch_size:
+            if not batch_size % mask.shape[0] == 0:
+                raise ValueError(
+                    "The passed mask and the required batch size don't match. Masks are supposed to be duplicated to"
+                    f" a total batch size of {batch_size}, but {mask.shape[0]} masks were passed. Make sure the number"
+                    " of masks that you pass is divisible by the total requested batch size."
+                )
+            mask = mask.repeat(batch_size // mask.shape[0], 1, 1, 1)
+
+        if masked_image is not None and masked_image.shape[1] == 4:
+            masked_image_latents = masked_image
+        else:
+            masked_image_latents = None
+
+        if masked_image is not None:
+            if masked_image_latents is None:
+                masked_image = masked_image.to(device=device, dtype=dtype)
+                masked_image_latents = self._encode_vae_image(components, masked_image, generator=generator)
+
+            if masked_image_latents.shape[0] < batch_size:
+                if not batch_size % masked_image_latents.shape[0] == 0:
+                    raise ValueError(
+                        "The passed images and the required batch size don't match. Images are supposed to be duplicated"
+                        f" to a total batch size of {batch_size}, but {masked_image_latents.shape[0]} images were passed."
+                        " Make sure the number of images that you pass is divisible by the total requested batch size."
+                    )
+                masked_image_latents = masked_image_latents.repeat(
+                    batch_size // masked_image_latents.shape[0], 1, 1, 1
+                )
+
+            # aligning device to prevent device errors when concating it with the latent model input
+            masked_image_latents = masked_image_latents.to(device=device, dtype=dtype)
+
+        return mask, masked_image_latents
+    
+     
+
+    @torch.no_grad()
+    def __call__(self, components: StableDiffusionXLModularLoader, state: PipelineState) -> PipelineState:
+
+        block_state = self.get_block_state(state)
+
+        block_state.dtype = block_state.dtype if block_state.dtype is not None else components.vae.dtype
+        block_state.device = components._execution_device
+
+        if block_state.padding_mask_crop is not None:
+            block_state.crops_coords = components.mask_processor.get_crop_region(block_state.mask_image, block_state.width, block_state.height, pad=block_state.padding_mask_crop)
+            block_state.resize_mode = "fill"
+        else:
+            block_state.crops_coords = None
+            block_state.resize_mode = "default"
+        
+        block_state.image = components.image_processor.preprocess(block_state.image, height=block_state.height, width=block_state.width, crops_coords=block_state.crops_coords, resize_mode=block_state.resize_mode)
+        block_state.image = block_state.image.to(dtype=torch.float32)
+
+        block_state.mask = components.mask_processor.preprocess(block_state.mask_image, height=block_state.height, width=block_state.width, resize_mode=block_state.resize_mode, crops_coords=block_state.crops_coords)
+        block_state.masked_image = block_state.image * (block_state.mask < 0.5)
+
+        block_state.batch_size = block_state.image.shape[0]
+        block_state.image = block_state.image.to(device=block_state.device, dtype=block_state.dtype)
+        block_state.image_latents = self._encode_vae_image(components, image=block_state.image, generator=block_state.generator)
+
+        # 7. Prepare mask latent variables
+        block_state.mask, block_state.masked_image_latents = self.prepare_mask_latents(
+            components,
+            block_state.mask,
+            block_state.masked_image,
+            block_state.batch_size,
+            block_state.height,
+            block_state.width,
+            block_state.dtype,
+            block_state.device,
+            block_state.generator,
+        )
+
+        self.add_block_state(state, block_state)
+
+
+        return components, state
+
+
+
+# auto blocks (YiYi TODO: maybe move all the auto blocks to a separate file)
+# Encode
+class StableDiffusionXLAutoVaeEncoderStep(AutoPipelineBlocks): 
+    block_classes = [StableDiffusionXLInpaintVaeEncoderStep, StableDiffusionXLVaeEncoderStep]
+    block_names = ["inpaint", "img2img"]
+    block_trigger_inputs = ["mask_image", "image"]
+
+    @property
+    def description(self):
+        return "Vae encoder step that encode the image inputs into their latent representations.\n" + \
+               "This is an auto pipeline block that works for both inpainting and img2img tasks.\n" + \
+               " - `StableDiffusionXLInpaintVaeEncoderStep` (inpaint) is used when both `mask_image` and `image` are provided.\n" + \
+               " - `StableDiffusionXLVaeEncoderStep` (img2img) is used when only `image` is provided."
+
+
+class StableDiffusionXLAutoIPAdapterStep(AutoPipelineBlocks, ModularIPAdapterMixin):
+    block_classes = [StableDiffusionXLIPAdapterStep]
+    block_names = ["ip_adapter"]
+    block_trigger_inputs = ["ip_adapter_image"]
+
+    @property
+    def description(self):
+        return "Run IP Adapter step if `ip_adapter_image` is provided."
+

src/diffusers/modular_pipelines/stable_diffusion_xl/modular_loader.py

@@ -0,0 +1,174 @@
+# Copyright 2024 The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from typing import Any, List, Optional, Tuple, Union, Dict
+import PIL
+import torch
+import numpy as np
+
+from ...loaders import StableDiffusionXLLoraLoaderMixin, TextualInversionLoaderMixin, ModularIPAdapterMixin
+from ...image_processor import PipelineImageInput
+from ...pipelines.pipeline_utils import StableDiffusionMixin
+from ...pipelines.stable_diffusion_xl.pipeline_output import StableDiffusionXLPipelineOutput
+from ...utils import logging
+
+from ..modular_pipeline import ModularLoader
+from ..modular_pipeline_utils import InputParam, OutputParam
+
+logger = logging.get_logger(__name__)  # pylint: disable=invalid-name
+
+
+
+# YiYi TODO: move to a different file? stable_diffusion_xl_module should have its own folder?
+# YiYi Notes: model specific components:
+## (1) it should inherit from ModularLoader
+## (2) acts like a container that holds components and configs
+## (3) define default config (related to components), e.g. default_sample_size, vae_scale_factor, num_channels_unet, num_channels_latents
+## (4) inherit from model-specic loader class (e.g. StableDiffusionXLLoraLoaderMixin)
+## (5) how to use together with Components_manager?
+class StableDiffusionXLModularLoader(
+    ModularLoader,
+    StableDiffusionMixin,
+    TextualInversionLoaderMixin,
+    StableDiffusionXLLoraLoaderMixin,
+    ModularIPAdapterMixin,
+):
+    @property
+    def default_sample_size(self):
+        default_sample_size = 128
+        if hasattr(self, "unet") and self.unet is not None:
+            default_sample_size = self.unet.config.sample_size
+        return default_sample_size
+
+    @property
+    def vae_scale_factor(self):
+        vae_scale_factor = 8
+        if hasattr(self, "vae") and self.vae is not None:
+            vae_scale_factor = 2 ** (len(self.vae.config.block_out_channels) - 1)
+        return vae_scale_factor
+
+    @property
+    def num_channels_unet(self):
+        num_channels_unet = 4
+        if hasattr(self, "unet") and self.unet is not None:
+            num_channels_unet = self.unet.config.in_channels
+        return num_channels_unet
+
+    @property
+    def num_channels_latents(self):
+        num_channels_latents = 4
+        if hasattr(self, "vae") and self.vae is not None:
+            num_channels_latents = self.vae.config.latent_channels
+        return num_channels_latents
+
+
+
+# YiYi Notes: not used yet, maintain a list of schema that can be used across all pipeline blocks
+SDXL_INPUTS_SCHEMA = {
+    "prompt": InputParam("prompt", type_hint=Union[str, List[str]], description="The prompt or prompts to guide the image generation"),
+    "prompt_2": InputParam("prompt_2", type_hint=Union[str, List[str]], description="The prompt or prompts to be sent to the tokenizer_2 and text_encoder_2"),
+    "negative_prompt": InputParam("negative_prompt", type_hint=Union[str, List[str]], description="The prompt or prompts not to guide the image generation"),
+    "negative_prompt_2": InputParam("negative_prompt_2", type_hint=Union[str, List[str]], description="The negative prompt or prompts for text_encoder_2"),
+    "cross_attention_kwargs": InputParam("cross_attention_kwargs", type_hint=Optional[dict], description="Kwargs dictionary passed to the AttentionProcessor"),
+    "clip_skip": InputParam("clip_skip", type_hint=Optional[int], description="Number of layers to skip in CLIP text encoder"),
+    "image": InputParam("image", type_hint=PipelineImageInput, required=True, description="The image(s) to modify for img2img or inpainting"),
+    "mask_image": InputParam("mask_image", type_hint=PipelineImageInput, required=True, description="Mask image for inpainting, white pixels will be repainted"),
+    "generator": InputParam("generator", type_hint=Optional[Union[torch.Generator, List[torch.Generator]]], description="Generator(s) for deterministic generation"),
+    "height": InputParam("height", type_hint=Optional[int], description="Height in pixels of the generated image"),
+    "width": InputParam("width", type_hint=Optional[int], description="Width in pixels of the generated image"),
+    "num_images_per_prompt": InputParam("num_images_per_prompt", type_hint=int, default=1, description="Number of images to generate per prompt"),
+    "num_inference_steps": InputParam("num_inference_steps", type_hint=int, default=50, description="Number of denoising steps"),
+    "timesteps": InputParam("timesteps", type_hint=Optional[torch.Tensor], description="Custom timesteps for the denoising process"),
+    "sigmas": InputParam("sigmas", type_hint=Optional[torch.Tensor], description="Custom sigmas for the denoising process"),
+    "denoising_end": InputParam("denoising_end", type_hint=Optional[float], description="Fraction of denoising process to complete before termination"),
+    # YiYi Notes: img2img defaults to 0.3, inpainting defaults to 0.9999
+    "strength": InputParam("strength", type_hint=float, default=0.3, description="How much to transform the reference image"),
+    "denoising_start": InputParam("denoising_start", type_hint=Optional[float], description="Starting point of the denoising process"),
+    "latents": InputParam("latents", type_hint=Optional[torch.Tensor], description="Pre-generated noisy latents for image generation"),
+    "padding_mask_crop": InputParam("padding_mask_crop", type_hint=Optional[Tuple[int, int]], description="Size of margin in crop for image and mask"),
+    "original_size": InputParam("original_size", type_hint=Optional[Tuple[int, int]], description="Original size of the image for SDXL's micro-conditioning"),
+    "target_size": InputParam("target_size", type_hint=Optional[Tuple[int, int]], description="Target size for SDXL's micro-conditioning"),
+    "negative_original_size": InputParam("negative_original_size", type_hint=Optional[Tuple[int, int]], description="Negative conditioning based on image resolution"),
+    "negative_target_size": InputParam("negative_target_size", type_hint=Optional[Tuple[int, int]], description="Negative conditioning based on target resolution"),
+    "crops_coords_top_left": InputParam("crops_coords_top_left", type_hint=Tuple[int, int], default=(0, 0), description="Top-left coordinates for SDXL's micro-conditioning"),
+    "negative_crops_coords_top_left": InputParam("negative_crops_coords_top_left", type_hint=Tuple[int, int], default=(0, 0), description="Negative conditioning crop coordinates"),
+    "aesthetic_score": InputParam("aesthetic_score", type_hint=float, default=6.0, description="Simulates aesthetic score of generated image"),
+    "negative_aesthetic_score": InputParam("negative_aesthetic_score", type_hint=float, default=2.0, description="Simulates negative aesthetic score"),
+    "eta": InputParam("eta", type_hint=float, default=0.0, description="Parameter η in the DDIM paper"),
+    "output_type": InputParam("output_type", type_hint=str, default="pil", description="Output format (pil/tensor/np.array)"),
+    "ip_adapter_image": InputParam("ip_adapter_image", type_hint=PipelineImageInput, required=True, description="Image(s) to be used as IP adapter"),
+    "control_image": InputParam("control_image", type_hint=PipelineImageInput, required=True, description="ControlNet input condition"),
+    "control_guidance_start": InputParam("control_guidance_start", type_hint=Union[float, List[float]], default=0.0, description="When ControlNet starts applying"),
+    "control_guidance_end": InputParam("control_guidance_end", type_hint=Union[float, List[float]], default=1.0, description="When ControlNet stops applying"),
+    "controlnet_conditioning_scale": InputParam("controlnet_conditioning_scale", type_hint=Union[float, List[float]], default=1.0, description="Scale factor for ControlNet outputs"),
+    "guess_mode": InputParam("guess_mode", type_hint=bool, default=False, description="Enables ControlNet encoder to recognize input without prompts"),
+    "control_mode": InputParam("control_mode", type_hint=List[int], required=True, description="Control mode for union controlnet")
+}
+
+
+SDXL_INTERMEDIATE_INPUTS_SCHEMA = {
+    "prompt_embeds": InputParam("prompt_embeds", type_hint=torch.Tensor, required=True, description="Text embeddings used to guide image generation"),
+    "negative_prompt_embeds": InputParam("negative_prompt_embeds", type_hint=torch.Tensor, description="Negative text embeddings"),
+    "pooled_prompt_embeds": InputParam("pooled_prompt_embeds", type_hint=torch.Tensor, required=True, description="Pooled text embeddings"),
+    "negative_pooled_prompt_embeds": InputParam("negative_pooled_prompt_embeds", type_hint=torch.Tensor, description="Negative pooled text embeddings"),
+    "batch_size": InputParam("batch_size", type_hint=int, required=True, description="Number of prompts"),
+    "dtype": InputParam("dtype", type_hint=torch.dtype, description="Data type of model tensor inputs"),
+    "preprocess_kwargs": InputParam("preprocess_kwargs", type_hint=Optional[dict], description="Kwargs for ImageProcessor"),
+    "latents": InputParam("latents", type_hint=torch.Tensor, required=True, description="Initial latents for denoising process"),
+    "timesteps": InputParam("timesteps", type_hint=torch.Tensor, required=True, description="Timesteps for inference"),
+    "num_inference_steps": InputParam("num_inference_steps", type_hint=int, required=True, description="Number of denoising steps"),
+    "latent_timestep": InputParam("latent_timestep", type_hint=torch.Tensor, required=True, description="Initial noise level timestep"),
+    "image_latents": InputParam("image_latents", type_hint=torch.Tensor, required=True, description="Latents representing reference image"),
+    "mask": InputParam("mask", type_hint=torch.Tensor, required=True, description="Mask for inpainting"),
+    "masked_image_latents": InputParam("masked_image_latents", type_hint=torch.Tensor, description="Masked image latents for inpainting"),
+    "add_time_ids": InputParam("add_time_ids", type_hint=torch.Tensor, required=True, description="Time ids for conditioning"),
+    "negative_add_time_ids": InputParam("negative_add_time_ids", type_hint=torch.Tensor, description="Negative time ids"),
+    "timestep_cond": InputParam("timestep_cond", type_hint=torch.Tensor, description="Timestep conditioning for LCM"),
+    "noise": InputParam("noise", type_hint=torch.Tensor, description="Noise added to image latents"),
+    "crops_coords": InputParam("crops_coords", type_hint=Optional[Tuple[int]], description="Crop coordinates"),
+    "ip_adapter_embeds": InputParam("ip_adapter_embeds", type_hint=List[torch.Tensor], description="Image embeddings for IP-Adapter"),
+    "negative_ip_adapter_embeds": InputParam("negative_ip_adapter_embeds", type_hint=List[torch.Tensor], description="Negative image embeddings for IP-Adapter"),
+    "images": InputParam("images", type_hint=Union[List[PIL.Image.Image], List[torch.Tensor], List[np.array]], required=True, description="Generated images")
+}
+
+
+SDXL_INTERMEDIATE_OUTPUTS_SCHEMA = {
+    "prompt_embeds": OutputParam("prompt_embeds", type_hint=torch.Tensor, description="Text embeddings used to guide image generation"),
+    "negative_prompt_embeds": OutputParam("negative_prompt_embeds", type_hint=torch.Tensor, description="Negative text embeddings"),
+    "pooled_prompt_embeds": OutputParam("pooled_prompt_embeds", type_hint=torch.Tensor, description="Pooled text embeddings"),
+    "negative_pooled_prompt_embeds": OutputParam("negative_pooled_prompt_embeds", type_hint=torch.Tensor, description="Negative pooled text embeddings"),
+    "batch_size": OutputParam("batch_size", type_hint=int, description="Number of prompts"),
+    "dtype": OutputParam("dtype", type_hint=torch.dtype, description="Data type of model tensor inputs"),
+    "image_latents": OutputParam("image_latents", type_hint=torch.Tensor, description="Latents representing reference image"),
+    "mask": OutputParam("mask", type_hint=torch.Tensor, description="Mask for inpainting"),
+    "masked_image_latents": OutputParam("masked_image_latents", type_hint=torch.Tensor, description="Masked image latents for inpainting"),
+    "crops_coords": OutputParam("crops_coords", type_hint=Optional[Tuple[int]], description="Crop coordinates"),
+    "timesteps": OutputParam("timesteps", type_hint=torch.Tensor, description="Timesteps for inference"),
+    "num_inference_steps": OutputParam("num_inference_steps", type_hint=int, description="Number of denoising steps"),
+    "latent_timestep": OutputParam("latent_timestep", type_hint=torch.Tensor, description="Initial noise level timestep"),
+    "add_time_ids": OutputParam("add_time_ids", type_hint=torch.Tensor, description="Time ids for conditioning"),
+    "negative_add_time_ids": OutputParam("negative_add_time_ids", type_hint=torch.Tensor, description="Negative time ids"),
+    "timestep_cond": OutputParam("timestep_cond", type_hint=torch.Tensor, description="Timestep conditioning for LCM"),
+    "latents": OutputParam("latents", type_hint=torch.Tensor, description="Denoised latents"),
+    "noise": OutputParam("noise", type_hint=torch.Tensor, description="Noise added to image latents"),
+    "ip_adapter_embeds": OutputParam("ip_adapter_embeds", type_hint=List[torch.Tensor], description="Image embeddings for IP-Adapter"),
+    "negative_ip_adapter_embeds": OutputParam("negative_ip_adapter_embeds", type_hint=List[torch.Tensor], description="Negative image embeddings for IP-Adapter"),
+    "images": OutputParam("images", type_hint=Union[List[PIL.Image.Image], List[torch.Tensor], List[np.array]], description="Generated images")
+}
+
+
+SDXL_OUTPUTS_SCHEMA = {
+    "images": OutputParam("images", type_hint=Union[Tuple[Union[List[PIL.Image.Image], List[torch.Tensor], List[np.array]]], StableDiffusionXLPipelineOutput], description="The final generated images")
+}
+

src/diffusers/modular_pipelines/stable_diffusion_xl/modular_pipeline_block_mappings.py

@@ -0,0 +1,128 @@
+# Copyright 2024 The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from collections import OrderedDict
+
+# Import all the necessary block classes
+from .denoise import (
+    StableDiffusionXLAutoDenoiseStep,
+    StableDiffusionXLDenoiseStep,
+    StableDiffusionXLControlNetDenoiseStep
+)
+from .before_denoise import (
+    StableDiffusionXLAutoBeforeDenoiseStep,
+    StableDiffusionXLInputStep,
+    StableDiffusionXLSetTimestepsStep,
+    StableDiffusionXLPrepareLatentsStep,
+    StableDiffusionXLPrepareAdditionalConditioningStep,
+    StableDiffusionXLImg2ImgSetTimestepsStep,
+    StableDiffusionXLImg2ImgPrepareLatentsStep,
+    StableDiffusionXLImg2ImgPrepareAdditionalConditioningStep,
+    StableDiffusionXLInpaintPrepareLatentsStep,
+    StableDiffusionXLControlNetInputStep,
+    StableDiffusionXLControlNetUnionInputStep
+)
+from .encoders import (
+    StableDiffusionXLTextEncoderStep,
+    StableDiffusionXLAutoIPAdapterStep,
+    StableDiffusionXLAutoVaeEncoderStep,
+    StableDiffusionXLVaeEncoderStep,
+    StableDiffusionXLInpaintVaeEncoderStep,
+    StableDiffusionXLIPAdapterStep
+)
+from .decoders import (
+    StableDiffusionXLDecodeStep,
+    StableDiffusionXLInpaintDecodeStep,
+    StableDiffusionXLAutoDecodeStep
+)
+
+
+# YiYi notes: comment out for now, work on this later
+# block mapping 
+TEXT2IMAGE_BLOCKS = OrderedDict([
+    ("text_encoder", StableDiffusionXLTextEncoderStep),
+    ("ip_adapter", StableDiffusionXLAutoIPAdapterStep),
+    ("input", StableDiffusionXLInputStep),
+    ("set_timesteps", StableDiffusionXLSetTimestepsStep),
+    ("prepare_latents", StableDiffusionXLPrepareLatentsStep),
+    ("prepare_add_cond", StableDiffusionXLPrepareAdditionalConditioningStep),
+    ("denoise", StableDiffusionXLDenoiseStep),
+    ("decode", StableDiffusionXLDecodeStep)
+])
+
+IMAGE2IMAGE_BLOCKS = OrderedDict([
+    ("text_encoder", StableDiffusionXLTextEncoderStep),
+    ("ip_adapter", StableDiffusionXLAutoIPAdapterStep),
+    ("image_encoder", StableDiffusionXLVaeEncoderStep),
+    ("input", StableDiffusionXLInputStep),
+    ("set_timesteps", StableDiffusionXLImg2ImgSetTimestepsStep),
+    ("prepare_latents", StableDiffusionXLImg2ImgPrepareLatentsStep),
+    ("prepare_add_cond", StableDiffusionXLImg2ImgPrepareAdditionalConditioningStep),
+    ("denoise", StableDiffusionXLDenoiseStep),
+    ("decode", StableDiffusionXLDecodeStep)
+])
+
+INPAINT_BLOCKS = OrderedDict([
+    ("text_encoder", StableDiffusionXLTextEncoderStep),
+    ("ip_adapter", StableDiffusionXLAutoIPAdapterStep),
+    ("image_encoder", StableDiffusionXLInpaintVaeEncoderStep),
+    ("input", StableDiffusionXLInputStep),
+    ("set_timesteps", StableDiffusionXLImg2ImgSetTimestepsStep),
+    ("prepare_latents", StableDiffusionXLInpaintPrepareLatentsStep),
+    ("prepare_add_cond", StableDiffusionXLImg2ImgPrepareAdditionalConditioningStep),
+    ("denoise", StableDiffusionXLDenoiseStep),
+    ("decode", StableDiffusionXLInpaintDecodeStep)
+])
+
+CONTROLNET_BLOCKS = OrderedDict([
+    ("controlnet_input", StableDiffusionXLControlNetInputStep),
+    ("denoise", StableDiffusionXLControlNetDenoiseStep),
+])
+
+CONTROLNET_UNION_BLOCKS = OrderedDict([
+    ("controlnet_input", StableDiffusionXLControlNetUnionInputStep),
+    ("denoise", StableDiffusionXLControlNetDenoiseStep),
+])
+
+IP_ADAPTER_BLOCKS = OrderedDict([
+    ("ip_adapter", StableDiffusionXLIPAdapterStep),
+])
+
+AUTO_BLOCKS = OrderedDict([
+    ("text_encoder", StableDiffusionXLTextEncoderStep),
+    ("ip_adapter", StableDiffusionXLAutoIPAdapterStep),
+    ("image_encoder", StableDiffusionXLAutoVaeEncoderStep),
+    ("before_denoise", StableDiffusionXLAutoBeforeDenoiseStep),
+    ("denoise", StableDiffusionXLAutoDenoiseStep),
+    ("decode", StableDiffusionXLAutoDecodeStep)
+])
+
+AUTO_CORE_BLOCKS = OrderedDict([
+    ("before_denoise", StableDiffusionXLAutoBeforeDenoiseStep),
+    ("denoise", StableDiffusionXLAutoDenoiseStep),
+])
+
+
+SDXL_SUPPORTED_BLOCKS = {
+    "text2img": TEXT2IMAGE_BLOCKS,
+    "img2img": IMAGE2IMAGE_BLOCKS,
+    "inpaint": INPAINT_BLOCKS,
+    "controlnet": CONTROLNET_BLOCKS,
+    "controlnet_union": CONTROLNET_UNION_BLOCKS,
+    "ip_adapter": IP_ADAPTER_BLOCKS,
+    "auto": AUTO_BLOCKS
+}
+
+
+

src/diffusers/modular_pipelines/stable_diffusion_xl/modular_pipeline_presets.py

@@ -0,0 +1,43 @@
+# Copyright 2024 The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from typing import Any, List, Optional, Tuple, Union, Dict
+from ...utils import logging
+from ..modular_pipeline import SequentialPipelineBlocks
+
+from .denoise import StableDiffusionXLAutoDenoiseStep
+from .before_denoise import StableDiffusionXLAutoBeforeDenoiseStep
+from .decoders import StableDiffusionXLAutoDecodeStep
+from .encoders import StableDiffusionXLTextEncoderStep, StableDiffusionXLAutoIPAdapterStep, StableDiffusionXLAutoVaeEncoderStep
+
+logger = logging.get_logger(__name__)  # pylint: disable=invalid-name
+
+
+class StableDiffusionXLAutoPipeline(SequentialPipelineBlocks):
+    block_classes = [StableDiffusionXLTextEncoderStep, StableDiffusionXLAutoIPAdapterStep, StableDiffusionXLAutoVaeEncoderStep, StableDiffusionXLAutoBeforeDenoiseStep, StableDiffusionXLAutoDenoiseStep, StableDiffusionXLAutoDecodeStep]
+    block_names = ["text_encoder", "ip_adapter", "image_encoder", "before_denoise", "denoise", "decoder"]
+
+    @property
+    def description(self):
+        return "Auto Modular pipeline for text-to-image, image-to-image, inpainting, and controlnet tasks using Stable Diffusion XL.\n" + \
+               "- for image-to-image generation, you need to provide either `image` or `image_latents`\n" + \
+               "- for inpainting, you need to provide `mask_image` and `image`, optionally you can provide `padding_mask_crop` \n" + \
+               "- to run the controlnet workflow, you need to provide `control_image`\n" + \
+               "- to run the controlnet_union workflow, you need to provide `control_image` and `control_mode`\n" + \
+               "- to run the ip_adapter workflow, you need to provide `ip_adapter_image`\n" + \
+               "- for text-to-image generation, all you need to provide is `prompt`"
+
+
+
+

src/diffusers/pipelines/auto_pipeline.py

@@ -246,14 +246,15 @@ def _get_connected_pipeline(pipeline_cls):
         return _get_task_class(AUTO_INPAINT_PIPELINES_MAPPING, pipeline_cls.__name__, throw_error_if_not_exist=False)
 
 
-def _get_task_class(mapping, pipeline_class_name, throw_error_if_not_exist: bool = True):
-    def get_model(pipeline_class_name):
-        for task_mapping in SUPPORTED_TASKS_MAPPINGS:
-            for model_name, pipeline in task_mapping.items():
-                if pipeline.__name__ == pipeline_class_name:
-                    return model_name
+def _get_model(pipeline_class_name):
+    for task_mapping in SUPPORTED_TASKS_MAPPINGS:
+        for model_name, pipeline in task_mapping.items():
+            if pipeline.__name__ == pipeline_class_name:
+                return model_name
 
-    model_name = get_model(pipeline_class_name)
+
+def _get_task_class(mapping, pipeline_class_name, throw_error_if_not_exist: bool = True):
+    model_name = _get_model(pipeline_class_name)
 
     if model_name is not None:
         task_class = mapping.get(model_name, None)

src/diffusers/pipelines/controlnet/pipeline_controlnet_sd_xl_img2img.py

@@ -912,12 +912,6 @@ class StableDiffusionXLControlNetImg2ImgPipeline(
                 f"`image` has to be of type `torch.Tensor`, `PIL.Image.Image` or list but is {type(image)}"
             )
 
-        latents_mean = latents_std = None
-        if hasattr(self.vae.config, "latents_mean") and self.vae.config.latents_mean is not None:
-            latents_mean = torch.tensor(self.vae.config.latents_mean).view(1, 4, 1, 1)
-        if hasattr(self.vae.config, "latents_std") and self.vae.config.latents_std is not None:
-            latents_std = torch.tensor(self.vae.config.latents_std).view(1, 4, 1, 1)
-
         # Offload text encoder if `enable_model_cpu_offload` was enabled
         if hasattr(self, "final_offload_hook") and self.final_offload_hook is not None:
             self.text_encoder_2.to("cpu")
@@ -931,6 +925,11 @@ class StableDiffusionXLControlNetImg2ImgPipeline(
             init_latents = image
 
         else:
+            latents_mean = latents_std = None
+            if hasattr(self.vae.config, "latents_mean") and self.vae.config.latents_mean is not None:
+                latents_mean = torch.tensor(self.vae.config.latents_mean).view(1, 4, 1, 1)
+            if hasattr(self.vae.config, "latents_std") and self.vae.config.latents_std is not None:
+                latents_std = torch.tensor(self.vae.config.latents_std).view(1, 4, 1, 1)
             # make sure the VAE is in float32 mode, as it overflows in float16
             if self.vae.config.force_upcast:
                 image = image.float()

src/diffusers/pipelines/controlnet/pipeline_controlnet_union_sd_xl_img2img.py

@@ -867,12 +867,6 @@ class StableDiffusionXLControlNetUnionImg2ImgPipeline(
                 f"`image` has to be of type `torch.Tensor`, `PIL.Image.Image` or list but is {type(image)}"
             )
 
-        latents_mean = latents_std = None
-        if hasattr(self.vae.config, "latents_mean") and self.vae.config.latents_mean is not None:
-            latents_mean = torch.tensor(self.vae.config.latents_mean).view(1, 4, 1, 1)
-        if hasattr(self.vae.config, "latents_std") and self.vae.config.latents_std is not None:
-            latents_std = torch.tensor(self.vae.config.latents_std).view(1, 4, 1, 1)
-
         # Offload text encoder if `enable_model_cpu_offload` was enabled
         if hasattr(self, "final_offload_hook") and self.final_offload_hook is not None:
             self.text_encoder_2.to("cpu")
@@ -886,6 +880,11 @@ class StableDiffusionXLControlNetUnionImg2ImgPipeline(
             init_latents = image
 
         else:
+            latents_mean = latents_std = None
+            if hasattr(self.vae.config, "latents_mean") and self.vae.config.latents_mean is not None:
+                latents_mean = torch.tensor(self.vae.config.latents_mean).view(1, 4, 1, 1)
+            if hasattr(self.vae.config, "latents_std") and self.vae.config.latents_std is not None:
+                latents_std = torch.tensor(self.vae.config.latents_std).view(1, 4, 1, 1)
             # make sure the VAE is in float32 mode, as it overflows in float16
             if self.vae.config.force_upcast:
                 image = image.float()

src/diffusers/pipelines/kolors/pipeline_kolors_img2img.py

@@ -609,12 +609,6 @@ class KolorsImg2ImgPipeline(DiffusionPipeline, StableDiffusionMixin, StableDiffu
                 f"`image` has to be of type `torch.Tensor`, `PIL.Image.Image` or list but is {type(image)}"
             )
 
-        latents_mean = latents_std = None
-        if hasattr(self.vae.config, "latents_mean") and self.vae.config.latents_mean is not None:
-            latents_mean = torch.tensor(self.vae.config.latents_mean).view(1, 4, 1, 1)
-        if hasattr(self.vae.config, "latents_std") and self.vae.config.latents_std is not None:
-            latents_std = torch.tensor(self.vae.config.latents_std).view(1, 4, 1, 1)
-
         # Offload text encoder if `enable_model_cpu_offload` was enabled
         if hasattr(self, "final_offload_hook") and self.final_offload_hook is not None:
             self.text_encoder_2.to("cpu")
@@ -628,6 +622,11 @@ class KolorsImg2ImgPipeline(DiffusionPipeline, StableDiffusionMixin, StableDiffu
             init_latents = image
 
         else:
+            latents_mean = latents_std = None
+            if hasattr(self.vae.config, "latents_mean") and self.vae.config.latents_mean is not None:
+                latents_mean = torch.tensor(self.vae.config.latents_mean).view(1, 4, 1, 1)
+            if hasattr(self.vae.config, "latents_std") and self.vae.config.latents_std is not None:
+                latents_std = torch.tensor(self.vae.config.latents_std).view(1, 4, 1, 1)
             # make sure the VAE is in float32 mode, as it overflows in float16
             if self.vae.config.force_upcast:
                 image = image.float()

src/diffusers/pipelines/pag/pipeline_pag_controlnet_sd_xl_img2img.py

@@ -917,12 +917,6 @@ class StableDiffusionXLControlNetPAGImg2ImgPipeline(
                 f"`image` has to be of type `torch.Tensor`, `PIL.Image.Image` or list but is {type(image)}"
             )
 
-        latents_mean = latents_std = None
-        if hasattr(self.vae.config, "latents_mean") and self.vae.config.latents_mean is not None:
-            latents_mean = torch.tensor(self.vae.config.latents_mean).view(1, 4, 1, 1)
-        if hasattr(self.vae.config, "latents_std") and self.vae.config.latents_std is not None:
-            latents_std = torch.tensor(self.vae.config.latents_std).view(1, 4, 1, 1)
-
         # Offload text encoder if `enable_model_cpu_offload` was enabled
         if hasattr(self, "final_offload_hook") and self.final_offload_hook is not None:
             self.text_encoder_2.to("cpu")
@@ -936,6 +930,11 @@ class StableDiffusionXLControlNetPAGImg2ImgPipeline(
             init_latents = image
 
         else:
+            latents_mean = latents_std = None
+            if hasattr(self.vae.config, "latents_mean") and self.vae.config.latents_mean is not None:
+                latents_mean = torch.tensor(self.vae.config.latents_mean).view(1, 4, 1, 1)
+            if hasattr(self.vae.config, "latents_std") and self.vae.config.latents_std is not None:
+                latents_std = torch.tensor(self.vae.config.latents_std).view(1, 4, 1, 1)
             # make sure the VAE is in float32 mode, as it overflows in float16
             if self.vae.config.force_upcast:
                 image = image.float()

src/diffusers/pipelines/pag/pipeline_pag_sd_xl_img2img.py

@@ -707,12 +707,6 @@ class StableDiffusionXLPAGImg2ImgPipeline(
                 f"`image` has to be of type `torch.Tensor`, `PIL.Image.Image` or list but is {type(image)}"
             )
 
-        latents_mean = latents_std = None
-        if hasattr(self.vae.config, "latents_mean") and self.vae.config.latents_mean is not None:
-            latents_mean = torch.tensor(self.vae.config.latents_mean).view(1, 4, 1, 1)
-        if hasattr(self.vae.config, "latents_std") and self.vae.config.latents_std is not None:
-            latents_std = torch.tensor(self.vae.config.latents_std).view(1, 4, 1, 1)
-
         # Offload text encoder if `enable_model_cpu_offload` was enabled
         if hasattr(self, "final_offload_hook") and self.final_offload_hook is not None:
             self.text_encoder_2.to("cpu")
@@ -726,6 +720,11 @@ class StableDiffusionXLPAGImg2ImgPipeline(
             init_latents = image
 
         else:
+            latents_mean = latents_std = None
+            if hasattr(self.vae.config, "latents_mean") and self.vae.config.latents_mean is not None:
+                latents_mean = torch.tensor(self.vae.config.latents_mean).view(1, 4, 1, 1)
+            if hasattr(self.vae.config, "latents_std") and self.vae.config.latents_std is not None:
+                latents_std = torch.tensor(self.vae.config.latents_std).view(1, 4, 1, 1)
             # make sure the VAE is in float32 mode, as it overflows in float16
             if self.vae.config.force_upcast:
                 image = image.float()

src/diffusers/pipelines/pipeline_loading_utils.py

@@ -331,6 +331,20 @@ def maybe_raise_or_warn(
         )
 
 
+# a simpler version of get_class_obj_and_candidates, it won't work with custom code
+def simple_get_class_obj(library_name, class_name):
+    from diffusers import pipelines
+    is_pipeline_module = hasattr(pipelines, library_name)
+
+    if is_pipeline_module:
+        pipeline_module = getattr(pipelines, library_name)
+        class_obj = getattr(pipeline_module, class_name)
+    else:
+        library = importlib.import_module(library_name)
+        class_obj = getattr(library, class_name)
+
+    return class_obj
+
 def get_class_obj_and_candidates(
     library_name, class_name, importable_classes, pipelines, is_pipeline_module, component_name=None, cache_dir=None
 ):
@@ -839,7 +853,10 @@ def _fetch_class_library_tuple(module):
         library = not_compiled_module.__module__
 
     # retrieve class_name
-    class_name = not_compiled_module.__class__.__name__
+    if isinstance(not_compiled_module, type):
+        class_name = not_compiled_module.__name__
+    else:
+        class_name = not_compiled_module.__class__.__name__
 
     return (library, class_name)
 

src/diffusers/pipelines/pipeline_utils.py

@@ -427,7 +427,7 @@ class DiffusionPipeline(ConfigMixin, PushToHubMixin):
             module_is_sequentially_offloaded(module) for _, module in self.components.items()
         )
 
-        is_pipeline_device_mapped = self.hf_device_map is not None and len(self.hf_device_map) > 1
+        is_pipeline_device_mapped = hasattr(self, "hf_device_map") and self.hf_device_map is not None and len(self.hf_device_map) > 1
         if is_pipeline_device_mapped:
             raise ValueError(
                 "It seems like you have activated a device mapping strategy on the pipeline which doesn't allow explicit device placement using `to()`. You can call `reset_device_map()` to remove the existing device map from the pipeline."
@@ -444,6 +444,7 @@ class DiffusionPipeline(ConfigMixin, PushToHubMixin):
                     "You are trying to call `.to('cuda')` on a pipeline that has models quantized with `bitsandbytes`. Your current `accelerate` installation does not support it. Please upgrade the installation."
                 )
 
+
         # Display a warning in this case (the operation succeeds but the benefits are lost)
         pipeline_is_offloaded = any(module_is_offloaded(module) for _, module in self.components.items())
         if pipeline_is_offloaded and device_type in ["cuda", "xpu"]:
@@ -1119,9 +1120,11 @@ class DiffusionPipeline(ConfigMixin, PushToHubMixin):
                 The PyTorch device type of the accelerator that shall be used in inference. If not specified, it will
                 automatically detect the available accelerator and use.
         """
+ 
         self._maybe_raise_error_if_group_offload_active(raise_error=True)
 
-        is_pipeline_device_mapped = self.hf_device_map is not None and len(self.hf_device_map) > 1
+        is_pipeline_device_mapped = hasattr(self, "hf_device_map") and self.hf_device_map is not None and len(self.hf_device_map) > 1
+
         if is_pipeline_device_mapped:
             raise ValueError(
                 "It seems like you have activated a device mapping strategy on the pipeline so calling `enable_model_cpu_offload() isn't allowed. You can call `reset_device_map()` first and then call `enable_model_cpu_offload()`."
@@ -1245,7 +1248,7 @@ class DiffusionPipeline(ConfigMixin, PushToHubMixin):
             raise ImportError("`enable_sequential_cpu_offload` requires `accelerate v0.14.0` or higher")
         self.remove_all_hooks()
 
-        is_pipeline_device_mapped = self.hf_device_map is not None and len(self.hf_device_map) > 1
+        is_pipeline_device_mapped = hasattr(self, "hf_device_map") and self.hf_device_map is not None and len(self.hf_device_map) > 1
         if is_pipeline_device_mapped:
             raise ValueError(
                 "It seems like you have activated a device mapping strategy on the pipeline so calling `enable_sequential_cpu_offload() isn't allowed. You can call `reset_device_map()` first and then call `enable_sequential_cpu_offload()`."
@@ -1945,9 +1948,10 @@ class DiffusionPipeline(ConfigMixin, PushToHubMixin):
             f"{'' if k.startswith('_') else '_'}{k}": v for k, v in original_config.items() if k not in pipeline_kwargs
         }
 
+        optional_components = pipeline._optional_components if hasattr(pipeline, "_optional_components") and pipeline._optional_components else []
         missing_modules = (
             set(expected_modules)
-            - set(pipeline._optional_components)
+            - set(optional_components)
             - set(pipeline_kwargs.keys())
             - set(true_optional_modules)
         )

src/diffusers/pipelines/stable_diffusion_xl/pipeline_stable_diffusion_xl_img2img.py

@@ -695,12 +695,6 @@ class StableDiffusionXLImg2ImgPipeline(
                 f"`image` has to be of type `torch.Tensor`, `PIL.Image.Image` or list but is {type(image)}"
             )
 
-        latents_mean = latents_std = None
-        if hasattr(self.vae.config, "latents_mean") and self.vae.config.latents_mean is not None:
-            latents_mean = torch.tensor(self.vae.config.latents_mean).view(1, 4, 1, 1)
-        if hasattr(self.vae.config, "latents_std") and self.vae.config.latents_std is not None:
-            latents_std = torch.tensor(self.vae.config.latents_std).view(1, 4, 1, 1)
-
         # Offload text encoder if `enable_model_cpu_offload` was enabled
         if hasattr(self, "final_offload_hook") and self.final_offload_hook is not None:
             self.text_encoder_2.to("cpu")
@@ -714,6 +708,11 @@ class StableDiffusionXLImg2ImgPipeline(
             init_latents = image
 
         else:
+            latents_mean = latents_std = None
+            if hasattr(self.vae.config, "latents_mean") and self.vae.config.latents_mean is not None:
+                latents_mean = torch.tensor(self.vae.config.latents_mean).view(1, 4, 1, 1)
+            if hasattr(self.vae.config, "latents_std") and self.vae.config.latents_std is not None:
+                latents_std = torch.tensor(self.vae.config.latents_std).view(1, 4, 1, 1)
             # make sure the VAE is in float32 mode, as it overflows in float16
             if self.vae.config.force_upcast:
                 image = image.float()

src/diffusers/utils/dummy_pt_objects.py

@@ -1388,6 +1388,21 @@ class LDMSuperResolutionPipeline(metaclass=DummyObject):
         requires_backends(cls, ["torch"])
 
 
+class ModularLoader(metaclass=DummyObject):
+    _backends = ["torch"]
+
+    def __init__(self, *args, **kwargs):
+        requires_backends(self, ["torch"])
+
+    @classmethod
+    def from_config(cls, *args, **kwargs):
+        requires_backends(cls, ["torch"])
+
+    @classmethod
+    def from_pretrained(cls, *args, **kwargs):
+        requires_backends(cls, ["torch"])
+
+
 class PNDMPipeline(metaclass=DummyObject):
     _backends = ["torch"]
 

src/diffusers/utils/dummy_torch_and_transformers_objects.py

@@ -2432,6 +2432,21 @@ class StableDiffusionXLInstructPix2PixPipeline(metaclass=DummyObject):
         requires_backends(cls, ["torch", "transformers"])
 
 
+class StableDiffusionXLModularLoader(metaclass=DummyObject):
+    _backends = ["torch", "transformers"]
+
+    def __init__(self, *args, **kwargs):
+        requires_backends(self, ["torch", "transformers"])
+
+    @classmethod
+    def from_config(cls, *args, **kwargs):
+        requires_backends(cls, ["torch", "transformers"])
+
+    @classmethod
+    def from_pretrained(cls, *args, **kwargs):
+        requires_backends(cls, ["torch", "transformers"])
+
+
 class StableDiffusionXLPAGImg2ImgPipeline(metaclass=DummyObject):
     _backends = ["torch", "transformers"]
 

src/diffusers/utils/torch_utils.py

@@ -90,6 +90,11 @@ def is_compiled_module(module) -> bool:
     return isinstance(module, torch._dynamo.eval_frame.OptimizedModule)
 
 
+def unwrap_module(module):
+    """Unwraps a module if it was compiled with torch.compile()"""
+    return module._orig_mod if is_compiled_module(module) else module
+
+
 def fourier_filter(x_in: "torch.Tensor", threshold: int, scale: int) -> "torch.Tensor":
     """Fourier filter as introduced in FreeU (https://arxiv.org/abs/2309.11497).