Merge branch 'main' into pipeline-specific-mixins

src/diffusers/__init__.py

@@ -564,6 +564,7 @@ else:
             "QwenImageEditPlusPipeline",
             "QwenImageImg2ImgPipeline",
             "QwenImageInpaintPipeline",
+            "QwenImageLayeredPipeline",
             "QwenImagePipeline",
             "ReduxImageEncoder",
             "SanaControlNetPipeline",
@@ -1272,6 +1273,7 @@ if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT:
             QwenImageEditPlusPipeline,
             QwenImageImg2ImgPipeline,
             QwenImageInpaintPipeline,
+            QwenImageLayeredPipeline,
             QwenImagePipeline,
             ReduxImageEncoder,
             SanaControlNetPipeline,

src/diffusers/models/autoencoders/autoencoder_kl_qwenimage.py

@@ -394,6 +394,7 @@ class QwenImageEncoder3d(nn.Module):
         attn_scales=[],
         temperal_downsample=[True, True, False],
         dropout=0.0,
+        input_channels=3,
         non_linearity: str = "silu",
     ):
         super().__init__()
@@ -410,7 +411,7 @@ class QwenImageEncoder3d(nn.Module):
         scale = 1.0
 
         # init block
-        self.conv_in = QwenImageCausalConv3d(3, dims[0], 3, padding=1)
+        self.conv_in = QwenImageCausalConv3d(input_channels, dims[0], 3, padding=1)
 
         # downsample blocks
         self.down_blocks = nn.ModuleList([])
@@ -570,6 +571,7 @@ class QwenImageDecoder3d(nn.Module):
         attn_scales=[],
         temperal_upsample=[False, True, True],
         dropout=0.0,
+        input_channels=3,
         non_linearity: str = "silu",
     ):
         super().__init__()
@@ -621,7 +623,7 @@ class QwenImageDecoder3d(nn.Module):
 
         # output blocks
         self.norm_out = QwenImageRMS_norm(out_dim, images=False)
-        self.conv_out = QwenImageCausalConv3d(out_dim, 3, 3, padding=1)
+        self.conv_out = QwenImageCausalConv3d(out_dim, input_channels, 3, padding=1)
 
         self.gradient_checkpointing = False
 
@@ -684,6 +686,7 @@ class AutoencoderKLQwenImage(ModelMixin, AutoencoderMixin, ConfigMixin, FromOrig
         attn_scales: List[float] = [],
         temperal_downsample: List[bool] = [False, True, True],
         dropout: float = 0.0,
+        input_channels: int = 3,
         latents_mean: List[float] = [-0.7571, -0.7089, -0.9113, 0.1075, -0.1745, 0.9653, -0.1517, 1.5508, 0.4134, -0.0715, 0.5517, -0.3632, -0.1922, -0.9497, 0.2503, -0.2921],
         latents_std: List[float] = [2.8184, 1.4541, 2.3275, 2.6558, 1.2196, 1.7708, 2.6052, 2.0743, 3.2687, 2.1526, 2.8652, 1.5579, 1.6382, 1.1253, 2.8251, 1.9160],
     ) -> None:
@@ -695,13 +698,13 @@ class AutoencoderKLQwenImage(ModelMixin, AutoencoderMixin, ConfigMixin, FromOrig
         self.temperal_upsample = temperal_downsample[::-1]
 
         self.encoder = QwenImageEncoder3d(
-            base_dim, z_dim * 2, dim_mult, num_res_blocks, attn_scales, self.temperal_downsample, dropout
+            base_dim, z_dim * 2, dim_mult, num_res_blocks, attn_scales, self.temperal_downsample, dropout, input_channels
         )
         self.quant_conv = QwenImageCausalConv3d(z_dim * 2, z_dim * 2, 1)
         self.post_quant_conv = QwenImageCausalConv3d(z_dim, z_dim, 1)
 
         self.decoder = QwenImageDecoder3d(
-            base_dim, z_dim, dim_mult, num_res_blocks, attn_scales, self.temperal_upsample, dropout
+            base_dim, z_dim, dim_mult, num_res_blocks, attn_scales, self.temperal_upsample, dropout, input_channels
         )
 
         self.spatial_compression_ratio = 2 ** len(self.temperal_downsample)

src/diffusers/models/transformers/transformer_qwenimage.py

@@ -143,17 +143,26 @@ def apply_rotary_emb_qwen(
 
 
 class QwenTimestepProjEmbeddings(nn.Module):
-    def __init__(self, embedding_dim):
+    def __init__(self, embedding_dim, use_additional_t_cond=False):
         super().__init__()
 
         self.time_proj = Timesteps(num_channels=256, flip_sin_to_cos=True, downscale_freq_shift=0, scale=1000)
         self.timestep_embedder = TimestepEmbedding(in_channels=256, time_embed_dim=embedding_dim)
+        self.use_additional_t_cond = use_additional_t_cond
+        if use_additional_t_cond:
+            self.addition_t_embedding = nn.Embedding(2, embedding_dim)
 
-    def forward(self, timestep, hidden_states):
+    def forward(self, timestep, hidden_states, addition_t_cond=None):
         timesteps_proj = self.time_proj(timestep)
         timesteps_emb = self.timestep_embedder(timesteps_proj.to(dtype=hidden_states.dtype))  # (N, D)
 
         conditioning = timesteps_emb
+        if self.use_additional_t_cond:
+            if addition_t_cond is None:
+                raise ValueError("When additional_t_cond is True, addition_t_cond must be provided.")
+            addition_t_emb = self.addition_t_embedding(addition_t_cond)
+            addition_t_emb = addition_t_emb.to(dtype=hidden_states.dtype)
+            conditioning = conditioning + addition_t_emb
 
         return conditioning
 
@@ -259,6 +268,120 @@ class QwenEmbedRope(nn.Module):
         return freqs.clone().contiguous()
 
 
+class QwenEmbedLayer3DRope(nn.Module):
+    def __init__(self, theta: int, axes_dim: List[int], scale_rope=False):
+        super().__init__()
+        self.theta = theta
+        self.axes_dim = axes_dim
+        pos_index = torch.arange(4096)
+        neg_index = torch.arange(4096).flip(0) * -1 - 1
+        self.pos_freqs = torch.cat(
+            [
+                self.rope_params(pos_index, self.axes_dim[0], self.theta),
+                self.rope_params(pos_index, self.axes_dim[1], self.theta),
+                self.rope_params(pos_index, self.axes_dim[2], self.theta),
+            ],
+            dim=1,
+        )
+        self.neg_freqs = torch.cat(
+            [
+                self.rope_params(neg_index, self.axes_dim[0], self.theta),
+                self.rope_params(neg_index, self.axes_dim[1], self.theta),
+                self.rope_params(neg_index, self.axes_dim[2], self.theta),
+            ],
+            dim=1,
+        )
+
+        self.scale_rope = scale_rope
+
+    def rope_params(self, index, dim, theta=10000):
+        """
+        Args:
+            index: [0, 1, 2, 3] 1D Tensor representing the position index of the token
+        """
+        assert dim % 2 == 0
+        freqs = torch.outer(index, 1.0 / torch.pow(theta, torch.arange(0, dim, 2).to(torch.float32).div(dim)))
+        freqs = torch.polar(torch.ones_like(freqs), freqs)
+        return freqs
+
+    def forward(self, video_fhw, txt_seq_lens, device):
+        """
+        Args: video_fhw: [frame, height, width] a list of 3 integers representing the shape of the video Args:
+        txt_length: [bs] a list of 1 integers representing the length of the text
+        """
+        if self.pos_freqs.device != device:
+            self.pos_freqs = self.pos_freqs.to(device)
+            self.neg_freqs = self.neg_freqs.to(device)
+
+        if isinstance(video_fhw, list):
+            video_fhw = video_fhw[0]
+        if not isinstance(video_fhw, list):
+            video_fhw = [video_fhw]
+
+        vid_freqs = []
+        max_vid_index = 0
+        layer_num = len(video_fhw) - 1
+        for idx, fhw in enumerate(video_fhw):
+            frame, height, width = fhw
+            if idx != layer_num:
+                video_freq = self._compute_video_freqs(frame, height, width, idx)
+            else:
+                ### For the condition image, we set the layer index to -1
+                video_freq = self._compute_condition_freqs(frame, height, width)
+            video_freq = video_freq.to(device)
+            vid_freqs.append(video_freq)
+
+            if self.scale_rope:
+                max_vid_index = max(height // 2, width // 2, max_vid_index)
+            else:
+                max_vid_index = max(height, width, max_vid_index)
+
+        max_vid_index = max(max_vid_index, layer_num)
+        max_len = max(txt_seq_lens)
+        txt_freqs = self.pos_freqs[max_vid_index : max_vid_index + max_len, ...]
+        vid_freqs = torch.cat(vid_freqs, dim=0)
+
+        return vid_freqs, txt_freqs
+
+    @functools.lru_cache(maxsize=None)
+    def _compute_video_freqs(self, frame, height, width, idx=0):
+        seq_lens = frame * height * width
+        freqs_pos = self.pos_freqs.split([x // 2 for x in self.axes_dim], dim=1)
+        freqs_neg = self.neg_freqs.split([x // 2 for x in self.axes_dim], dim=1)
+
+        freqs_frame = freqs_pos[0][idx : idx + frame].view(frame, 1, 1, -1).expand(frame, height, width, -1)
+        if self.scale_rope:
+            freqs_height = torch.cat([freqs_neg[1][-(height - height // 2) :], freqs_pos[1][: height // 2]], dim=0)
+            freqs_height = freqs_height.view(1, height, 1, -1).expand(frame, height, width, -1)
+            freqs_width = torch.cat([freqs_neg[2][-(width - width // 2) :], freqs_pos[2][: width // 2]], dim=0)
+            freqs_width = freqs_width.view(1, 1, width, -1).expand(frame, height, width, -1)
+        else:
+            freqs_height = freqs_pos[1][:height].view(1, height, 1, -1).expand(frame, height, width, -1)
+            freqs_width = freqs_pos[2][:width].view(1, 1, width, -1).expand(frame, height, width, -1)
+
+        freqs = torch.cat([freqs_frame, freqs_height, freqs_width], dim=-1).reshape(seq_lens, -1)
+        return freqs.clone().contiguous()
+
+    @functools.lru_cache(maxsize=None)
+    def _compute_condition_freqs(self, frame, height, width):
+        seq_lens = frame * height * width
+        freqs_pos = self.pos_freqs.split([x // 2 for x in self.axes_dim], dim=1)
+        freqs_neg = self.neg_freqs.split([x // 2 for x in self.axes_dim], dim=1)
+
+        freqs_frame = freqs_neg[0][-1:].view(frame, 1, 1, -1).expand(frame, height, width, -1)
+        if self.scale_rope:
+            freqs_height = torch.cat([freqs_neg[1][-(height - height // 2) :], freqs_pos[1][: height // 2]], dim=0)
+            freqs_height = freqs_height.view(1, height, 1, -1).expand(frame, height, width, -1)
+            freqs_width = torch.cat([freqs_neg[2][-(width - width // 2) :], freqs_pos[2][: width // 2]], dim=0)
+            freqs_width = freqs_width.view(1, 1, width, -1).expand(frame, height, width, -1)
+        else:
+            freqs_height = freqs_pos[1][:height].view(1, height, 1, -1).expand(frame, height, width, -1)
+            freqs_width = freqs_pos[2][:width].view(1, 1, width, -1).expand(frame, height, width, -1)
+
+        freqs = torch.cat([freqs_frame, freqs_height, freqs_width], dim=-1).reshape(seq_lens, -1)
+        return freqs.clone().contiguous()
+
+
 class QwenDoubleStreamAttnProcessor2_0:
     """
     Attention processor for Qwen double-stream architecture, matching DoubleStreamLayerMegatron logic. This processor
@@ -578,14 +701,21 @@ class QwenImageTransformer2DModel(
         guidance_embeds: bool = False,  # TODO: this should probably be removed
         axes_dims_rope: Tuple[int, int, int] = (16, 56, 56),
         zero_cond_t: bool = False,
+        use_additional_t_cond: bool = False,
+        use_layer3d_rope: bool = False,
     ):
         super().__init__()
         self.out_channels = out_channels or in_channels
         self.inner_dim = num_attention_heads * attention_head_dim
 
-        self.pos_embed = QwenEmbedRope(theta=10000, axes_dim=list(axes_dims_rope), scale_rope=True)
+        if not use_layer3d_rope:
+            self.pos_embed = QwenEmbedRope(theta=10000, axes_dim=list(axes_dims_rope), scale_rope=True)
+        else:
+            self.pos_embed = QwenEmbedLayer3DRope(theta=10000, axes_dim=list(axes_dims_rope), scale_rope=True)
 
-        self.time_text_embed = QwenTimestepProjEmbeddings(embedding_dim=self.inner_dim)
+        self.time_text_embed = QwenTimestepProjEmbeddings(
+            embedding_dim=self.inner_dim, use_additional_t_cond=use_additional_t_cond
+        )
 
         self.txt_norm = RMSNorm(joint_attention_dim, eps=1e-6)
 
@@ -621,6 +751,7 @@ class QwenImageTransformer2DModel(
         guidance: torch.Tensor = None,  # TODO: this should probably be removed
         attention_kwargs: Optional[Dict[str, Any]] = None,
         controlnet_block_samples=None,
+        additional_t_cond=None,
         return_dict: bool = True,
     ) -> Union[torch.Tensor, Transformer2DModelOutput]:
         """
@@ -683,9 +814,9 @@ class QwenImageTransformer2DModel(
             guidance = guidance.to(hidden_states.dtype) * 1000
 
         temb = (
-            self.time_text_embed(timestep, hidden_states)
+            self.time_text_embed(timestep, hidden_states, additional_t_cond)
             if guidance is None
-            else self.time_text_embed(timestep, guidance, hidden_states)
+            else self.time_text_embed(timestep, guidance, hidden_states, additional_t_cond)
         )
 
         image_rotary_emb = self.pos_embed(img_shapes, txt_seq_lens, device=hidden_states.device)

src/diffusers/pipelines/__init__.py

@@ -422,6 +422,7 @@ else:
         "QwenImageEditInpaintPipeline",
         "QwenImageControlNetInpaintPipeline",
         "QwenImageControlNetPipeline",
+        "QwenImageLayeredPipeline",
     ]
     _import_structure["chronoedit"] = ["ChronoEditPipeline"]
 try:
@@ -764,6 +765,7 @@ if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT:
             QwenImageEditPlusPipeline,
             QwenImageImg2ImgPipeline,
             QwenImageInpaintPipeline,
+            QwenImageLayeredPipeline,
             QwenImagePipeline,
         )
         from .sana import (

src/diffusers/pipelines/qwenimage/__init__.py

@@ -31,6 +31,7 @@ else:
     _import_structure["pipeline_qwenimage_edit_plus"] = ["QwenImageEditPlusPipeline"]
     _import_structure["pipeline_qwenimage_img2img"] = ["QwenImageImg2ImgPipeline"]
     _import_structure["pipeline_qwenimage_inpaint"] = ["QwenImageInpaintPipeline"]
+    _import_structure["pipeline_qwenimage_layered"] = ["QwenImageLayeredPipeline"]
 
 if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT:
     try:
@@ -47,6 +48,7 @@ if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT:
         from .pipeline_qwenimage_edit_plus import QwenImageEditPlusPipeline
         from .pipeline_qwenimage_img2img import QwenImageImg2ImgPipeline
         from .pipeline_qwenimage_inpaint import QwenImageInpaintPipeline
+        from .pipeline_qwenimage_layered import QwenImageLayeredPipeline
 else:
     import sys
 

src/diffusers/pipelines/qwenimage/pipeline_qwenimage_layered.py

@@ -0,0 +1,905 @@
+# Copyright 2025 Qwen-Image Team and 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
+import math
+from typing import Any, Callable, Dict, List, Optional, Union
+
+import numpy as np
+import torch
+from transformers import Qwen2_5_VLForConditionalGeneration, Qwen2Tokenizer, Qwen2VLProcessor
+
+from ...image_processor import PipelineImageInput, VaeImageProcessor
+from ...loaders import QwenImageLoraLoaderMixin
+from ...models import AutoencoderKLQwenImage, QwenImageTransformer2DModel
+from ...schedulers import FlowMatchEulerDiscreteScheduler
+from ...utils import is_torch_xla_available, logging, replace_example_docstring
+from ...utils.torch_utils import randn_tensor
+from ..pipeline_utils import DiffusionPipeline
+from .pipeline_output import QwenImagePipelineOutput
+
+
+if is_torch_xla_available():
+    import torch_xla.core.xla_model as xm
+
+    XLA_AVAILABLE = True
+else:
+    XLA_AVAILABLE = False
+
+
+logger = logging.get_logger(__name__)  # pylint: disable=invalid-name
+
+EXAMPLE_DOC_STRING = """
+    Examples:
+        ```py
+        >>> import torch
+        >>> from PIL import Image
+        >>> from diffusers import QwenImageLayeredPipeline
+        >>> from diffusers.utils import load_image
+
+        >>> pipe = QwenImageLayeredPipeline.from_pretrained("Qwen/Qwen-Image-Layered", torch_dtype=torch.bfloat16)
+        >>> pipe.to("cuda")
+        >>> image = load_image(
+        ...     "https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/diffusers/yarn-art-pikachu.png"
+        ... ).convert("RGBA")
+        >>> prompt = ""
+        >>> # Depending on the variant being used, the pipeline call will slightly vary.
+        >>> # Refer to the pipeline documentation for more details.
+        >>> images = pipe(
+        ...     image,
+        ...     prompt,
+        ...     num_inference_steps=50,
+        ...     true_cfg_scale=4.0,
+        ...     layers=4,
+        ...     resolution=640,
+        ...     cfg_normalize=False,
+        ...     use_en_prompt=True,
+        ... ).images[0]
+        >>> for i, image in enumerate(images):
+        ...     image.save(f"{i}.out.png")
+        ```
+"""
+
+
+# Copied from diffusers.pipelines.qwenimage.pipeline_qwenimage.calculate_shift
+def calculate_shift(
+    image_seq_len,
+    base_seq_len: int = 256,
+    max_seq_len: int = 4096,
+    base_shift: float = 0.5,
+    max_shift: float = 1.15,
+):
+    m = (max_shift - base_shift) / (max_seq_len - base_seq_len)
+    b = base_shift - m * base_seq_len
+    mu = image_seq_len * m + b
+    return mu
+
+
+# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.retrieve_timesteps
+def retrieve_timesteps(
+    scheduler,
+    num_inference_steps: Optional[int] = None,
+    device: Optional[Union[str, torch.device]] = None,
+    timesteps: Optional[List[int]] = None,
+    sigmas: Optional[List[float]] = None,
+    **kwargs,
+):
+    r"""
+    Calls the scheduler's `set_timesteps` method and retrieves timesteps from the scheduler after the call. Handles
+    custom timesteps. Any kwargs will be supplied to `scheduler.set_timesteps`.
+
+    Args:
+        scheduler (`SchedulerMixin`):
+            The scheduler to get timesteps from.
+        num_inference_steps (`int`):
+            The number of diffusion steps used when generating samples with a pre-trained model. If used, `timesteps`
+            must be `None`.
+        device (`str` or `torch.device`, *optional*):
+            The device to which the timesteps should be moved to. If `None`, the timesteps are not moved.
+        timesteps (`List[int]`, *optional*):
+            Custom timesteps used to override the timestep spacing strategy of the scheduler. If `timesteps` is passed,
+            `num_inference_steps` and `sigmas` must be `None`.
+        sigmas (`List[float]`, *optional*):
+            Custom sigmas used to override the timestep spacing strategy of the scheduler. If `sigmas` is passed,
+            `num_inference_steps` and `timesteps` must be `None`.
+
+    Returns:
+        `Tuple[torch.Tensor, int]`: A tuple where the first element is the timestep schedule from the scheduler and the
+        second element is the number of inference steps.
+    """
+    if timesteps is not None and sigmas is not None:
+        raise ValueError("Only one of `timesteps` or `sigmas` can be passed. Please choose one to set custom values")
+    if timesteps is not None:
+        accepts_timesteps = "timesteps" in set(inspect.signature(scheduler.set_timesteps).parameters.keys())
+        if not accepts_timesteps:
+            raise ValueError(
+                f"The current scheduler class {scheduler.__class__}'s `set_timesteps` does not support custom"
+                f" timestep schedules. Please check whether you are using the correct scheduler."
+            )
+        scheduler.set_timesteps(timesteps=timesteps, device=device, **kwargs)
+        timesteps = scheduler.timesteps
+        num_inference_steps = len(timesteps)
+    elif sigmas is not None:
+        accept_sigmas = "sigmas" in set(inspect.signature(scheduler.set_timesteps).parameters.keys())
+        if not accept_sigmas:
+            raise ValueError(
+                f"The current scheduler class {scheduler.__class__}'s `set_timesteps` does not support custom"
+                f" sigmas schedules. Please check whether you are using the correct scheduler."
+            )
+        scheduler.set_timesteps(sigmas=sigmas, device=device, **kwargs)
+        timesteps = scheduler.timesteps
+        num_inference_steps = len(timesteps)
+    else:
+        scheduler.set_timesteps(num_inference_steps, device=device, **kwargs)
+        timesteps = scheduler.timesteps
+    return timesteps, num_inference_steps
+
+
+# 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")
+
+
+# Copied from diffusers.pipelines.qwenimage.pipeline_qwenimage_edit_plus.calculate_dimensions
+def calculate_dimensions(target_area, ratio):
+    width = math.sqrt(target_area * ratio)
+    height = width / ratio
+
+    width = round(width / 32) * 32
+    height = round(height / 32) * 32
+
+    return width, height
+
+
+class QwenImageLayeredPipeline(DiffusionPipeline, QwenImageLoraLoaderMixin):
+    r"""
+    The Qwen-Image-Layered pipeline for image decomposing.
+
+    Args:
+        transformer ([`QwenImageTransformer2DModel`]):
+            Conditional Transformer (MMDiT) architecture to denoise the encoded image latents.
+        scheduler ([`FlowMatchEulerDiscreteScheduler`]):
+            A scheduler to be used in combination with `transformer` to denoise the encoded image latents.
+        vae ([`AutoencoderKL`]):
+            Variational Auto-Encoder (VAE) Model to encode and decode images to and from latent representations.
+        text_encoder ([`Qwen2.5-VL-7B-Instruct`]):
+            [Qwen2.5-VL-7B-Instruct](https://huggingface.co/Qwen/Qwen2.5-VL-7B-Instruct), specifically the
+            [Qwen2.5-VL-7B-Instruct](https://huggingface.co/Qwen/Qwen2.5-VL-7B-Instruct) variant.
+        tokenizer (`QwenTokenizer`):
+            Tokenizer of class
+            [CLIPTokenizer](https://huggingface.co/docs/transformers/en/model_doc/clip#transformers.CLIPTokenizer).
+    """
+
+    model_cpu_offload_seq = "text_encoder->transformer->vae"
+    _callback_tensor_inputs = ["latents", "prompt_embeds"]
+
+    def __init__(
+        self,
+        scheduler: FlowMatchEulerDiscreteScheduler,
+        vae: AutoencoderKLQwenImage,
+        text_encoder: Qwen2_5_VLForConditionalGeneration,
+        tokenizer: Qwen2Tokenizer,
+        processor: Qwen2VLProcessor,
+        transformer: QwenImageTransformer2DModel,
+    ):
+        super().__init__()
+
+        self.register_modules(
+            vae=vae,
+            text_encoder=text_encoder,
+            tokenizer=tokenizer,
+            processor=processor,
+            transformer=transformer,
+            scheduler=scheduler,
+        )
+        self.vae_scale_factor = 2 ** len(self.vae.temperal_downsample) if getattr(self, "vae", None) else 8
+        self.latent_channels = self.vae.config.z_dim if getattr(self, "vae", None) else 16
+        # QwenImage latents are turned into 2x2 patches and packed. This means the latent width and height has to be divisible
+        # by the patch size. So the vae scale factor is multiplied by the patch size to account for this
+        self.image_processor = VaeImageProcessor(vae_scale_factor=self.vae_scale_factor * 2)
+        self.vl_processor = processor
+        self.tokenizer_max_length = 1024
+
+        self.prompt_template_encode = "<|im_start|>system\nDescribe the image by detailing the color, shape, size, texture, quantity, text, spatial relationships of the objects and background:<|im_end|>\n<|im_start|>user\n{}<|im_end|>\n<|im_start|>assistant\n"
+        self.prompt_template_encode_start_idx = 34
+        self.image_caption_prompt_cn = """<|im_start|>system\nYou are a helpful assistant.<|im_end|>\n<|im_start|>user\n# 图像标注器\n你是一个专业的图像标注器。请基于输入图像，撰写图注:\n1.
+使用自然、描述性的语言撰写图注，不要使用结构化形式或富文本形式。\n2. 通过加入以下内容，丰富图注细节：\n - 对象的属性：如数量、颜色、形状、大小、位置、材质、状态、动作等\n -
+对象间的视觉关系：如空间关系、功能关系、动作关系、从属关系、比较关系、因果关系等\n - 环境细节：例如天气、光照、颜色、纹理、气氛等\n - 文字内容：识别图像中清晰可见的文字，不做翻译和解释，用引号在图注中强调\n3.
+保持真实性与准确性：\n - 不要使用笼统的描述\n -
+描述图像中所有可见的信息，但不要加入没有在图像中出现的内容\n<|vision_start|><|image_pad|><|vision_end|><|im_end|>\n<|im_start|>assistant\n"""
+        self.image_caption_prompt_en = """<|im_start|>system\nYou are a helpful assistant.<|im_end|>\n<|im_start|>user\n# Image Annotator\nYou are a professional
+image annotator. Please write an image caption based on the input image:\n1. Write the caption using natural,
+descriptive language without structured formats or rich text.\n2. Enrich caption details by including: \n - Object
+attributes, such as quantity, color, shape, size, material, state, position, actions, and so on\n - Vision Relations
+between objects, such as spatial relations, functional relations, possessive relations, attachment relations, action
+relations, comparative relations, causal relations, and so on\n - Environmental details, such as weather, lighting,
+colors, textures, atmosphere, and so on\n - Identify the text clearly visible in the image, without translation or
+explanation, and highlight it in the caption with quotation marks\n3. Maintain authenticity and accuracy:\n - Avoid
+generalizations\n - Describe all visible information in the image, while do not add information not explicitly shown in
+the image\n<|vision_start|><|image_pad|><|vision_end|><|im_end|>\n<|im_start|>assistant\n"""
+        self.default_sample_size = 128
+
+    # Copied from diffusers.pipelines.qwenimage.pipeline_qwenimage.QwenImagePipeline._extract_masked_hidden
+    def _extract_masked_hidden(self, hidden_states: torch.Tensor, mask: torch.Tensor):
+        bool_mask = mask.bool()
+        valid_lengths = bool_mask.sum(dim=1)
+        selected = hidden_states[bool_mask]
+        split_result = torch.split(selected, valid_lengths.tolist(), dim=0)
+
+        return split_result
+
+    def _get_qwen_prompt_embeds(
+        self,
+        prompt: Union[str, List[str]] = None,
+        device: Optional[torch.device] = None,
+        dtype: Optional[torch.dtype] = None,
+    ):
+        device = device or self._execution_device
+        dtype = dtype or self.text_encoder.dtype
+
+        prompt = [prompt] if isinstance(prompt, str) else prompt
+
+        template = self.prompt_template_encode
+        drop_idx = self.prompt_template_encode_start_idx
+        txt = [template.format(e) for e in prompt]
+        txt_tokens = self.tokenizer(
+            txt,
+            padding=True,
+            return_tensors="pt",
+        ).to(device)
+        encoder_hidden_states = self.text_encoder(
+            input_ids=txt_tokens.input_ids,
+            attention_mask=txt_tokens.attention_mask,
+            output_hidden_states=True,
+        )
+        hidden_states = encoder_hidden_states.hidden_states[-1]
+        split_hidden_states = self._extract_masked_hidden(hidden_states, txt_tokens.attention_mask)
+        split_hidden_states = [e[drop_idx:] for e in split_hidden_states]
+        attn_mask_list = [torch.ones(e.size(0), dtype=torch.long, device=e.device) for e in split_hidden_states]
+        max_seq_len = max([e.size(0) for e in split_hidden_states])
+        prompt_embeds = torch.stack(
+            [torch.cat([u, u.new_zeros(max_seq_len - u.size(0), u.size(1))]) for u in split_hidden_states]
+        )
+        encoder_attention_mask = torch.stack(
+            [torch.cat([u, u.new_zeros(max_seq_len - u.size(0))]) for u in attn_mask_list]
+        )
+
+        prompt_embeds = prompt_embeds.to(dtype=dtype, device=device)
+
+        return prompt_embeds, encoder_attention_mask
+
+    # Copied from diffusers.pipelines.qwenimage.pipeline_qwenimage.QwenImagePipeline.encode_prompt
+    def encode_prompt(
+        self,
+        prompt: Union[str, List[str]],
+        device: Optional[torch.device] = None,
+        num_images_per_prompt: int = 1,
+        prompt_embeds: Optional[torch.Tensor] = None,
+        prompt_embeds_mask: Optional[torch.Tensor] = None,
+        max_sequence_length: int = 1024,
+    ):
+        r"""
+
+        Args:
+            prompt (`str` or `List[str]`, *optional*):
+                prompt to be encoded
+            device: (`torch.device`):
+                torch device
+            num_images_per_prompt (`int`):
+                number of images that should be generated per prompt
+            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.
+        """
+        device = device or self._execution_device
+
+        prompt = [prompt] if isinstance(prompt, str) else prompt
+        batch_size = len(prompt) if prompt_embeds is None else prompt_embeds.shape[0]
+
+        if prompt_embeds is None:
+            prompt_embeds, prompt_embeds_mask = self._get_qwen_prompt_embeds(prompt, device)
+
+        prompt_embeds = prompt_embeds[:, :max_sequence_length]
+        prompt_embeds_mask = prompt_embeds_mask[:, :max_sequence_length]
+
+        _, seq_len, _ = prompt_embeds.shape
+        prompt_embeds = prompt_embeds.repeat(1, num_images_per_prompt, 1)
+        prompt_embeds = prompt_embeds.view(batch_size * num_images_per_prompt, seq_len, -1)
+        prompt_embeds_mask = prompt_embeds_mask.repeat(1, num_images_per_prompt, 1)
+        prompt_embeds_mask = prompt_embeds_mask.view(batch_size * num_images_per_prompt, seq_len)
+
+        return prompt_embeds, prompt_embeds_mask
+
+    def get_image_caption(self, prompt_image, use_en_prompt=True, device=None):
+        if use_en_prompt:
+            prompt = self.image_caption_prompt_en
+        else:
+            prompt = self.image_caption_prompt_cn
+        model_inputs = self.vl_processor(
+            text=prompt,
+            images=prompt_image,
+            padding=True,
+            return_tensors="pt",
+        ).to(device)
+        generated_ids = self.text_encoder.generate(**model_inputs, max_new_tokens=512)
+        generated_ids_trimmed = [
+            out_ids[len(in_ids) :] for in_ids, out_ids in zip(model_inputs.input_ids, generated_ids)
+        ]
+        output_text = self.vl_processor.batch_decode(
+            generated_ids_trimmed, skip_special_tokens=True, clean_up_tokenization_spaces=False
+        )[0]
+        return output_text.strip()
+
+    def check_inputs(
+        self,
+        height,
+        width,
+        negative_prompt=None,
+        prompt_embeds=None,
+        negative_prompt_embeds=None,
+        prompt_embeds_mask=None,
+        negative_prompt_embeds_mask=None,
+        callback_on_step_end_tensor_inputs=None,
+        max_sequence_length=None,
+    ):
+        if height % (self.vae_scale_factor * 2) != 0 or width % (self.vae_scale_factor * 2) != 0:
+            logger.warning(
+                f"`height` and `width` have to be divisible by {self.vae_scale_factor * 2} but are {height} and {width}. Dimensions will be resized accordingly"
+            )
+
+        if callback_on_step_end_tensor_inputs is not None and not all(
+            k in self._callback_tensor_inputs for k in callback_on_step_end_tensor_inputs
+        ):
+            raise ValueError(
+                f"`callback_on_step_end_tensor_inputs` has to be in {self._callback_tensor_inputs}, but found {[k for k in callback_on_step_end_tensor_inputs if k not in self._callback_tensor_inputs]}"
+            )
+
+        if negative_prompt is not None and negative_prompt_embeds is not None:
+            raise ValueError(
+                f"Cannot forward both `negative_prompt`: {negative_prompt} and `negative_prompt_embeds`:"
+                f" {negative_prompt_embeds}. Please make sure to only forward one of the two."
+            )
+
+        if prompt_embeds is not None and prompt_embeds_mask is None:
+            raise ValueError(
+                "If `prompt_embeds` are provided, `prompt_embeds_mask` also have to be passed. Make sure to generate `prompt_embeds_mask` from the same text encoder that was used to generate `prompt_embeds`."
+            )
+        if negative_prompt_embeds is not None and negative_prompt_embeds_mask is None:
+            raise ValueError(
+                "If `negative_prompt_embeds` are provided, `negative_prompt_embeds_mask` also have to be passed. Make sure to generate `negative_prompt_embeds_mask` from the same text encoder that was used to generate `negative_prompt_embeds`."
+            )
+
+        if max_sequence_length is not None and max_sequence_length > 1024:
+            raise ValueError(f"`max_sequence_length` cannot be greater than 1024 but is {max_sequence_length}")
+
+    @staticmethod
+    def _pack_latents(latents, batch_size, num_channels_latents, height, width, layers):
+        latents = latents.view(batch_size, layers, num_channels_latents, height // 2, 2, width // 2, 2)
+        latents = latents.permute(0, 1, 3, 5, 2, 4, 6)
+        latents = latents.reshape(batch_size, layers * (height // 2) * (width // 2), num_channels_latents * 4)
+
+        return latents
+
+    @staticmethod
+    def _unpack_latents(latents, height, width, layers, vae_scale_factor):
+        batch_size, num_patches, channels = latents.shape
+
+        # VAE applies 8x compression on images but we must also account for packing which requires
+        # latent height and width to be divisible by 2.
+        height = 2 * (int(height) // (vae_scale_factor * 2))
+        width = 2 * (int(width) // (vae_scale_factor * 2))
+
+        latents = latents.view(batch_size, layers + 1, height // 2, width // 2, channels // 4, 2, 2)
+        latents = latents.permute(0, 1, 4, 2, 5, 3, 6)
+
+        latents = latents.reshape(batch_size, layers + 1, channels // (2 * 2), height, width)
+        latents = latents.permute(0, 2, 1, 3, 4)  # (b, c, f, h, w)
+
+        return latents
+
+    # Copied from diffusers.pipelines.qwenimage.pipeline_qwenimage_edit.QwenImageEditPipeline._encode_vae_image
+    def _encode_vae_image(self, image: torch.Tensor, generator: torch.Generator):
+        if isinstance(generator, list):
+            image_latents = [
+                retrieve_latents(self.vae.encode(image[i : i + 1]), generator=generator[i], sample_mode="argmax")
+                for i in range(image.shape[0])
+            ]
+            image_latents = torch.cat(image_latents, dim=0)
+        else:
+            image_latents = retrieve_latents(self.vae.encode(image), generator=generator, sample_mode="argmax")
+        latents_mean = (
+            torch.tensor(self.vae.config.latents_mean)
+            .view(1, self.latent_channels, 1, 1, 1)
+            .to(image_latents.device, image_latents.dtype)
+        )
+        latents_std = (
+            torch.tensor(self.vae.config.latents_std)
+            .view(1, self.latent_channels, 1, 1, 1)
+            .to(image_latents.device, image_latents.dtype)
+        )
+        image_latents = (image_latents - latents_mean) / latents_std
+
+        return image_latents
+
+    def prepare_latents(
+        self,
+        image,
+        batch_size,
+        num_channels_latents,
+        height,
+        width,
+        layers,
+        dtype,
+        device,
+        generator,
+        latents=None,
+    ):
+        # VAE applies 8x compression on images but we must also account for packing which requires
+        # latent height and width to be divisible by 2.
+        height = 2 * (int(height) // (self.vae_scale_factor * 2))
+        width = 2 * (int(width) // (self.vae_scale_factor * 2))
+
+        shape = (
+            batch_size,
+            layers + 1,
+            num_channels_latents,
+            height,
+            width,
+        )  ### the generated first image is combined image
+
+        image_latents = None
+        if image is not None:
+            image = image.to(device=device, dtype=dtype)
+            if image.shape[1] != self.latent_channels:
+                image_latents = self._encode_vae_image(image=image, generator=generator)
+            else:
+                image_latents = image
+            if batch_size > image_latents.shape[0] and batch_size % image_latents.shape[0] == 0:
+                # expand init_latents for batch_size
+                additional_image_per_prompt = batch_size // image_latents.shape[0]
+                image_latents = torch.cat([image_latents] * additional_image_per_prompt, dim=0)
+            elif batch_size > image_latents.shape[0] and batch_size % image_latents.shape[0] != 0:
+                raise ValueError(
+                    f"Cannot duplicate `image` of batch size {image_latents.shape[0]} to {batch_size} text prompts."
+                )
+            else:
+                image_latents = torch.cat([image_latents], dim=0)
+
+            image_latent_height, image_latent_width = image_latents.shape[3:]
+            image_latents = image_latents.permute(0, 2, 1, 3, 4)  # (b, c, f, h, w) -> (b, f, c, h, w)
+            image_latents = self._pack_latents(
+                image_latents, batch_size, num_channels_latents, image_latent_height, image_latent_width, 1
+            )
+
+        if isinstance(generator, list) and len(generator) != batch_size:
+            raise ValueError(
+                f"You have passed a list of generators of length {len(generator)}, but requested an effective batch"
+                f" size of {batch_size}. Make sure the batch size matches the length of the generators."
+            )
+        if latents is None:
+            latents = randn_tensor(shape, generator=generator, device=device, dtype=dtype)
+            latents = self._pack_latents(latents, batch_size, num_channels_latents, height, width, layers + 1)
+        else:
+            latents = latents.to(device=device, dtype=dtype)
+
+        return latents, image_latents
+
+    @property
+    def guidance_scale(self):
+        return self._guidance_scale
+
+    @property
+    def attention_kwargs(self):
+        return self._attention_kwargs
+
+    @property
+    def num_timesteps(self):
+        return self._num_timesteps
+
+    @property
+    def current_timestep(self):
+        return self._current_timestep
+
+    @property
+    def interrupt(self):
+        return self._interrupt
+
+    @torch.no_grad()
+    @replace_example_docstring(EXAMPLE_DOC_STRING)
+    def __call__(
+        self,
+        image: Optional[PipelineImageInput] = None,
+        prompt: Union[str, List[str]] = None,
+        negative_prompt: Union[str, List[str]] = None,
+        true_cfg_scale: float = 4.0,
+        layers: Optional[int] = 4,
+        num_inference_steps: int = 50,
+        sigmas: Optional[List[float]] = None,
+        guidance_scale: Optional[float] = None,
+        num_images_per_prompt: int = 1,
+        generator: Optional[Union[torch.Generator, List[torch.Generator]]] = None,
+        latents: Optional[torch.Tensor] = None,
+        prompt_embeds: Optional[torch.Tensor] = None,
+        prompt_embeds_mask: Optional[torch.Tensor] = None,
+        negative_prompt_embeds: Optional[torch.Tensor] = None,
+        negative_prompt_embeds_mask: Optional[torch.Tensor] = None,
+        output_type: Optional[str] = "pil",
+        return_dict: bool = True,
+        attention_kwargs: Optional[Dict[str, Any]] = None,
+        callback_on_step_end: Optional[Callable[[int, int, Dict], None]] = None,
+        callback_on_step_end_tensor_inputs: List[str] = ["latents"],
+        max_sequence_length: int = 512,
+        resolution: int = 640,
+        cfg_normalize: bool = False,
+        use_en_prompt: bool = False,
+    ):
+        r"""
+        Function invoked when calling the pipeline for generation.
+
+        Args:
+            image (`torch.Tensor`, `PIL.Image.Image`, `np.ndarray`, `List[torch.Tensor]`, `List[PIL.Image.Image]`, or `List[np.ndarray]`):
+                `Image`, numpy array or tensor representing an image batch to be used as the starting point. For both
+                numpy array and pytorch tensor, the expected value range is between `[0, 1]` If it's a tensor or a list
+                or tensors, the expected shape should be `(B, C, H, W)` or `(C, H, W)`. If it is a numpy array or a
+                list of arrays, the expected shape should be `(B, H, W, C)` or `(H, W, C)` It can also accept image
+                latents as `image`, but if passing latents directly it is not encoded again.
+            prompt (`str` or `List[str]`, *optional*):
+                The prompt or prompts to guide the image generation. If not defined, one has to pass `prompt_embeds`.
+                instead.
+            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 `true_cfg_scale` is
+                not greater than `1`).
+            true_cfg_scale (`float`, *optional*, defaults to 1.0):
+                true_cfg_scale (`float`, *optional*, defaults to 1.0): Guidance scale as defined in [Classifier-Free
+                Diffusion Guidance](https://huggingface.co/papers/2207.12598). `true_cfg_scale` is defined as `w` of
+                equation 2. of [Imagen Paper](https://huggingface.co/papers/2205.11487). Classifier-free guidance is
+                enabled by setting `true_cfg_scale > 1` and a provided `negative_prompt`. Higher guidance scale
+                encourages to generate images that are closely linked to the text `prompt`, usually at the expense of
+                lower image quality.
+            num_inference_steps (`int`, *optional*, defaults to 50):
+                The number of denoising steps. More denoising steps usually lead to a higher quality image at the
+                expense of slower inference.
+            sigmas (`List[float]`, *optional*):
+                Custom sigmas to use for the denoising process with schedulers which support a `sigmas` argument in
+                their `set_timesteps` method. If not defined, the default behavior when `num_inference_steps` is passed
+                will be used.
+            guidance_scale (`float`, *optional*, defaults to None):
+                A guidance scale value for guidance distilled models. Unlike the traditional classifier-free guidance
+                where the guidance scale is applied during inference through noise prediction rescaling, guidance
+                distilled models take the guidance scale directly as an input parameter during forward pass. Guidance
+                scale is enabled by setting `guidance_scale > 1`. Higher guidance scale encourages to generate images
+                that are closely linked to the text `prompt`, usually at the expense of lower image quality. This
+                parameter in the pipeline is there to support future guidance-distilled models when they come up. It is
+                ignored when not using guidance distilled models. To enable traditional classifier-free guidance,
+                please pass `true_cfg_scale > 1.0` and `negative_prompt` (even an empty negative prompt like " " should
+                enable classifier-free guidance computations).
+            num_images_per_prompt (`int`, *optional*, defaults to 1):
+                The number of images to generate per prompt.
+            generator (`torch.Generator` or `List[torch.Generator]`, *optional*):
+                One or a list of [torch generator(s)](https://pytorch.org/docs/stable/generated/torch.Generator.html)
+                to make generation deterministic.
+            latents (`torch.Tensor`, *optional*):
+                Pre-generated noisy latents, sampled from a Gaussian distribution, to be used as inputs for image
+                generation. Can be used to tweak the same generation with different prompts. If not provided, a latents
+                tensor will be generated by sampling using the supplied random `generator`.
+            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.
+            output_type (`str`, *optional*, defaults to `"pil"`):
+                The output format of the generate image. Choose between
+                [PIL](https://pillow.readthedocs.io/en/stable/): `PIL.Image.Image` or `np.array`.
+            return_dict (`bool`, *optional*, defaults to `True`):
+                Whether or not to return a [`~pipelines.qwenimage.QwenImagePipelineOutput`] instead of a plain tuple.
+            attention_kwargs (`dict`, *optional*):
+                A kwargs dictionary that if specified is passed along to the `AttentionProcessor` as defined under
+                `self.processor` in
+                [diffusers.models.attention_processor](https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/attention_processor.py).
+            callback_on_step_end (`Callable`, *optional*):
+                A function that calls at the end of each denoising steps during the inference. The function is called
+                with the following arguments: `callback_on_step_end(self: DiffusionPipeline, step: int, timestep: int,
+                callback_kwargs: Dict)`. `callback_kwargs` will include a list of all tensors as specified by
+                `callback_on_step_end_tensor_inputs`.
+            callback_on_step_end_tensor_inputs (`List`, *optional*):
+                The list of tensor inputs for the `callback_on_step_end` function. The tensors specified in the list
+                will be passed as `callback_kwargs` argument. You will only be able to include variables listed in the
+                `._callback_tensor_inputs` attribute of your pipeline class.
+            max_sequence_length (`int` defaults to 512): Maximum sequence length to use with the `prompt`.
+            resolution (`int`, *optional*, defaults to 640):
+                using different bucket in (640, 1024) to determin the condition and output resolution
+            cfg_normalize (`bool`, *optional*, defaults to `False`)
+                whether enable cfg normalization.
+            use_en_prompt (`bool`, *optional*, defaults to `False`)
+                automatic caption language if user does not provide caption
+
+        Examples:
+
+        Returns:
+            [`~pipelines.qwenimage.QwenImagePipelineOutput`] or `tuple`:
+            [`~pipelines.qwenimage.QwenImagePipelineOutput`] if `return_dict` is True, otherwise a `tuple`. When
+            returning a tuple, the first element is a list with the generated images.
+        """
+        image_size = image[0].size if isinstance(image, list) else image.size
+        assert resolution in [640, 1024], f"resolution must be either 640 or 1024, but got {resolution}"
+        calculated_width, calculated_height = calculate_dimensions(
+            resolution * resolution, image_size[0] / image_size[1]
+        )
+        height = calculated_height
+        width = calculated_width
+
+        multiple_of = self.vae_scale_factor * 2
+        width = width // multiple_of * multiple_of
+        height = height // multiple_of * multiple_of
+
+        # 1. Check inputs. Raise error if not correct
+        self.check_inputs(
+            height,
+            width,
+            negative_prompt=negative_prompt,
+            prompt_embeds=prompt_embeds,
+            negative_prompt_embeds=negative_prompt_embeds,
+            prompt_embeds_mask=prompt_embeds_mask,
+            negative_prompt_embeds_mask=negative_prompt_embeds_mask,
+            callback_on_step_end_tensor_inputs=callback_on_step_end_tensor_inputs,
+            max_sequence_length=max_sequence_length,
+        )
+
+        self._guidance_scale = guidance_scale
+        self._attention_kwargs = attention_kwargs
+        self._current_timestep = None
+        self._interrupt = False
+
+        device = self._execution_device
+        # 2. Preprocess image
+        if image is not None and not (isinstance(image, torch.Tensor) and image.size(1) == self.latent_channels):
+            image = self.image_processor.resize(image, calculated_height, calculated_width)
+            prompt_image = image
+            image = self.image_processor.preprocess(image, calculated_height, calculated_width)
+            image = image.unsqueeze(2)
+            image = image.to(dtype=self.text_encoder.dtype)
+
+        if prompt is None or prompt == "" or prompt == " ":
+            prompt = self.get_image_caption(prompt_image, use_en_prompt=use_en_prompt, device=device)
+
+        # 3. Define call parameters
+        if prompt is not None and isinstance(prompt, str):
+            batch_size = 1
+        elif prompt is not None and isinstance(prompt, list):
+            batch_size = len(prompt)
+        else:
+            batch_size = prompt_embeds.shape[0]
+
+        has_neg_prompt = negative_prompt is not None or (
+            negative_prompt_embeds is not None and negative_prompt_embeds_mask is not None
+        )
+
+        if true_cfg_scale > 1 and not has_neg_prompt:
+            logger.warning(
+                f"true_cfg_scale is passed as {true_cfg_scale}, but classifier-free guidance is not enabled since no negative_prompt is provided."
+            )
+        elif true_cfg_scale <= 1 and has_neg_prompt:
+            logger.warning(
+                " negative_prompt is passed but classifier-free guidance is not enabled since true_cfg_scale <= 1"
+            )
+
+        do_true_cfg = true_cfg_scale > 1 and has_neg_prompt
+        prompt_embeds, prompt_embeds_mask = self.encode_prompt(
+            prompt=prompt,
+            prompt_embeds=prompt_embeds,
+            prompt_embeds_mask=prompt_embeds_mask,
+            device=device,
+            num_images_per_prompt=num_images_per_prompt,
+            max_sequence_length=max_sequence_length,
+        )
+        if do_true_cfg:
+            negative_prompt_embeds, negative_prompt_embeds_mask = self.encode_prompt(
+                prompt=negative_prompt,
+                prompt_embeds=negative_prompt_embeds,
+                prompt_embeds_mask=negative_prompt_embeds_mask,
+                device=device,
+                num_images_per_prompt=num_images_per_prompt,
+                max_sequence_length=max_sequence_length,
+            )
+
+        # 4. Prepare latent variables
+        num_channels_latents = self.transformer.config.in_channels // 4
+        latents, image_latents = self.prepare_latents(
+            image,
+            batch_size * num_images_per_prompt,
+            num_channels_latents,
+            height,
+            width,
+            layers,
+            prompt_embeds.dtype,
+            device,
+            generator,
+            latents,
+        )
+        img_shapes = [
+            [
+                *[
+                    (1, height // self.vae_scale_factor // 2, width // self.vae_scale_factor // 2)
+                    for _ in range(layers + 1)
+                ],
+                (1, calculated_height // self.vae_scale_factor // 2, calculated_width // self.vae_scale_factor // 2),
+            ]
+        ] * batch_size
+
+        # 5. Prepare timesteps
+        sigmas = np.linspace(1.0, 0, num_inference_steps + 1)[:-1] if sigmas is None else sigmas
+        image_seq_len = latents.shape[1]
+        base_seqlen = 256 * 256 / 16 / 16
+        mu = (image_latents.shape[1] / base_seqlen) ** 0.5
+        timesteps, num_inference_steps = retrieve_timesteps(
+            self.scheduler,
+            num_inference_steps,
+            device,
+            sigmas=sigmas,
+            mu=mu,
+        )
+        num_warmup_steps = max(len(timesteps) - num_inference_steps * self.scheduler.order, 0)
+        self._num_timesteps = len(timesteps)
+
+        # handle guidance
+        if self.transformer.config.guidance_embeds and guidance_scale is None:
+            raise ValueError("guidance_scale is required for guidance-distilled model.")
+        elif self.transformer.config.guidance_embeds:
+            guidance = torch.full([1], guidance_scale, device=device, dtype=torch.float32)
+            guidance = guidance.expand(latents.shape[0])
+        elif not self.transformer.config.guidance_embeds and guidance_scale is not None:
+            logger.warning(
+                f"guidance_scale is passed as {guidance_scale}, but ignored since the model is not guidance-distilled."
+            )
+            guidance = None
+        elif not self.transformer.config.guidance_embeds and guidance_scale is None:
+            guidance = None
+
+        if self.attention_kwargs is None:
+            self._attention_kwargs = {}
+
+        txt_seq_lens = prompt_embeds_mask.sum(dim=1).tolist() if prompt_embeds_mask is not None else None
+        negative_txt_seq_lens = (
+            negative_prompt_embeds_mask.sum(dim=1).tolist() if negative_prompt_embeds_mask is not None else None
+        )
+        is_rgb = torch.tensor([0] * batch_size).to(device=device, dtype=torch.long)
+        # 6. Denoising loop
+        self.scheduler.set_begin_index(0)
+        with self.progress_bar(total=num_inference_steps) as progress_bar:
+            for i, t in enumerate(timesteps):
+                if self.interrupt:
+                    continue
+
+                self._current_timestep = t
+
+                latent_model_input = latents
+                if image_latents is not None:
+                    latent_model_input = torch.cat([latents, image_latents], dim=1)
+
+                # broadcast to batch dimension in a way that's compatible with ONNX/Core ML
+                timestep = t.expand(latents.shape[0]).to(latents.dtype)
+                with self.transformer.cache_context("cond"):
+                    noise_pred = self.transformer(
+                        hidden_states=latent_model_input,
+                        timestep=timestep / 1000,
+                        guidance=guidance,
+                        encoder_hidden_states_mask=prompt_embeds_mask,
+                        encoder_hidden_states=prompt_embeds,
+                        img_shapes=img_shapes,
+                        txt_seq_lens=txt_seq_lens,
+                        attention_kwargs=self.attention_kwargs,
+                        additional_t_cond=is_rgb,
+                        return_dict=False,
+                    )[0]
+                    noise_pred = noise_pred[:, : latents.size(1)]
+
+                if do_true_cfg:
+                    with self.transformer.cache_context("uncond"):
+                        neg_noise_pred = self.transformer(
+                            hidden_states=latent_model_input,
+                            timestep=timestep / 1000,
+                            guidance=guidance,
+                            encoder_hidden_states_mask=negative_prompt_embeds_mask,
+                            encoder_hidden_states=negative_prompt_embeds,
+                            img_shapes=img_shapes,
+                            txt_seq_lens=negative_txt_seq_lens,
+                            attention_kwargs=self.attention_kwargs,
+                            additional_t_cond=is_rgb,
+                            return_dict=False,
+                        )[0]
+                    neg_noise_pred = neg_noise_pred[:, : latents.size(1)]
+                    comb_pred = neg_noise_pred + true_cfg_scale * (noise_pred - neg_noise_pred)
+
+                    if cfg_normalize:
+                        cond_norm = torch.norm(noise_pred, dim=-1, keepdim=True)
+                        noise_norm = torch.norm(comb_pred, dim=-1, keepdim=True)
+                        noise_pred = comb_pred * (cond_norm / noise_norm)
+                    else:
+                        noise_pred = comb_pred
+
+                # compute the previous noisy sample x_t -> x_t-1
+                latents_dtype = latents.dtype
+                latents = self.scheduler.step(noise_pred, t, latents, return_dict=False)[0]
+
+                if latents.dtype != latents_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
+                        latents = latents.to(latents_dtype)
+
+                if callback_on_step_end is not None:
+                    callback_kwargs = {}
+                    for k in callback_on_step_end_tensor_inputs:
+                        callback_kwargs[k] = locals()[k]
+                    callback_outputs = callback_on_step_end(self, i, t, callback_kwargs)
+
+                    latents = callback_outputs.pop("latents", latents)
+                    prompt_embeds = callback_outputs.pop("prompt_embeds", prompt_embeds)
+
+                # call the callback, if provided
+                if i == len(timesteps) - 1 or ((i + 1) > num_warmup_steps and (i + 1) % self.scheduler.order == 0):
+                    progress_bar.update()
+
+                if XLA_AVAILABLE:
+                    xm.mark_step()
+
+        self._current_timestep = None
+        if output_type == "latent":
+            image = latents
+        else:
+            latents = self._unpack_latents(latents, height, width, layers, self.vae_scale_factor)
+            latents = latents.to(self.vae.dtype)
+            latents_mean = (
+                torch.tensor(self.vae.config.latents_mean)
+                .view(1, self.vae.config.z_dim, 1, 1, 1)
+                .to(latents.device, latents.dtype)
+            )
+            latents_std = 1.0 / torch.tensor(self.vae.config.latents_std).view(1, self.vae.config.z_dim, 1, 1, 1).to(
+                latents.device, latents.dtype
+            )
+            latents = latents / latents_std + latents_mean
+
+            b, c, f, h, w = latents.shape
+
+            latents = latents[:, :, 1:]  # remove the first frame as it is the orgin input
+
+            latents = latents.permute(0, 2, 1, 3, 4).view(-1, c, 1, h, w)
+
+            image = self.vae.decode(latents, return_dict=False)[0]  # (b f) c 1 h w
+
+            image = image.squeeze(2)
+
+            image = self.image_processor.postprocess(image, output_type=output_type)
+            images = []
+            for bidx in range(b):
+                images.append(image[bidx * f : (bidx + 1) * f])
+
+        # Offload all models
+        self.maybe_free_model_hooks()
+
+        if not return_dict:
+            return (images,)
+
+        return QwenImagePipelineOutput(images=images)

src/diffusers/utils/dummy_torch_and_transformers_objects.py

@@ -2297,6 +2297,21 @@ class QwenImageInpaintPipeline(metaclass=DummyObject):
         requires_backends(cls, ["torch", "transformers"])
 
 
+class QwenImageLayeredPipeline(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 QwenImagePipeline(metaclass=DummyObject):
     _backends = ["torch", "transformers"]
 

tests/quantization/torchao/test_torchao.py

@@ -35,6 +35,7 @@ from diffusers.models.attention_processor import Attention
 from diffusers.quantizers import PipelineQuantizationConfig
 
 from ...testing_utils import (
+    Expectations,
     backend_empty_cache,
     backend_synchronize,
     enable_full_determinism,
@@ -497,8 +498,23 @@ class TorchAoTest(unittest.TestCase):
 
     def test_model_memory_usage(self):
         model_id = "hf-internal-testing/tiny-flux-pipe"
-        expected_memory_saving_ratio = 2.0
-
+        expected_memory_saving_ratios = Expectations(
+            {
+                # XPU: For this tiny model, per-tensor overheads (alignment, fragmentation, metadata) become visible.
+                # While XPU doesn't have the large fixed cuBLAS workspace of A100, these small overheads prevent reaching the ideal 2.0 ratio.
+                # Observed ~1.27x (158k vs 124k) for model size.
+                # The runtime memory overhead is ~88k for both bf16 and int8wo. Adding this to model size: (158k+88k)/(124k+88k) ≈ 1.15.
+                ("xpu", None): 1.15,
+                # On Ampere, the cuBLAS kernels used for matrix multiplication often allocate a fixed-size workspace.
+                # Since the tiny-flux model weights are likely smaller than or comparable to this workspace, the total memory is dominated by the workspace.
+                ("cuda", 8): 1.02,
+                # On Hopper, TorchAO utilizes newer, highly optimized kernels (via Triton or CUTLASS 3.x) that are designed to be workspace-free or use negligible extra memory.
+                # Additionally, Triton kernels often handle unaligned memory better, avoiding the padding overhead seen on other backends for tiny tensors.
+                # This allows it to achieve the near-ideal 2.0x compression ratio.
+                ("cuda", 9): 2.0,
+            }
+        )
+        expected_memory_saving_ratio = expected_memory_saving_ratios.get_expectation()
         inputs = self.get_dummy_tensor_inputs(device=torch_device)
 
         transformer_bf16 = self.get_dummy_components(None, model_id=model_id)["transformer"]