quick fix to make sure sdxl conversion works

Signed-off-by: Raphael <oOraph@users.noreply.github.com>
Co-authored-by: Raphael <oOraph@users.noreply.github.com>

docs/source/en/api/pipelines/audio_diffusion.mdx

@@ -43,7 +43,7 @@ pipe = DiffusionPipeline.from_pretrained("teticio/audio-diffusion-256").to(devic
 
 output = pipe()
 display(output.images[0])
-display(Audio(output.audios[0], rate=mel.get_sample_rate()))
+display(Audio(output.audios[0], rate=pipe.mel.get_sample_rate()))
 ```
 
 ### Latent Audio Diffusion

docs/source/en/api/pipelines/stable_diffusion/stable_diffusion_xl.mdx

@@ -21,7 +21,7 @@ The abstract of the paper is the following:
 ## Tips
 
 - Stable Diffusion XL works especially well with images between 768 and 1024.
-- Stable Diffusion XL output image can be improved by making use of a refiner as shown below
+- Stable Diffusion XL output image can be improved by making use of a refiner as shown below.
 
 ### Available checkpoints:
 
@@ -40,7 +40,7 @@ pip install safetensors
 pip install invisible-watermark>=2.0
 ```
 
-### *Text-to-Image*
+### Text-to-Image
 
 You can use SDXL as follows for *text-to-image*:
 
@@ -71,6 +71,7 @@ pipe = StableDiffusionXLPipeline.from_pretrained(
 )
 pipe.to("cuda")
 
+use_refiner = True
 refiner = StableDiffusionXLImg2ImgPipeline.from_pretrained(
     "stabilityai/stable-diffusion-xl-refiner-0.9", torch_dtype=torch.float16, use_safetensors=True, variant="fp16"
 )
@@ -82,7 +83,29 @@ image = pipe(prompt=prompt, output_type="latent" if use_refiner else "pil").imag
 image = refiner(prompt=prompt, image=image[None, :]).images[0]
 ```
 
-### Loading single file checkpoitns / original file format
+### Image-to-image 
+
+```py 
+import torch
+from diffusers import StableDiffusionXLImg2ImgPipeline
+from diffusers.utils import load_image
+
+pipe = StableDiffusionXLImg2ImgPipeline.from_pretrained(
+    "stabilityai/stable-diffusion-xl-refiner-0.9", torch_dtype=torch.float16
+)
+pipe = pipe.to("cuda")
+url = "https://huggingface.co/datasets/patrickvonplaten/images/resolve/main/aa_xl/000000009.png"
+
+init_image = load_image(url).convert("RGB")
+prompt = "a photo of an astronaut riding a horse on mars"
+image = pipe(prompt, image=init_image).images[0]
+```
+
+| Original Image | Refined Image |
+|---|---|
+| ![](https://huggingface.co/datasets/diffusers/docs-images/resolve/main/sd_xl/init_image.png) | ![](https://huggingface.co/datasets/diffusers/docs-images/resolve/main/sd_xl/refined_image.png) |
+
+### Loading single file checkpoints / original file format
 
 By making use of [`~diffusers.loaders.FromSingleFileMixin.from_single_file`] you can also load the 
 original file format into `diffusers`:
@@ -127,7 +150,7 @@ You can speed up inference by making use of `torch.compile`. This should give yo
 + refiner.unet = torch.compile(refiner.unet, mode="reduce-overhead", fullgraph=True)
 ```
 
-### Running with `torch` < 2.0
+### Running with `torch` \< 2.0
 
 **Note** that if you want to run Stable Diffusion XL with `torch` < 2.0, please make sure to enable xformers 
 attention:

examples/controlnet/train_controlnet.py

@@ -56,7 +56,7 @@ if is_wandb_available():
     import wandb
 
 # Will error if the minimal version of diffusers is not installed. Remove at your own risks.
-check_min_version("0.18.0.dev0")
+check_min_version("0.18.0")
 
 logger = get_logger(__name__)
 

examples/controlnet/train_controlnet_flax.py

@@ -59,7 +59,7 @@ if is_wandb_available():
     import wandb
 
 # Will error if the minimal version of diffusers is not installed. Remove at your own risks.
-check_min_version("0.18.0.dev0")
+check_min_version("0.18.0")
 
 logger = logging.getLogger(__name__)
 

examples/custom_diffusion/train_custom_diffusion.py

@@ -57,7 +57,7 @@ from diffusers.utils.import_utils import is_xformers_available
 
 
 # Will error if the minimal version of diffusers is not installed. Remove at your own risks.
-check_min_version("0.18.0.dev0")
+check_min_version("0.18.0")
 
 logger = get_logger(__name__)
 

examples/dreambooth/train_dreambooth.py

@@ -59,7 +59,7 @@ if is_wandb_available():
     import wandb
 
 # Will error if the minimal version of diffusers is not installed. Remove at your own risks.
-check_min_version("0.18.0.dev0")
+check_min_version("0.18.0")
 
 logger = get_logger(__name__)
 

examples/dreambooth/train_dreambooth_flax.py

@@ -36,7 +36,7 @@ from diffusers.utils import check_min_version
 
 
 # Will error if the minimal version of diffusers is not installed. Remove at your own risks.
-check_min_version("0.18.0.dev0")
+check_min_version("0.18.0")
 
 # Cache compiled models across invocations of this script.
 cc.initialize_cache(os.path.expanduser("~/.cache/jax/compilation_cache"))

examples/dreambooth/train_dreambooth_lora.py

@@ -65,7 +65,7 @@ from diffusers.utils.import_utils import is_xformers_available
 
 
 # Will error if the minimal version of diffusers is not installed. Remove at your own risks.
-check_min_version("0.18.0.dev0")
+check_min_version("0.18.0")
 
 logger = get_logger(__name__)
 
@@ -436,6 +436,12 @@ def parse_args(input_args=None):
         default=None,
         help="The optional `class_label` conditioning to pass to the unet, available values are `timesteps`.",
     )
+    parser.add_argument(
+        "--rank",
+        type=int,
+        default=4,
+        help=("The dimension of the LoRA update matrices."),
+    )
 
     if input_args is not None:
         args = parser.parse_args(input_args)
@@ -845,7 +851,9 @@ def main(args):
                 LoRAAttnProcessor2_0 if hasattr(F, "scaled_dot_product_attention") else LoRAAttnProcessor
             )
         unet_lora_attn_procs[name] = lora_attn_processor_class(
-            hidden_size=hidden_size, cross_attention_dim=cross_attention_dim
+            hidden_size=hidden_size,
+            cross_attention_dim=cross_attention_dim,
+            rank=args.rank,
         )
 
     unet.set_attn_processor(unet_lora_attn_procs)
@@ -860,7 +868,9 @@ def main(args):
         for name, module in text_encoder.named_modules():
             if name.endswith(TEXT_ENCODER_ATTN_MODULE):
                 text_lora_attn_procs[name] = LoRAAttnProcessor(
-                    hidden_size=module.out_proj.out_features, cross_attention_dim=None
+                    hidden_size=module.out_proj.out_features,
+                    cross_attention_dim=None,
+                    rank=args.rank,
                 )
         text_encoder_lora_layers = AttnProcsLayers(text_lora_attn_procs)
         temp_pipeline = DiffusionPipeline.from_pretrained(

examples/instruct_pix2pix/train_instruct_pix2pix.py

@@ -52,7 +52,7 @@ from diffusers.utils.import_utils import is_xformers_available
 
 
 # Will error if the minimal version of diffusers is not installed. Remove at your own risks.
-check_min_version("0.18.0.dev0")
+check_min_version("0.18.0")
 
 logger = get_logger(__name__, log_level="INFO")
 

examples/text_to_image/train_text_to_image.py

@@ -54,7 +54,7 @@ if is_wandb_available():
 
 
 # Will error if the minimal version of diffusers is not installed. Remove at your own risks.
-check_min_version("0.18.0.dev0")
+check_min_version("0.18.0")
 
 logger = get_logger(__name__, log_level="INFO")
 

examples/text_to_image/train_text_to_image_flax.py

@@ -33,7 +33,7 @@ from diffusers.utils import check_min_version
 
 
 # Will error if the minimal version of diffusers is not installed. Remove at your own risks.
-check_min_version("0.18.0.dev0")
+check_min_version("0.18.0")
 
 logger = logging.getLogger(__name__)
 

examples/text_to_image/train_text_to_image_lora.py

@@ -48,7 +48,7 @@ from diffusers.utils.import_utils import is_xformers_available
 
 
 # Will error if the minimal version of diffusers is not installed. Remove at your own risks.
-check_min_version("0.18.0.dev0")
+check_min_version("0.18.0")
 
 logger = get_logger(__name__, log_level="INFO")
 

examples/textual_inversion/textual_inversion.py

@@ -78,7 +78,7 @@ else:
 
 
 # Will error if the minimal version of diffusers is not installed. Remove at your own risks.
-check_min_version("0.18.0.dev0")
+check_min_version("0.18.0")
 
 logger = get_logger(__name__)
 

examples/textual_inversion/textual_inversion_flax.py

@@ -56,7 +56,7 @@ else:
 # ------------------------------------------------------------------------------
 
 # Will error if the minimal version of diffusers is not installed. Remove at your own risks.
-check_min_version("0.18.0.dev0")
+check_min_version("0.18.0")
 
 logger = logging.getLogger(__name__)
 

examples/unconditional_image_generation/train_unconditional.py

@@ -29,7 +29,7 @@ from diffusers.utils.import_utils import is_xformers_available
 
 
 # Will error if the minimal version of diffusers is not installed. Remove at your own risks.
-check_min_version("0.18.0.dev0")
+check_min_version("0.18.0")
 
 logger = get_logger(__name__, log_level="INFO")
 

setup.py

@@ -232,7 +232,7 @@ install_requires = [
 
 setup(
     name="diffusers",
-    version="0.18.0.dev0",  # expected format is one of x.y.z.dev0, or x.y.z.rc1 or x.y.z (no to dashes, yes to dots)
+    version="0.18.2",  # expected format is one of x.y.z.dev0, or x.y.z.rc1 or x.y.z (no to dashes, yes to dots)
     description="Diffusers",
     long_description=open("README.md", "r", encoding="utf-8").read(),
     long_description_content_type="text/markdown",

src/diffusers/__init__.py

@@ -1,4 +1,4 @@
-__version__ = "0.18.0.dev0"
+__version__ = "0.18.2"
 
 from .configuration_utils import ConfigMixin
 from .utils import (

src/diffusers/configuration_utils.py

@@ -607,7 +607,7 @@ def register_to_config(init):
 
         # Take note of the parameters that were not present in the loaded config
         if len(set(new_kwargs.keys()) - set(init_kwargs)) > 0:
-            new_kwargs["_use_default_values"] = set(new_kwargs.keys()) - set(init_kwargs)
+            new_kwargs["_use_default_values"] = list(set(new_kwargs.keys()) - set(init_kwargs))
 
         new_kwargs = {**config_init_kwargs, **new_kwargs}
         getattr(self, "register_to_config")(**new_kwargs)
@@ -655,7 +655,7 @@ def flax_register_to_config(cls):
 
         # Take note of the parameters that were not present in the loaded config
         if len(set(new_kwargs.keys()) - set(init_kwargs)) > 0:
-            new_kwargs["_use_default_values"] = set(new_kwargs.keys()) - set(init_kwargs)
+            new_kwargs["_use_default_values"] = list(set(new_kwargs.keys()) - set(init_kwargs))
 
         getattr(self, "register_to_config")(**new_kwargs)
         original_init(self, *args, **kwargs)

src/diffusers/loaders.py

@@ -177,7 +177,7 @@ class UNet2DConditionLoadersMixin:
 
         if use_safetensors and not is_safetensors_available():
             raise ValueError(
-                "`use_safetensors`=True but safetensors is not installed. Please install safetensors with `pip install safetenstors"
+                "`use_safetensors`=True but safetensors is not installed. Please install safetensors with `pip install safetensors"
             )
 
         allow_pickle = False
@@ -589,7 +589,7 @@ class TextualInversionLoaderMixin:
 
         if use_safetensors and not is_safetensors_available():
             raise ValueError(
-                "`use_safetensors`=True but safetensors is not installed. Please install safetensors with `pip install safetenstors"
+                "`use_safetensors`=True but safetensors is not installed. Please install safetensors with `pip install safetensors"
             )
 
         allow_pickle = False
@@ -806,7 +806,7 @@ class LoraLoaderMixin:
 
         if use_safetensors and not is_safetensors_available():
             raise ValueError(
-                "`use_safetensors`=True but safetensors is not installed. Please install safetensors with `pip install safetenstors"
+                "`use_safetensors`=True but safetensors is not installed. Please install safetensors with `pip install safetensors"
             )
 
         allow_pickle = False
@@ -1054,7 +1054,7 @@ class LoraLoaderMixin:
 
         if use_safetensors and not is_safetensors_available():
             raise ValueError(
-                "`use_safetensors`=True but safetensors is not installed. Please install safetensors with `pip install safetenstors"
+                "`use_safetensors`=True but safetensors is not installed. Please install safetensors with `pip install safetensors"
             )
 
         allow_pickle = False
@@ -1394,7 +1394,7 @@ class FromSingleFileMixin:
         use_auth_token = kwargs.pop("use_auth_token", None)
         revision = kwargs.pop("revision", None)
         extract_ema = kwargs.pop("extract_ema", False)
-        image_size = kwargs.pop("image_size", 512)
+        image_size = kwargs.pop("image_size", None)
         scheduler_type = kwargs.pop("scheduler_type", "pndm")
         num_in_channels = kwargs.pop("num_in_channels", None)
         upcast_attention = kwargs.pop("upcast_attention", None)

src/diffusers/models/attention_flax.py

@@ -152,6 +152,7 @@ class FlaxAttention(nn.Module):
         self.value = nn.Dense(inner_dim, use_bias=False, dtype=self.dtype, name="to_v")
 
         self.proj_attn = nn.Dense(self.query_dim, dtype=self.dtype, name="to_out_0")
+        self.dropout_layer = nn.Dropout(rate=self.dropout)
 
     def reshape_heads_to_batch_dim(self, tensor):
         batch_size, seq_len, dim = tensor.shape
@@ -214,7 +215,7 @@ class FlaxAttention(nn.Module):
 
         hidden_states = self.reshape_batch_dim_to_heads(hidden_states)
         hidden_states = self.proj_attn(hidden_states)
-        return hidden_states
+        return self.dropout_layer(hidden_states, deterministic=deterministic)
 
 
 class FlaxBasicTransformerBlock(nn.Module):
@@ -260,6 +261,7 @@ class FlaxBasicTransformerBlock(nn.Module):
         self.norm1 = nn.LayerNorm(epsilon=1e-5, dtype=self.dtype)
         self.norm2 = nn.LayerNorm(epsilon=1e-5, dtype=self.dtype)
         self.norm3 = nn.LayerNorm(epsilon=1e-5, dtype=self.dtype)
+        self.dropout_layer = nn.Dropout(rate=self.dropout)
 
     def __call__(self, hidden_states, context, deterministic=True):
         # self attention
@@ -280,7 +282,7 @@ class FlaxBasicTransformerBlock(nn.Module):
         hidden_states = self.ff(self.norm3(hidden_states), deterministic=deterministic)
         hidden_states = hidden_states + residual
 
-        return hidden_states
+        return self.dropout_layer(hidden_states, deterministic=deterministic)
 
 
 class FlaxTransformer2DModel(nn.Module):
@@ -356,6 +358,8 @@ class FlaxTransformer2DModel(nn.Module):
                 dtype=self.dtype,
             )
 
+        self.dropout_layer = nn.Dropout(rate=self.dropout)
+
     def __call__(self, hidden_states, context, deterministic=True):
         batch, height, width, channels = hidden_states.shape
         residual = hidden_states
@@ -378,7 +382,7 @@ class FlaxTransformer2DModel(nn.Module):
             hidden_states = self.proj_out(hidden_states)
 
         hidden_states = hidden_states + residual
-        return hidden_states
+        return self.dropout_layer(hidden_states, deterministic=deterministic)
 
 
 class FlaxFeedForward(nn.Module):
@@ -409,7 +413,7 @@ class FlaxFeedForward(nn.Module):
         self.net_2 = nn.Dense(self.dim, dtype=self.dtype)
 
     def __call__(self, hidden_states, deterministic=True):
-        hidden_states = self.net_0(hidden_states)
+        hidden_states = self.net_0(hidden_states, deterministic=deterministic)
         hidden_states = self.net_2(hidden_states)
         return hidden_states
 
@@ -434,8 +438,9 @@ class FlaxGEGLU(nn.Module):
     def setup(self):
         inner_dim = self.dim * 4
         self.proj = nn.Dense(inner_dim * 2, dtype=self.dtype)
+        self.dropout_layer = nn.Dropout(rate=self.dropout)
 
     def __call__(self, hidden_states, deterministic=True):
         hidden_states = self.proj(hidden_states)
         hidden_linear, hidden_gelu = jnp.split(hidden_states, 2, axis=2)
-        return hidden_linear * nn.gelu(hidden_gelu)
+        return self.dropout_layer(hidden_linear * nn.gelu(hidden_gelu), deterministic=deterministic)

src/diffusers/models/modeling_utils.py

@@ -456,7 +456,7 @@ class ModelMixin(torch.nn.Module):
 
         if use_safetensors and not is_safetensors_available():
             raise ValueError(
-                "`use_safetensors`=True but safetensors is not installed. Please install safetensors with `pip install safetenstors"
+                "`use_safetensors`=True but safetensors is not installed. Please install safetensors with `pip install safetensors"
             )
 
         allow_pickle = False

src/diffusers/pipelines/pipeline_utils.py

@@ -204,7 +204,7 @@ def variant_compatible_siblings(filenames, variant=None) -> Union[List[os.PathLi
     transformers_index_format = r"\d{5}-of-\d{5}"
 
     if variant is not None:
-        # `diffusion_pytorch_model.fp16.bin` as well as `model.fp16-00001-of-00002.safetenstors`
+        # `diffusion_pytorch_model.fp16.bin` as well as `model.fp16-00001-of-00002.safetensors`
         variant_file_re = re.compile(
             rf"({'|'.join(weight_prefixes)})\.({variant}|{variant}-{transformers_index_format})\.({'|'.join(weight_suffixs)})$"
         )
@@ -213,7 +213,7 @@ def variant_compatible_siblings(filenames, variant=None) -> Union[List[os.PathLi
             rf"({'|'.join(weight_prefixes)})\.({'|'.join(weight_suffixs)})\.index\.{variant}\.json$"
         )
 
-    # `diffusion_pytorch_model.bin` as well as `model-00001-of-00002.safetenstors`
+    # `diffusion_pytorch_model.bin` as well as `model-00001-of-00002.safetensors`
     non_variant_file_re = re.compile(
         rf"({'|'.join(weight_prefixes)})(-{transformers_index_format})?\.({'|'.join(weight_suffixs)})$"
     )
@@ -1168,7 +1168,7 @@ class DiffusionPipeline(ConfigMixin):
 
         if use_safetensors and not is_safetensors_available():
             raise ValueError(
-                "`use_safetensors`=True but safetensors is not installed. Please install safetensors with `pip install safetenstors"
+                "`use_safetensors`=True but safetensors is not installed. Please install safetensors with `pip install safetensors"
             )
 
         allow_pickle = False
@@ -1213,6 +1213,15 @@ class DiffusionPipeline(ConfigMixin):
             filenames = {sibling.rfilename for sibling in info.siblings}
             model_filenames, variant_filenames = variant_compatible_siblings(filenames, variant=variant)
 
+            if len(variant_filenames) == 0 and variant is not None:
+                deprecation_message = (
+                    f"You are trying to load the model files of the `variant={variant}`, but no such modeling files are available."
+                    f"The default model files: {model_filenames} will be loaded instead. Make sure to not load from `variant={variant}`"
+                    "if such variant modeling files are not available. Doing so will lead to an error in v0.22.0 as defaulting to non-variant"
+                    "modeling files is deprecated."
+                )
+                deprecate("no variant default", "0.22.0", deprecation_message, standard_warn=False)
+
             # remove ignored filenames
             model_filenames = set(model_filenames) - set(ignore_filenames)
             variant_filenames = set(variant_filenames) - set(ignore_filenames)
@@ -1302,7 +1311,7 @@ class DiffusionPipeline(ConfigMixin):
             snapshot_folder = Path(config_file).parent
             pipeline_is_cached = all((snapshot_folder / f).is_file() for f in expected_files)
 
-            if pipeline_is_cached:
+            if pipeline_is_cached and not force_download:
                 # if the pipeline is cached, we can directly return it
                 # else call snapshot_download
                 return snapshot_folder

src/diffusers/pipelines/stable_diffusion/convert_from_ckpt.py

@@ -24,6 +24,7 @@ from transformers import (
     AutoFeatureExtractor,
     BertTokenizerFast,
     CLIPImageProcessor,
+    CLIPTextConfig,
     CLIPTextModel,
     CLIPTextModelWithProjection,
     CLIPTokenizer,
@@ -48,7 +49,7 @@ from ...schedulers import (
     PNDMScheduler,
     UnCLIPScheduler,
 )
-from ...utils import is_omegaconf_available, is_safetensors_available, logging
+from ...utils import is_accelerate_available, is_omegaconf_available, is_safetensors_available, logging
 from ...utils.import_utils import BACKENDS_MAPPING
 from ..latent_diffusion.pipeline_latent_diffusion import LDMBertConfig, LDMBertModel
 from ..paint_by_example import PaintByExampleImageEncoder
@@ -57,6 +58,10 @@ from .safety_checker import StableDiffusionSafetyChecker
 from .stable_unclip_image_normalizer import StableUnCLIPImageNormalizer
 
 
+if is_accelerate_available():
+    from accelerate import init_empty_weights
+    from accelerate.utils import set_module_tensor_to_device
+
 logger = logging.get_logger(__name__)  # pylint: disable=invalid-name
 
 
@@ -391,8 +396,8 @@ def convert_ldm_unet_checkpoint(
 
         # at least a 100 parameters have to start with `model_ema` in order for the checkpoint to be EMA
         if sum(k.startswith("model_ema") for k in keys) > 100 and extract_ema:
-            print(f"Checkpoint {path} has both EMA and non-EMA weights.")
-            print(
+            logger.warning(f"Checkpoint {path} has both EMA and non-EMA weights.")
+            logger.warning(
                 "In this conversion only the EMA weights are extracted. If you want to instead extract the non-EMA"
                 " weights (useful to continue fine-tuning), please make sure to remove the `--extract_ema` flag."
             )
@@ -402,7 +407,7 @@ def convert_ldm_unet_checkpoint(
                     unet_state_dict[key.replace(unet_key, "")] = checkpoint.pop(flat_ema_key)
         else:
             if sum(k.startswith("model_ema") for k in keys) > 100:
-                print(
+                logger.warning(
                     "In this conversion only the non-EMA weights are extracted. If you want to instead extract the EMA"
                     " weights (usually better for inference), please make sure to add the `--extract_ema` flag."
                 )
@@ -770,11 +775,12 @@ def convert_ldm_bert_checkpoint(checkpoint, config):
 
 
 def convert_ldm_clip_checkpoint(checkpoint, local_files_only=False, text_encoder=None):
-    text_model = (
-        CLIPTextModel.from_pretrained("openai/clip-vit-large-patch14", local_files_only=local_files_only)
-        if text_encoder is None
-        else text_encoder
-    )
+    if text_encoder is None:
+        config_name = "openai/clip-vit-large-patch14"
+        config = CLIPTextConfig.from_pretrained(config_name)
+
+        with init_empty_weights():
+            text_model = CLIPTextModel(config)
 
     keys = list(checkpoint.keys())
 
@@ -787,7 +793,8 @@ def convert_ldm_clip_checkpoint(checkpoint, local_files_only=False, text_encoder
             if key.startswith(prefix):
                 text_model_dict[key[len(prefix + ".") :]] = checkpoint[key]
 
-    text_model.load_state_dict(text_model_dict)
+    for param_name, param in text_model_dict.items():
+        set_module_tensor_to_device(text_model, param_name, "cpu", value=param)
 
     return text_model
 
@@ -884,14 +891,26 @@ def convert_paint_by_example_checkpoint(checkpoint):
     return model
 
 
-def convert_open_clip_checkpoint(checkpoint, prefix="cond_stage_model.model."):
+def convert_open_clip_checkpoint(
+    checkpoint, config_name, prefix="cond_stage_model.model.", has_projection=False, **config_kwargs
+):
     # text_model = CLIPTextModel.from_pretrained("stabilityai/stable-diffusion-2", subfolder="text_encoder")
-    text_model = CLIPTextModelWithProjection.from_pretrained(
-        "laion/CLIP-ViT-bigG-14-laion2B-39B-b160k", projection_dim=1280
-    )
+    # text_model = CLIPTextModelWithProjection.from_pretrained(
+    #    "laion/CLIP-ViT-bigG-14-laion2B-39B-b160k", projection_dim=1280
+    # )
+    config = CLIPTextConfig.from_pretrained(config_name, **config_kwargs)
+
+    with init_empty_weights():
+        text_model = CLIPTextModelWithProjection(config) if has_projection else CLIPTextModel(config)
 
     keys = list(checkpoint.keys())
 
+    keys_to_ignore = []
+    if config_name == "stabilityai/stable-diffusion-2" and config.num_hidden_layers == 23:
+        # make sure to remove all keys > 22
+        keys_to_ignore += [k for k in keys if k.startswith("cond_stage_model.model.transformer.resblocks.23")]
+        keys_to_ignore += ["cond_stage_model.model.text_projection"]
+
     text_model_dict = {}
 
     if prefix + "text_projection" in checkpoint:
@@ -902,8 +921,8 @@ def convert_open_clip_checkpoint(checkpoint, prefix="cond_stage_model.model."):
     text_model_dict["text_model.embeddings.position_ids"] = text_model.text_model.embeddings.get_buffer("position_ids")
 
     for key in keys:
-        # if "resblocks.23" in key:  # Diffusers drops the final layer and only uses the penultimate layer
-        #     continue
+        if key in keys_to_ignore:
+            continue
         if key[len(prefix) :] in textenc_conversion_map:
             if key.endswith("text_projection"):
                 value = checkpoint[key].T
@@ -931,7 +950,8 @@ def convert_open_clip_checkpoint(checkpoint, prefix="cond_stage_model.model."):
 
                 text_model_dict[new_key] = checkpoint[key]
 
-    text_model.load_state_dict(text_model_dict)
+    for param_name, param in text_model_dict.items():
+        set_module_tensor_to_device(text_model, param_name, "cpu", value=param)
 
     return text_model
 
@@ -1061,7 +1081,7 @@ def convert_controlnet_checkpoint(
 def download_from_original_stable_diffusion_ckpt(
     checkpoint_path: str,
     original_config_file: str = None,
-    image_size: int = 512,
+    image_size: Optional[int] = None,
     prediction_type: str = None,
     model_type: str = None,
     extract_ema: bool = False,
@@ -1144,6 +1164,7 @@ def download_from_original_stable_diffusion_ckpt(
         LDMTextToImagePipeline,
         PaintByExamplePipeline,
         StableDiffusionControlNetPipeline,
+        StableDiffusionInpaintPipeline,
         StableDiffusionPipeline,
         StableDiffusionXLImg2ImgPipeline,
         StableDiffusionXLPipeline,
@@ -1166,12 +1187,9 @@ def download_from_original_stable_diffusion_ckpt(
         if not is_safetensors_available():
             raise ValueError(BACKENDS_MAPPING["safetensors"][1])
 
-        from safetensors import safe_open
+        from safetensors.torch import load_file as safe_load
 
-        checkpoint = {}
-        with safe_open(checkpoint_path, framework="pt", device="cpu") as f:
-            for key in f.keys():
-                checkpoint[key] = f.get_tensor(key)
+        checkpoint = safe_load(checkpoint_path, device="cpu")
     else:
         if device is None:
             device = "cuda" if torch.cuda.is_available() else "cpu"
@@ -1183,7 +1201,7 @@ def download_from_original_stable_diffusion_ckpt(
     if "global_step" in checkpoint:
         global_step = checkpoint["global_step"]
     else:
-        print("global_step key not found in model")
+        logger.debug("global_step key not found in model")
         global_step = None
 
     # NOTE: this while loop isn't great but this controlnet checkpoint has one additional
@@ -1230,8 +1248,15 @@ def download_from_original_stable_diffusion_ckpt(
             model_type = "SDXL"
         else:
             model_type = "SDXL-Refiner"
+        if image_size is None:
+            image_size = 1024
 
-    if num_in_channels is not None:
+    if num_in_channels is None and pipeline_class == StableDiffusionInpaintPipeline:
+        num_in_channels = 9
+    elif num_in_channels is None:
+        num_in_channels = 4
+
+    if "unet_config" in original_config.model.params:
         original_config["model"]["params"]["unet_config"]["params"]["in_channels"] = num_in_channels
 
     if (
@@ -1263,7 +1288,6 @@ def download_from_original_stable_diffusion_ckpt(
     num_train_timesteps = getattr(original_config.model.params, "timesteps", None) or 1000
 
     if model_type in ["SDXL", "SDXL-Refiner"]:
-        image_size = 1024
         scheduler_dict = {
             "beta_schedule": "scaled_linear",
             "beta_start": 0.00085,
@@ -1279,7 +1303,6 @@ def download_from_original_stable_diffusion_ckpt(
         }
         scheduler = EulerDiscreteScheduler.from_config(scheduler_dict)
         scheduler_type = "euler"
-        vae_path = "stabilityai/sdxl-vae"
     else:
         beta_start = getattr(original_config.model.params, "linear_start", None) or 0.02
         beta_end = getattr(original_config.model.params, "linear_end", None) or 0.085
@@ -1318,25 +1341,45 @@ def download_from_original_stable_diffusion_ckpt(
     # Convert the UNet2DConditionModel model.
     unet_config = create_unet_diffusers_config(original_config, image_size=image_size)
     unet_config["upcast_attention"] = upcast_attention
-    unet = UNet2DConditionModel(**unet_config)
+    with init_empty_weights():
+        unet = UNet2DConditionModel(**unet_config)
 
     converted_unet_checkpoint = convert_ldm_unet_checkpoint(
         checkpoint, unet_config, path=checkpoint_path, extract_ema=extract_ema
     )
-    unet.load_state_dict(converted_unet_checkpoint)
+
+    for param_name, param in converted_unet_checkpoint.items():
+        set_module_tensor_to_device(unet, param_name, "cpu", value=param)
 
     # Convert the VAE model.
     if vae_path is None:
         vae_config = create_vae_diffusers_config(original_config, image_size=image_size)
         converted_vae_checkpoint = convert_ldm_vae_checkpoint(checkpoint, vae_config)
 
-        vae = AutoencoderKL(**vae_config)
-        vae.load_state_dict(converted_vae_checkpoint)
+        if (
+            "model" in original_config
+            and "params" in original_config.model
+            and "scale_factor" in original_config.model.params
+        ):
+            vae_scaling_factor = original_config.model.params.scale_factor
+        else:
+            vae_scaling_factor = 0.18215  # default SD scaling factor
+
+        vae_config["scaling_factor"] = vae_scaling_factor
+
+        with init_empty_weights():
+            vae = AutoencoderKL(**vae_config)
+
+        for param_name, param in converted_vae_checkpoint.items():
+            set_module_tensor_to_device(vae, param_name, "cpu", value=param)
     else:
         vae = AutoencoderKL.from_pretrained(vae_path)
 
     if model_type == "FrozenOpenCLIPEmbedder":
-        text_model = convert_open_clip_checkpoint(checkpoint)
+        config_name = "stabilityai/stable-diffusion-2"
+        config_kwargs = {"subfolder": "text_encoder"}
+
+        text_model = convert_open_clip_checkpoint(checkpoint, config_name, **config_kwargs)
         tokenizer = CLIPTokenizer.from_pretrained("stabilityai/stable-diffusion-2", subfolder="tokenizer")
 
         if stable_unclip is None:
@@ -1469,7 +1512,12 @@ def download_from_original_stable_diffusion_ckpt(
             tokenizer = CLIPTokenizer.from_pretrained("openai/clip-vit-large-patch14")
             text_encoder = convert_ldm_clip_checkpoint(checkpoint, local_files_only=local_files_only)
             tokenizer_2 = CLIPTokenizer.from_pretrained("laion/CLIP-ViT-bigG-14-laion2B-39B-b160k", pad_token="!")
-            text_encoder_2 = convert_open_clip_checkpoint(checkpoint, prefix="conditioner.embedders.1.model.")
+
+            config_name = "laion/CLIP-ViT-bigG-14-laion2B-39B-b160k"
+            config_kwargs = {"projection_dim": 1280}
+            text_encoder_2 = convert_open_clip_checkpoint(
+                checkpoint, config_name, prefix="conditioner.embedders.1.model.", has_projection=True, **config_kwargs
+            )
 
             pipe = StableDiffusionXLPipeline(
                 vae=vae,
@@ -1485,7 +1533,12 @@ def download_from_original_stable_diffusion_ckpt(
             tokenizer = None
             text_encoder = None
             tokenizer_2 = CLIPTokenizer.from_pretrained("laion/CLIP-ViT-bigG-14-laion2B-39B-b160k", pad_token="!")
-            text_encoder_2 = convert_open_clip_checkpoint(checkpoint, prefix="conditioner.embedders.0.model.")
+
+            config_name = "laion/CLIP-ViT-bigG-14-laion2B-39B-b160k"
+            config_kwargs = {"projection_dim": 1280}
+            text_encoder_2 = convert_open_clip_checkpoint(
+                checkpoint, config_name, prefix="conditioner.embedders.0.model.", has_projection=True, **config_kwargs
+            )
 
             pipe = StableDiffusionXLImg2ImgPipeline(
                 vae=vae,

src/diffusers/pipelines/stable_diffusion/pipeline_stable_diffusion_inpaint.py

@@ -24,7 +24,7 @@ from transformers import CLIPImageProcessor, CLIPTextModel, CLIPTokenizer
 
 from ...configuration_utils import FrozenDict
 from ...image_processor import VaeImageProcessor
-from ...loaders import LoraLoaderMixin, TextualInversionLoaderMixin
+from ...loaders import FromSingleFileMixin, LoraLoaderMixin, TextualInversionLoaderMixin
 from ...models import AutoencoderKL, UNet2DConditionModel
 from ...schedulers import KarrasDiffusionSchedulers
 from ...utils import deprecate, is_accelerate_available, is_accelerate_version, logging, randn_tensor
@@ -153,7 +153,9 @@ def prepare_mask_and_masked_image(image, mask, height, width, return_image: bool
     return mask, masked_image
 
 
-class StableDiffusionInpaintPipeline(DiffusionPipeline, TextualInversionLoaderMixin, LoraLoaderMixin):
+class StableDiffusionInpaintPipeline(
+    DiffusionPipeline, TextualInversionLoaderMixin, LoraLoaderMixin, FromSingleFileMixin
+):
     r"""
     Pipeline for text-guided image inpainting using Stable Diffusion.
 

src/diffusers/pipelines/stable_diffusion/pipeline_stable_diffusion_upscale.py

@@ -748,15 +748,19 @@ class StableDiffusionUpscalePipeline(DiffusionPipeline, TextualInversionLoaderMi
         # make sure the VAE is in float32 mode, as it overflows in float16
         self.vae.to(dtype=torch.float32)
 
-        use_torch_2_0_or_xformers = self.vae.decoder.mid_block.attentions[0].processor in [
-            AttnProcessor2_0,
-            XFormersAttnProcessor,
-            LoRAXFormersAttnProcessor,
-            LoRAAttnProcessor2_0,
-        ]
+        use_torch_2_0_or_xformers = isinstance(
+            self.vae.decoder.mid_block.attentions[0].processor,
+            (
+                AttnProcessor2_0,
+                XFormersAttnProcessor,
+                LoRAXFormersAttnProcessor,
+                LoRAAttnProcessor2_0,
+            ),
+        )
+
         # if xformers or torch_2_0 is used attention block does not need
         # to be in float32 which can save lots of memory
-        if not use_torch_2_0_or_xformers:
+        if use_torch_2_0_or_xformers:
             self.vae.post_quant_conv.to(latents.dtype)
             self.vae.decoder.conv_in.to(latents.dtype)
             self.vae.decoder.mid_block.to(latents.dtype)

src/diffusers/pipelines/stable_diffusion_xl/__init__.py

@@ -8,7 +8,6 @@ from ...utils import BaseOutput, is_invisible_watermark_available, is_torch_avai
 
 
 @dataclass
-# Copied from diffusers.pipelines.stable_diffusion.__init__.StableDiffusionPipelineOutput with StableDiffusion->StableDiffusionXL
 class StableDiffusionXLPipelineOutput(BaseOutput):
     """
     Output class for Stable Diffusion pipelines.
@@ -17,13 +16,9 @@ class StableDiffusionXLPipelineOutput(BaseOutput):
         images (`List[PIL.Image.Image]` or `np.ndarray`)
             List of denoised PIL images of length `batch_size` or numpy array of shape `(batch_size, height, width,
             num_channels)`. PIL images or numpy array present the denoised images of the diffusion pipeline.
-        nsfw_content_detected (`List[bool]`)
-            List of flags denoting whether the corresponding generated image likely represents "not-safe-for-work"
-            (nsfw) content, or `None` if safety checking could not be performed.
     """
 
     images: Union[List[PIL.Image.Image], np.ndarray]
-    nsfw_content_detected: Optional[List[bool]]
 
 
 if is_transformers_available() and is_torch_available() and is_invisible_watermark_available():

src/diffusers/pipelines/stable_diffusion_xl/pipeline_stable_diffusion_xl.py

@@ -129,9 +129,11 @@ class StableDiffusionXLPipeline(DiffusionPipeline, FromSingleFileMixin):
         self.register_to_config(force_zeros_for_empty_prompt=force_zeros_for_empty_prompt)
         self.vae_scale_factor = 2 ** (len(self.vae.config.block_out_channels) - 1)
         self.image_processor = VaeImageProcessor(vae_scale_factor=self.vae_scale_factor)
+        self.default_sample_size = self.unet.config.sample_size
 
         self.watermark = StableDiffusionXLWatermarker()
 
+    # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.enable_vae_slicing
     def enable_vae_slicing(self):
         r"""
         Enable sliced VAE decoding.
@@ -141,6 +143,7 @@ class StableDiffusionXLPipeline(DiffusionPipeline, FromSingleFileMixin):
         """
         self.vae.enable_slicing()
 
+    # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.disable_vae_slicing
     def disable_vae_slicing(self):
         r"""
         Disable sliced VAE decoding. If `enable_vae_slicing` was previously invoked, this method will go back to
@@ -148,6 +151,7 @@ class StableDiffusionXLPipeline(DiffusionPipeline, FromSingleFileMixin):
         """
         self.vae.disable_slicing()
 
+    # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.enable_vae_tiling
     def enable_vae_tiling(self):
         r"""
         Enable tiled VAE decoding.
@@ -157,6 +161,7 @@ class StableDiffusionXLPipeline(DiffusionPipeline, FromSingleFileMixin):
         """
         self.vae.enable_tiling()
 
+    # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.disable_vae_tiling
     def disable_vae_tiling(self):
         r"""
         Disable tiled VAE decoding. If `enable_vae_tiling` was previously invoked, this method will go back to
@@ -183,7 +188,7 @@ class StableDiffusionXLPipeline(DiffusionPipeline, FromSingleFileMixin):
             self.to("cpu", silence_dtype_warnings=True)
             torch.cuda.empty_cache()  # otherwise we don't see the memory savings (but they probably exist)
 
-        for cpu_offloaded_model in [self.unet, self.text_encoder, self.vae]:
+        for cpu_offloaded_model in [self.unet, self.text_encoder, self.text_encoder_2, self.vae]:
             cpu_offload(cpu_offloaded_model, device)
 
     def enable_model_cpu_offload(self, gpu_id=0):
@@ -217,6 +222,7 @@ class StableDiffusionXLPipeline(DiffusionPipeline, FromSingleFileMixin):
         self.final_offload_hook = hook
 
     @property
+    # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline._execution_device
     def _execution_device(self):
         r"""
         Returns the device on which the pipeline's models will be executed. After calling
@@ -237,12 +243,14 @@ class StableDiffusionXLPipeline(DiffusionPipeline, FromSingleFileMixin):
     def encode_prompt(
         self,
         prompt,
-        device,
-        num_images_per_prompt,
-        do_classifier_free_guidance,
+        device: Optional[torch.device] = None,
+        num_images_per_prompt: int = 1,
+        do_classifier_free_guidance: bool = True,
         negative_prompt=None,
         prompt_embeds: Optional[torch.FloatTensor] = None,
         negative_prompt_embeds: Optional[torch.FloatTensor] = None,
+        pooled_prompt_embeds: Optional[torch.FloatTensor] = None,
+        negative_pooled_prompt_embeds: Optional[torch.FloatTensor] = None,
         lora_scale: Optional[float] = None,
     ):
         r"""
@@ -268,9 +276,18 @@ class StableDiffusionXLPipeline(DiffusionPipeline, FromSingleFileMixin):
                 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.FloatTensor`, *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.FloatTensor`, *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.
         """
+        device = device or self._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(self, LoraLoaderMixin):
@@ -399,6 +416,7 @@ class StableDiffusionXLPipeline(DiffusionPipeline, FromSingleFileMixin):
 
             negative_prompt_embeds = torch.concat(negative_prompt_embeds_list, dim=-1)
 
+        bs_embed = pooled_prompt_embeds.shape[0]
         pooled_prompt_embeds = pooled_prompt_embeds.repeat(1, num_images_per_prompt).view(
             bs_embed * num_images_per_prompt, -1
         )
@@ -408,20 +426,7 @@ class StableDiffusionXLPipeline(DiffusionPipeline, FromSingleFileMixin):
 
         return prompt_embeds, negative_prompt_embeds, pooled_prompt_embeds, negative_pooled_prompt_embeds
 
-    def run_safety_checker(self, image, device, dtype):
-        if self.safety_checker is None:
-            has_nsfw_concept = None
-        else:
-            if torch.is_tensor(image):
-                feature_extractor_input = self.image_processor.postprocess(image, output_type="pil")
-            else:
-                feature_extractor_input = self.image_processor.numpy_to_pil(image)
-            safety_checker_input = self.feature_extractor(feature_extractor_input, return_tensors="pt").to(device)
-            image, has_nsfw_concept = self.safety_checker(
-                images=image, clip_input=safety_checker_input.pixel_values.to(dtype)
-            )
-        return image, has_nsfw_concept
-
+    # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.prepare_extra_step_kwargs
     def prepare_extra_step_kwargs(self, generator, eta):
         # prepare extra kwargs for the scheduler step, since not all schedulers have the same signature
         # eta (η) is only used with the DDIMScheduler, it will be ignored for other schedulers.
@@ -448,6 +453,8 @@ class StableDiffusionXLPipeline(DiffusionPipeline, FromSingleFileMixin):
         negative_prompt=None,
         prompt_embeds=None,
         negative_prompt_embeds=None,
+        pooled_prompt_embeds=None,
+        negative_pooled_prompt_embeds=None,
     ):
         if height % 8 != 0 or width % 8 != 0:
             raise ValueError(f"`height` and `width` have to be divisible by 8 but are {height} and {width}.")
@@ -486,6 +493,17 @@ class StableDiffusionXLPipeline(DiffusionPipeline, FromSingleFileMixin):
                     f" {negative_prompt_embeds.shape}."
                 )
 
+        if prompt_embeds is not None and pooled_prompt_embeds is None:
+            raise ValueError(
+                "If `prompt_embeds` are provided, `pooled_prompt_embeds` also have to be passed. Make sure to generate `pooled_prompt_embeds` from the same text encoder that was used to generate `prompt_embeds`."
+            )
+
+        if negative_prompt_embeds is not None and negative_pooled_prompt_embeds is None:
+            raise ValueError(
+                "If `negative_prompt_embeds` are provided, `negative_pooled_prompt_embeds` also have to be passed. Make sure to generate `negative_pooled_prompt_embeds` from the same text encoder that was used to generate `negative_prompt_embeds`."
+            )
+
+    # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.prepare_latents
     def prepare_latents(self, batch_size, num_channels_latents, height, width, dtype, device, generator, latents=None):
         shape = (batch_size, num_channels_latents, height // self.vae_scale_factor, width // self.vae_scale_factor)
         if isinstance(generator, list) and len(generator) != batch_size:
@@ -535,6 +553,8 @@ class StableDiffusionXLPipeline(DiffusionPipeline, FromSingleFileMixin):
         latents: Optional[torch.FloatTensor] = None,
         prompt_embeds: Optional[torch.FloatTensor] = None,
         negative_prompt_embeds: Optional[torch.FloatTensor] = None,
+        pooled_prompt_embeds: Optional[torch.FloatTensor] = None,
+        negative_pooled_prompt_embeds: Optional[torch.FloatTensor] = None,
         output_type: Optional[str] = "pil",
         return_dict: bool = True,
         callback: Optional[Callable[[int, int, torch.FloatTensor], None]] = None,
@@ -588,6 +608,13 @@ class StableDiffusionXLPipeline(DiffusionPipeline, FromSingleFileMixin):
                 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.FloatTensor`, *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.FloatTensor`, *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.
             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`.
@@ -626,15 +653,23 @@ class StableDiffusionXLPipeline(DiffusionPipeline, FromSingleFileMixin):
             "not-safe-for-work" (nsfw) content, according to the `safety_checker`.
         """
         # 0. Default height and width to unet
-        height = height or self.unet.config.sample_size * self.vae_scale_factor
-        width = width or self.unet.config.sample_size * self.vae_scale_factor
+        height = height or self.default_sample_size * self.vae_scale_factor
+        width = width or self.default_sample_size * self.vae_scale_factor
 
         original_size = original_size or (height, width)
         target_size = target_size or (height, width)
 
         # 1. Check inputs. Raise error if not correct
         self.check_inputs(
-            prompt, height, width, callback_steps, negative_prompt, prompt_embeds, negative_prompt_embeds
+            prompt,
+            height,
+            width,
+            callback_steps,
+            negative_prompt,
+            prompt_embeds,
+            negative_prompt_embeds,
+            pooled_prompt_embeds,
+            negative_pooled_prompt_embeds,
         )
 
         # 2. Define call parameters
@@ -669,6 +704,8 @@ class StableDiffusionXLPipeline(DiffusionPipeline, FromSingleFileMixin):
             negative_prompt,
             prompt_embeds=prompt_embeds,
             negative_prompt_embeds=negative_prompt_embeds,
+            pooled_prompt_embeds=pooled_prompt_embeds,
+            negative_pooled_prompt_embeds=negative_pooled_prompt_embeds,
             lora_scale=text_encoder_lora_scale,
         )
 
@@ -749,15 +786,18 @@ class StableDiffusionXLPipeline(DiffusionPipeline, FromSingleFileMixin):
         # make sure the VAE is in float32 mode, as it overflows in float16
         self.vae.to(dtype=torch.float32)
 
-        use_torch_2_0_or_xformers = self.vae.decoder.mid_block.attentions[0].processor in [
-            AttnProcessor2_0,
-            XFormersAttnProcessor,
-            LoRAXFormersAttnProcessor,
-            LoRAAttnProcessor2_0,
-        ]
+        use_torch_2_0_or_xformers = isinstance(
+            self.vae.decoder.mid_block.attentions[0].processor,
+            (
+                AttnProcessor2_0,
+                XFormersAttnProcessor,
+                LoRAXFormersAttnProcessor,
+                LoRAAttnProcessor2_0,
+            ),
+        )
         # if xformers or torch_2_0 is used attention block does not need
         # to be in float32 which can save lots of memory
-        if not use_torch_2_0_or_xformers:
+        if use_torch_2_0_or_xformers:
             self.vae.post_quant_conv.to(latents.dtype)
             self.vae.decoder.conv_in.to(latents.dtype)
             self.vae.decoder.mid_block.to(latents.dtype)
@@ -765,27 +805,19 @@ class StableDiffusionXLPipeline(DiffusionPipeline, FromSingleFileMixin):
             latents = latents.float()
 
         if not output_type == "latent":
-            # CHECK there is problem here (PVP)
             image = self.vae.decode(latents / self.vae.config.scaling_factor, return_dict=False)[0]
-            has_nsfw_concept = None
         else:
             image = latents
-            has_nsfw_concept = None
-            return StableDiffusionXLPipelineOutput(images=image, nsfw_content_detected=None)
-
-        if has_nsfw_concept is None:
-            do_denormalize = [True] * image.shape[0]
-        else:
-            do_denormalize = [not has_nsfw for has_nsfw in has_nsfw_concept]
+            return StableDiffusionXLPipelineOutput(images=image)
 
         image = self.watermark.apply_watermark(image)
-        image = self.image_processor.postprocess(image, output_type=output_type, do_denormalize=do_denormalize)
+        image = self.image_processor.postprocess(image, output_type=output_type)
 
         # Offload last model to CPU
         if hasattr(self, "final_offload_hook") and self.final_offload_hook is not None:
             self.final_offload_hook.offload()
 
         if not return_dict:
-            return (image, has_nsfw_concept)
+            return (image,)
 
-        return StableDiffusionXLPipelineOutput(images=image, nsfw_content_detected=has_nsfw_concept)
+        return StableDiffusionXLPipelineOutput(images=image)

src/diffusers/pipelines/stable_diffusion_xl/pipeline_stable_diffusion_xl_img2img.py

@@ -140,6 +140,7 @@ class StableDiffusionXLImg2ImgPipeline(DiffusionPipeline, FromSingleFileMixin):
 
         self.watermark = StableDiffusionXLWatermarker()
 
+    # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.enable_vae_slicing
     def enable_vae_slicing(self):
         r"""
         Enable sliced VAE decoding.
@@ -149,6 +150,7 @@ class StableDiffusionXLImg2ImgPipeline(DiffusionPipeline, FromSingleFileMixin):
         """
         self.vae.enable_slicing()
 
+    # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.disable_vae_slicing
     def disable_vae_slicing(self):
         r"""
         Disable sliced VAE decoding. If `enable_vae_slicing` was previously invoked, this method will go back to
@@ -156,6 +158,7 @@ class StableDiffusionXLImg2ImgPipeline(DiffusionPipeline, FromSingleFileMixin):
         """
         self.vae.disable_slicing()
 
+    # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.enable_vae_tiling
     def enable_vae_tiling(self):
         r"""
         Enable tiled VAE decoding.
@@ -165,6 +168,7 @@ class StableDiffusionXLImg2ImgPipeline(DiffusionPipeline, FromSingleFileMixin):
         """
         self.vae.enable_tiling()
 
+    # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.disable_vae_tiling
     def disable_vae_tiling(self):
         r"""
         Disable tiled VAE decoding. If `enable_vae_tiling` was previously invoked, this method will go back to
@@ -172,6 +176,7 @@ class StableDiffusionXLImg2ImgPipeline(DiffusionPipeline, FromSingleFileMixin):
         """
         self.vae.disable_tiling()
 
+    # Copied from diffusers.pipelines.stable_diffusion_xl.pipeline_stable_diffusion_xl.StableDiffusionXLPipeline.enable_sequential_cpu_offload
     def enable_sequential_cpu_offload(self, gpu_id=0):
         r"""
         Offloads all models to CPU using accelerate, significantly reducing memory usage. When called, unet,
@@ -191,9 +196,10 @@ class StableDiffusionXLImg2ImgPipeline(DiffusionPipeline, FromSingleFileMixin):
             self.to("cpu", silence_dtype_warnings=True)
             torch.cuda.empty_cache()  # otherwise we don't see the memory savings (but they probably exist)
 
-        for cpu_offloaded_model in [self.unet, self.text_encoder, self.vae]:
+        for cpu_offloaded_model in [self.unet, self.text_encoder, self.text_encoder_2, self.vae]:
             cpu_offload(cpu_offloaded_model, device)
 
+    # Copied from diffusers.pipelines.stable_diffusion_xl.pipeline_stable_diffusion_xl.StableDiffusionXLPipeline.enable_model_cpu_offload
     def enable_model_cpu_offload(self, gpu_id=0):
         r"""
         Offloads all models to CPU using accelerate, reducing memory usage with a low impact on performance. Compared
@@ -225,6 +231,7 @@ class StableDiffusionXLImg2ImgPipeline(DiffusionPipeline, FromSingleFileMixin):
         self.final_offload_hook = hook
 
     @property
+    # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline._execution_device
     def _execution_device(self):
         r"""
         Returns the device on which the pipeline's models will be executed. After calling
@@ -242,15 +249,18 @@ class StableDiffusionXLImg2ImgPipeline(DiffusionPipeline, FromSingleFileMixin):
                 return torch.device(module._hf_hook.execution_device)
         return self.device
 
+    # Copied from diffusers.pipelines.stable_diffusion_xl.pipeline_stable_diffusion_xl.StableDiffusionXLPipeline.encode_prompt
     def encode_prompt(
         self,
         prompt,
-        device,
-        num_images_per_prompt,
-        do_classifier_free_guidance,
+        device: Optional[torch.device] = None,
+        num_images_per_prompt: int = 1,
+        do_classifier_free_guidance: bool = True,
         negative_prompt=None,
         prompt_embeds: Optional[torch.FloatTensor] = None,
         negative_prompt_embeds: Optional[torch.FloatTensor] = None,
+        pooled_prompt_embeds: Optional[torch.FloatTensor] = None,
+        negative_pooled_prompt_embeds: Optional[torch.FloatTensor] = None,
         lora_scale: Optional[float] = None,
     ):
         r"""
@@ -276,9 +286,18 @@ class StableDiffusionXLImg2ImgPipeline(DiffusionPipeline, FromSingleFileMixin):
                 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.FloatTensor`, *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.FloatTensor`, *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.
         """
+        device = device or self._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(self, LoraLoaderMixin):
@@ -327,13 +346,11 @@ class StableDiffusionXLImg2ImgPipeline(DiffusionPipeline, FromSingleFileMixin):
                     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]
-
                 prompt_embeds = prompt_embeds.hidden_states[-2]
 
-                prompt_embeds = prompt_embeds
-
                 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)
@@ -349,10 +366,9 @@ class StableDiffusionXLImg2ImgPipeline(DiffusionPipeline, FromSingleFileMixin):
             negative_prompt_embeds = torch.zeros_like(prompt_embeds)
             negative_pooled_prompt_embeds = torch.zeros_like(pooled_prompt_embeds)
         elif do_classifier_free_guidance and negative_prompt_embeds is None:
+            negative_prompt = negative_prompt or ""
             uncond_tokens: List[str]
-            if negative_prompt is None:
-                uncond_tokens = [""] * batch_size
-            elif prompt is not None and type(prompt) is not type(negative_prompt):
+            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)}."
@@ -389,7 +405,6 @@ class StableDiffusionXLImg2ImgPipeline(DiffusionPipeline, FromSingleFileMixin):
                 )
                 # 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]
 
                 if do_classifier_free_guidance:
@@ -411,6 +426,7 @@ class StableDiffusionXLImg2ImgPipeline(DiffusionPipeline, FromSingleFileMixin):
 
             negative_prompt_embeds = torch.concat(negative_prompt_embeds_list, dim=-1)
 
+        bs_embed = pooled_prompt_embeds.shape[0]
         pooled_prompt_embeds = pooled_prompt_embeds.repeat(1, num_images_per_prompt).view(
             bs_embed * num_images_per_prompt, -1
         )
@@ -420,20 +436,7 @@ class StableDiffusionXLImg2ImgPipeline(DiffusionPipeline, FromSingleFileMixin):
 
         return prompt_embeds, negative_prompt_embeds, pooled_prompt_embeds, negative_pooled_prompt_embeds
 
-    def run_safety_checker(self, image, device, dtype):
-        if self.safety_checker is None:
-            has_nsfw_concept = None
-        else:
-            if torch.is_tensor(image):
-                feature_extractor_input = self.image_processor.postprocess(image, output_type="pil")
-            else:
-                feature_extractor_input = self.image_processor.numpy_to_pil(image)
-            safety_checker_input = self.feature_extractor(feature_extractor_input, return_tensors="pt").to(device)
-            image, has_nsfw_concept = self.safety_checker(
-                images=image, clip_input=safety_checker_input.pixel_values.to(dtype)
-            )
-        return image, has_nsfw_concept
-
+    # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.prepare_extra_step_kwargs
     def prepare_extra_step_kwargs(self, generator, eta):
         # prepare extra kwargs for the scheduler step, since not all schedulers have the same signature
         # eta (η) is only used with the DDIMScheduler, it will be ignored for other schedulers.
@@ -624,6 +627,8 @@ class StableDiffusionXLImg2ImgPipeline(DiffusionPipeline, FromSingleFileMixin):
         latents: Optional[torch.FloatTensor] = None,
         prompt_embeds: Optional[torch.FloatTensor] = None,
         negative_prompt_embeds: Optional[torch.FloatTensor] = None,
+        pooled_prompt_embeds: Optional[torch.FloatTensor] = None,
+        negative_pooled_prompt_embeds: Optional[torch.FloatTensor] = None,
         output_type: Optional[str] = "pil",
         return_dict: bool = True,
         callback: Optional[Callable[[int, int, torch.FloatTensor], None]] = None,
@@ -683,6 +688,13 @@ class StableDiffusionXLImg2ImgPipeline(DiffusionPipeline, FromSingleFileMixin):
                 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.FloatTensor`, *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.FloatTensor`, *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.
             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`.
@@ -759,6 +771,8 @@ class StableDiffusionXLImg2ImgPipeline(DiffusionPipeline, FromSingleFileMixin):
             negative_prompt,
             prompt_embeds=prompt_embeds,
             negative_prompt_embeds=negative_prompt_embeds,
+            pooled_prompt_embeds=pooled_prompt_embeds,
+            negative_pooled_prompt_embeds=negative_pooled_prompt_embeds,
             lora_scale=text_encoder_lora_scale,
         )
 
@@ -845,15 +859,18 @@ class StableDiffusionXLImg2ImgPipeline(DiffusionPipeline, FromSingleFileMixin):
         # make sure the VAE is in float32 mode, as it overflows in float16
         self.vae.to(dtype=torch.float32)
 
-        use_torch_2_0_or_xformers = self.vae.decoder.mid_block.attentions[0].processor in [
-            AttnProcessor2_0,
-            XFormersAttnProcessor,
-            LoRAXFormersAttnProcessor,
-            LoRAAttnProcessor2_0,
-        ]
+        use_torch_2_0_or_xformers = isinstance(
+            self.vae.decoder.mid_block.attentions[0].processor,
+            (
+                AttnProcessor2_0,
+                XFormersAttnProcessor,
+                LoRAXFormersAttnProcessor,
+                LoRAAttnProcessor2_0,
+            ),
+        )
         # if xformers or torch_2_0 is used attention block does not need
         # to be in float32 which can save lots of memory
-        if not use_torch_2_0_or_xformers:
+        if use_torch_2_0_or_xformers:
             self.vae.post_quant_conv.to(latents.dtype)
             self.vae.decoder.conv_in.to(latents.dtype)
             self.vae.decoder.mid_block.to(latents.dtype)
@@ -862,24 +879,18 @@ class StableDiffusionXLImg2ImgPipeline(DiffusionPipeline, FromSingleFileMixin):
 
         if not output_type == "latent":
             image = self.vae.decode(latents / self.vae.config.scaling_factor, return_dict=False)[0]
-            has_nsfw_concept = None
         else:
             image = latents
-            return StableDiffusionXLPipelineOutput(images=image, nsfw_content_detected=None)
-
-        if has_nsfw_concept is None:
-            do_denormalize = [True] * image.shape[0]
-        else:
-            do_denormalize = [not has_nsfw for has_nsfw in has_nsfw_concept]
+            return StableDiffusionXLPipelineOutput(images=image)
 
         image = self.watermark.apply_watermark(image)
-        image = self.image_processor.postprocess(image, output_type=output_type, do_denormalize=do_denormalize)
+        image = self.image_processor.postprocess(image, output_type=output_type)
 
         # Offload last model to CPU
         if hasattr(self, "final_offload_hook") and self.final_offload_hook is not None:
             self.final_offload_hook.offload()
 
         if not return_dict:
-            return (image, has_nsfw_concept)
+            return (image,)
 
-        return StableDiffusionXLPipelineOutput(images=image, nsfw_content_detected=has_nsfw_concept)
+        return StableDiffusionXLPipelineOutput(images=image)

tests/others/test_config.py

@@ -264,7 +264,7 @@ class ConfigTester(unittest.TestCase):
         config_dict = {k: v for k, v in config.config.items() if not k.startswith("_")}
 
         # make sure that default config has all keys in `_use_default_values`
-        assert set(config_dict.keys()) == config.config._use_default_values
+        assert set(config_dict.keys()) == set(config.config._use_default_values)
 
         with tempfile.TemporaryDirectory() as tmpdirname:
             config.save_config(tmpdirname)

tests/pipelines/stable_diffusion/test_stable_diffusion_inpaint.py

@@ -20,17 +20,20 @@ import unittest
 
 import numpy as np
 import torch
+from huggingface_hub import hf_hub_download
 from PIL import Image
 from transformers import CLIPTextConfig, CLIPTextModel, CLIPTokenizer
 
 from diffusers import (
     AutoencoderKL,
+    DDIMScheduler,
     DPMSolverMultistepScheduler,
     LMSDiscreteScheduler,
     PNDMScheduler,
     StableDiffusionInpaintPipeline,
     UNet2DConditionModel,
 )
+from diffusers.models.attention_processor import AttnProcessor
 from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion_inpaint import prepare_mask_and_masked_image
 from diffusers.utils import floats_tensor, load_image, load_numpy, nightly, slow, torch_device
 from diffusers.utils.testing_utils import (
@@ -512,6 +515,42 @@ class StableDiffusionInpaintPipelineSlowTests(unittest.TestCase):
 
         assert np.abs(expected_slice - image_slice).max() < 6e-4
 
+    def test_download_local(self):
+        filename = hf_hub_download("runwayml/stable-diffusion-inpainting", filename="sd-v1-5-inpainting.ckpt")
+
+        pipe = StableDiffusionInpaintPipeline.from_single_file(filename, torch_dtype=torch.float16)
+        pipe.scheduler = DDIMScheduler.from_config(pipe.scheduler.config)
+        pipe.to("cuda")
+
+        inputs = self.get_inputs(torch_device)
+        inputs["num_inference_steps"] = 1
+        image_out = pipe(**inputs).images[0]
+
+        assert image_out.shape == (512, 512, 3)
+
+    def test_download_ckpt_diff_format_is_same(self):
+        ckpt_path = "https://huggingface.co/runwayml/stable-diffusion-inpainting/blob/main/sd-v1-5-inpainting.ckpt"
+
+        pipe = StableDiffusionInpaintPipeline.from_single_file(ckpt_path)
+        pipe.scheduler = DDIMScheduler.from_config(pipe.scheduler.config)
+        pipe.unet.set_attn_processor(AttnProcessor())
+        pipe.to("cuda")
+
+        inputs = self.get_inputs(torch_device)
+        inputs["num_inference_steps"] = 5
+        image_ckpt = pipe(**inputs).images[0]
+
+        pipe = StableDiffusionInpaintPipeline.from_pretrained("runwayml/stable-diffusion-inpainting")
+        pipe.scheduler = DDIMScheduler.from_config(pipe.scheduler.config)
+        pipe.unet.set_attn_processor(AttnProcessor())
+        pipe.to("cuda")
+
+        inputs = self.get_inputs(torch_device)
+        inputs["num_inference_steps"] = 5
+        image = pipe(**inputs).images[0]
+
+        assert np.max(np.abs(image - image_ckpt)) < 1e-4
+
 
 @nightly
 @require_torch_gpu

tests/pipelines/stable_diffusion_2/test_stable_diffusion_v_pred.py

@@ -19,6 +19,7 @@ import unittest
 
 import numpy as np
 import torch
+from huggingface_hub import hf_hub_download
 from transformers import CLIPTextConfig, CLIPTextModel, CLIPTokenizer
 
 from diffusers import (
@@ -29,6 +30,7 @@ from diffusers import (
     StableDiffusionPipeline,
     UNet2DConditionModel,
 )
+from diffusers.models.attention_processor import AttnProcessor
 from diffusers.utils import load_numpy, slow, torch_device
 from diffusers.utils.testing_utils import enable_full_determinism, require_torch_gpu
 
@@ -426,6 +428,40 @@ class StableDiffusion2VPredictionPipelineIntegrationTests(unittest.TestCase):
         assert image.shape == (768, 768, 3)
         assert np.abs(expected_image - image).max() < 7.5e-1
 
+    def test_download_local(self):
+        filename = hf_hub_download("stabilityai/stable-diffusion-2-1", filename="v2-1_768-ema-pruned.safetensors")
+
+        pipe = StableDiffusionPipeline.from_single_file(filename, torch_dtype=torch.float16)
+        pipe.scheduler = DDIMScheduler.from_config(pipe.scheduler.config)
+        pipe.to("cuda")
+
+        image_out = pipe("test", num_inference_steps=1, output_type="np").images[0]
+
+        assert image_out.shape == (768, 768, 3)
+
+    def test_download_ckpt_diff_format_is_same(self):
+        single_file_path = (
+            "https://huggingface.co/stabilityai/stable-diffusion-2-1/blob/main/v2-1_768-ema-pruned.safetensors"
+        )
+
+        pipe_single = StableDiffusionPipeline.from_single_file(single_file_path)
+        pipe_single.scheduler = DDIMScheduler.from_config(pipe_single.scheduler.config)
+        pipe_single.unet.set_attn_processor(AttnProcessor())
+        pipe_single.to("cuda")
+
+        generator = torch.Generator(device="cpu").manual_seed(0)
+        image_ckpt = pipe_single("a turtle", num_inference_steps=5, generator=generator, output_type="np").images[0]
+
+        pipe = StableDiffusionPipeline.from_pretrained("stabilityai/stable-diffusion-2-1")
+        pipe.scheduler = DDIMScheduler.from_config(pipe.scheduler.config)
+        pipe.unet.set_attn_processor(AttnProcessor())
+        pipe.to("cuda")
+
+        generator = torch.Generator(device="cpu").manual_seed(0)
+        image = pipe("a turtle", num_inference_steps=5, generator=generator, output_type="np").images[0]
+
+        assert np.max(np.abs(image - image_ckpt)) < 1e-3
+
     def test_stable_diffusion_text2img_intermediate_state_v_pred(self):
         number_of_steps = 0
 

tests/pipelines/stable_diffusion_xl/test_stable_diffusion_xl.py

@@ -144,6 +144,46 @@ class StableDiffusionXLPipelineFastTests(PipelineLatentTesterMixin, PipelineTest
 
         assert np.abs(image_slice.flatten() - expected_slice).max() < 1e-2
 
+    def test_stable_diffusion_xl_negative_prompt_embeds(self):
+        components = self.get_dummy_components()
+        sd_pipe = StableDiffusionXLPipeline(**components)
+        sd_pipe = sd_pipe.to(torch_device)
+        sd_pipe = sd_pipe.to(torch_device)
+        sd_pipe.set_progress_bar_config(disable=None)
+
+        # forward without prompt embeds
+        inputs = self.get_dummy_inputs(torch_device)
+        negative_prompt = 3 * ["this is a negative prompt"]
+        inputs["negative_prompt"] = negative_prompt
+        inputs["prompt"] = 3 * [inputs["prompt"]]
+
+        output = sd_pipe(**inputs)
+        image_slice_1 = output.images[0, -3:, -3:, -1]
+
+        # forward with prompt embeds
+        inputs = self.get_dummy_inputs(torch_device)
+        negative_prompt = 3 * ["this is a negative prompt"]
+        prompt = 3 * [inputs.pop("prompt")]
+
+        (
+            prompt_embeds,
+            negative_prompt_embeds,
+            pooled_prompt_embeds,
+            negative_pooled_prompt_embeds,
+        ) = sd_pipe.encode_prompt(prompt, negative_prompt=negative_prompt)
+
+        output = sd_pipe(
+            **inputs,
+            prompt_embeds=prompt_embeds,
+            negative_prompt_embeds=negative_prompt_embeds,
+            pooled_prompt_embeds=pooled_prompt_embeds,
+            negative_pooled_prompt_embeds=negative_pooled_prompt_embeds,
+        )
+        image_slice_2 = output.images[0, -3:, -3:, -1]
+
+        # make sure that it's equal
+        assert np.abs(image_slice_1.flatten() - image_slice_2.flatten()).max() < 1e-4
+
     def test_attention_slicing_forward_pass(self):
         super().test_attention_slicing_forward_pass(expected_max_diff=3e-3)
 

tests/pipelines/stable_diffusion_xl/test_stable_diffusion_xl_img2img.py

@@ -165,6 +165,46 @@ class StableDiffusionXLImg2ImgPipelineFastTests(PipelineLatentTesterMixin, Pipel
     def test_save_load_optional_components(self):
         pass
 
+    def test_stable_diffusion_xl_img2img_negative_prompt_embeds(self):
+        components = self.get_dummy_components()
+        sd_pipe = StableDiffusionXLImg2ImgPipeline(**components)
+        sd_pipe = sd_pipe.to(torch_device)
+        sd_pipe = sd_pipe.to(torch_device)
+        sd_pipe.set_progress_bar_config(disable=None)
+
+        # forward without prompt embeds
+        inputs = self.get_dummy_inputs(torch_device)
+        negative_prompt = 3 * ["this is a negative prompt"]
+        inputs["negative_prompt"] = negative_prompt
+        inputs["prompt"] = 3 * [inputs["prompt"]]
+
+        output = sd_pipe(**inputs)
+        image_slice_1 = output.images[0, -3:, -3:, -1]
+
+        # forward with prompt embeds
+        inputs = self.get_dummy_inputs(torch_device)
+        negative_prompt = 3 * ["this is a negative prompt"]
+        prompt = 3 * [inputs.pop("prompt")]
+
+        (
+            prompt_embeds,
+            negative_prompt_embeds,
+            pooled_prompt_embeds,
+            negative_pooled_prompt_embeds,
+        ) = sd_pipe.encode_prompt(prompt, negative_prompt=negative_prompt)
+
+        output = sd_pipe(
+            **inputs,
+            prompt_embeds=prompt_embeds,
+            negative_prompt_embeds=negative_prompt_embeds,
+            pooled_prompt_embeds=pooled_prompt_embeds,
+            negative_pooled_prompt_embeds=negative_pooled_prompt_embeds,
+        )
+        image_slice_2 = output.images[0, -3:, -3:, -1]
+
+        # make sure that it's equal
+        assert np.abs(image_slice_1.flatten() - image_slice_2.flatten()).max() < 1e-4
+
 
 @slow
 @require_torch_gpu

tests/pipelines/test_pipelines.py

@@ -14,6 +14,7 @@
 # limitations under the License.
 
 import gc
+import glob
 import json
 import os
 import random
@@ -56,6 +57,7 @@ from diffusers import (
     UniPCMultistepScheduler,
     logging,
 )
+from diffusers.pipelines.pipeline_utils import variant_compatible_siblings
 from diffusers.schedulers.scheduling_utils import SCHEDULER_CONFIG_NAME
 from diffusers.utils import (
     CONFIG_NAME,
@@ -1361,6 +1363,29 @@ class PipelineFastTests(unittest.TestCase):
             assert sd.config.safety_checker != (None, None)
             assert sd.config.feature_extractor != (None, None)
 
+    def test_warning_no_variant_available(self):
+        variant = "fp16"
+        with self.assertWarns(FutureWarning) as warning_context:
+            cached_folder = StableDiffusionPipeline.download(
+                "hf-internal-testing/diffusers-stable-diffusion-tiny-all", variant=variant
+            )
+
+        assert "but no such modeling files are available" in str(warning_context.warning)
+        assert variant in str(warning_context.warning)
+
+        def get_all_filenames(directory):
+            filenames = glob.glob(directory + "/**", recursive=True)
+            filenames = [f for f in filenames if os.path.isfile(f)]
+            return filenames
+
+        filenames = get_all_filenames(str(cached_folder))
+
+        all_model_files, variant_model_files = variant_compatible_siblings(filenames, variant=variant)
+
+        # make sure that none of the model names are variant model names
+        assert len(variant_model_files) == 0
+        assert len(all_model_files) > 0
+
 
 @slow
 @require_torch_gpu

tests/pipelines/test_pipelines_common.py

@@ -699,12 +699,16 @@ class PipelineTesterMixin:
 
         inputs = self.get_dummy_inputs(torch_device)
         output_without_offload = pipe(**inputs)[0]
-        output_without_offload.cpu() if torch.is_tensor(output_without_offload) else output_without_offload
+        output_without_offload = (
+            output_without_offload.cpu() if torch.is_tensor(output_without_offload) else output_without_offload
+        )
 
         pipe.enable_xformers_memory_efficient_attention()
         inputs = self.get_dummy_inputs(torch_device)
         output_with_offload = pipe(**inputs)[0]
-        output_with_offload.cpu() if torch.is_tensor(output_with_offload) else output_without_offload
+        output_with_offload = (
+            output_with_offload.cpu() if torch.is_tensor(output_with_offload) else output_without_offload
+        )
 
         if test_max_difference:
             max_diff = np.abs(output_with_offload - output_without_offload).max()

tests/pipelines/text_to_video/test_text_to_video.py

@@ -26,7 +26,7 @@ from diffusers import (
     TextToVideoSDPipeline,
     UNet3DConditionModel,
 )
-from diffusers.utils import load_numpy, skip_mps, slow
+from diffusers.utils import is_xformers_available, load_numpy, skip_mps, slow, torch_device
 from diffusers.utils.testing_utils import enable_full_determinism
 
 from ..pipeline_params import TEXT_TO_IMAGE_BATCH_PARAMS, TEXT_TO_IMAGE_PARAMS
@@ -143,6 +143,13 @@ class TextToVideoSDPipelineFastTests(PipelineTesterMixin, unittest.TestCase):
     def test_attention_slicing_forward_pass(self):
         self._test_attention_slicing_forward_pass(test_mean_pixel_difference=False, expected_max_diff=3e-3)
 
+    @unittest.skipIf(
+        torch_device != "cuda" or not is_xformers_available(),
+        reason="XFormers attention is only available with CUDA and `xformers` installed",
+    )
+    def test_xformers_attention_forwardGenerator_pass(self):
+        self._test_xformers_attention_forwardGenerator_pass(test_mean_pixel_difference=False, expected_max_diff=1e-2)
+
     # (todo): sayakpaul
     @unittest.skip(reason="Batching needs to be properly figured out first for this pipeline.")
     def test_inference_batch_consistent(self):