[SD-XL] Add new pipelines (#3859)

* Add new text encoder

* add transformers depth

* More

* Correct conversion script

* Fix more

* Fix more

* Correct more

* correct text encoder

* Finish all

* proof that in works in run local xl

* clean up

* Get refiner to work

* Add red castle

* Fix batch size

* Improve pipelines more

* Finish text2image tests

* Add img2img test

* Fix more

* fix import

* Fix embeddings for classic models (#3888)

Fix embeddings for classic SD models.

* Allow multiple prompts to be passed to the refiner (#3895)

* finish more

* Apply suggestions from code review

* add watermarker

* Model offload (#3889)

* Model offload.

* Model offload for refiner / img2img

* Hardcode encoder offload on img2img vae encode

Saves some GPU RAM in img2img / refiner tasks so it remains below 8 GB.

---------
Co-authored-by: Patrick von Platen <patrick.v.platen@gmail.com>

* correct

* fix

* clean print

* Update install warning for `invisible-watermark`

* add: missing docstrings.

* fix and simplify the usage example in img2img.

* fix setup for watermarking.

* Revert "fix setup for watermarking."

This reverts commit 491bc9f5a640bbf46a97a8e52d6eff7e70eb8e4b.

* fix: watermarking setup.

* fix: op.

* run make fix-copies.

* make sure tests pass

* improve convert

* make tests pass

* make tests pass

* better error message

* fiinsh

* finish

* Fix final test

---------
Co-authored-by: Pedro Cuenca <pedro@huggingface.co>
Co-authored-by: Sayak Paul <spsayakpaul@gmail.com>

.github/workflows/build_documentation.yml

@@ -10,12 +10,19 @@ on:
 
 jobs:
   build:
+    steps:
+      - name: Install dependencies
+        run: |
+          apt-get update && apt-get install libsndfile1-dev libgl1 -y
+
+      - name: Build doc
         uses: huggingface/doc-builder/.github/workflows/build_main_documentation.yml@main
         with:
           commit_sha: ${{ github.sha }}
           package: diffusers
           notebook_folder: diffusers_doc
           languages: en ko zh
+
     secrets:
       token: ${{ secrets.HUGGINGFACE_PUSH }}
       hf_token: ${{ secrets.HF_DOC_BUILD_PUSH }}

.github/workflows/build_pr_documentation.yml

@@ -9,9 +9,15 @@ concurrency:
 
 jobs:
   build:
+    steps:
+      - name: Install dependencies
+        run: |
+          apt-get update && apt-get install libsndfile1-dev libgl1 -y
+
+      - name: Build doc
         uses: huggingface/doc-builder/.github/workflows/build_pr_documentation.yml@main
         with:
           commit_sha: ${{ github.event.pull_request.head.sha }}
           pr_number: ${{ github.event.number }}
           package: diffusers
-      languages: en ko
+          languages: en ko zh

.github/workflows/pr_tests.yml

@@ -62,7 +62,7 @@ jobs:
 
     - name: Install dependencies
       run: |
-        apt-get update && apt-get install libsndfile1-dev -y
+        apt-get update && apt-get install libsndfile1-dev libgl1 -y
         python -m pip install -e .[quality,test]
 
     - name: Environment

docker/diffusers-pytorch-cpu/Dockerfile

@@ -14,6 +14,7 @@ RUN apt update && \
                    libsndfile1-dev \
                    python3.8 \
                    python3-pip \
+                   libgl1 \
                    python3.8-venv && \
     rm -rf /var/lib/apt/lists
 
@@ -27,6 +28,7 @@ RUN python3 -m pip install --no-cache-dir --upgrade pip && \
         torch \
         torchvision \
         torchaudio \
+        invisible_watermark \
         --extra-index-url https://download.pytorch.org/whl/cpu && \
     python3 -m pip install --no-cache-dir \
         accelerate \

docker/diffusers-pytorch-cuda/Dockerfile

@@ -12,6 +12,7 @@ RUN apt update && \
                    curl \
                    ca-certificates \
                    libsndfile1-dev \
+                   libgl1 \
                    python3.8 \
                    python3-pip \
                    python3.8-venv && \
@@ -26,7 +27,8 @@ RUN python3 -m pip install --no-cache-dir --upgrade pip && \
     python3 -m pip install --no-cache-dir \
         torch \
         torchvision \
-        torchaudio && \
+        torchaudio \
+        invisible_watermark && \
     python3 -m pip install --no-cache-dir \
         accelerate \
         datasets \

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

+<!--Copyright 2023 The HuggingFace Team. All rights reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with
+the License. You may obtain a copy of the License at
+
+http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on
+an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the
+specific language governing permissions and limitations under the License.
+-->
+
+# Stable diffusion XL
+
+Stable Diffusion 2 is a text-to-image _latent diffusion_ model built upon the work of [Stable Diffusion 1](https://stability.ai/blog/stable-diffusion-public-release). 
+The project to train Stable Diffusion 2 was led by Robin Rombach and Katherine Crowson from [Stability AI](https://stability.ai/) and [LAION](https://laion.ai/).
+
+*The Stable Diffusion 2.0 release includes robust text-to-image models trained using a brand new text encoder (OpenCLIP), developed by LAION with support from Stability AI, which greatly improves the quality of the generated images compared to earlier V1 releases. The text-to-image models in this release can generate images with default resolutions of both 512x512 pixels and 768x768 pixels. 
+These models are trained on an aesthetic subset of the [LAION-5B dataset](https://laion.ai/blog/laion-5b/) created by the DeepFloyd team at Stability AI, which is then further filtered to remove adult content using [LAION’s NSFW filter](https://openreview.net/forum?id=M3Y74vmsMcY).*
+
+For more details about how Stable Diffusion 2 works and how it differs from Stable Diffusion 1, please refer to the official [launch announcement post](https://stability.ai/blog/stable-diffusion-v2-release).
+
+## Tips
+
+### Available checkpoints:
+
+- *Text-to-Image (1024x1024 resolution)*: [stabilityai/stable-diffusion-xl-base-0.9](https://huggingface.co/stabilityai/stable-diffusion-xl-base-0.9) with [`StableDiffusionXLPipeline`]
+- *Image-to-Image / Refiner (1024x1024 resolution)*: [stabilityai/stable-diffusion-xl-refiner-0.9](https://huggingface.co/stabilityai/stable-diffusion-xl-refiner-0.9) with [`StableDiffusionXLImg2ImgPipeline`]
+
+TODO
+
+## StableDiffusionXLPipeline
+
+[[autodoc]] StableDiffusionXLPipeline
+	- all
+	- __call__
+
+## StableDiffusionXLImg2ImgPipeline
+
+[[autodoc]] StableDiffusionXLImg2ImgPipeline
+	- all
+	- __call__

scripts/convert_original_stable_diffusion_to_diffusers.py

@@ -126,6 +126,13 @@ if __name__ == "__main__":
         "--controlnet", action="store_true", default=None, help="Set flag if this is a controlnet checkpoint."
     )
     parser.add_argument("--half", action="store_true", help="Save weights in half precision.")
+    parser.add_argument(
+        "--vae_path",
+        type=str,
+        default=None,
+        required=False,
+        help="Set to a path, hub id to an already converted vae to not convert it again.",
+    )
     args = parser.parse_args()
 
     pipe = download_from_original_stable_diffusion_ckpt(
@@ -144,6 +151,7 @@ if __name__ == "__main__":
         stable_unclip_prior=args.stable_unclip_prior,
         clip_stats_path=args.clip_stats_path,
         controlnet=args.controlnet,
+        vae_path=args.vae_path,
     )
 
     if args.half:

setup.py

@@ -89,6 +89,7 @@ _deps = [
     "huggingface-hub>=0.13.2",
     "requests-mock==1.10.0",
     "importlib_metadata",
+    "invisible-watermark",
     "isort>=5.5.4",
     "jax>=0.2.8,!=0.3.2",
     "jaxlib>=0.1.65",
@@ -193,6 +194,7 @@ extras["test"] = deps_list(
     "compel",
     "datasets",
     "Jinja2",
+    "invisible-watermark",
     "k-diffusion",
     "librosa",
     "omegaconf",

src/diffusers/__init__.py

@@ -5,6 +5,7 @@ from .utils import (
     OptionalDependencyNotAvailable,
     is_flax_available,
     is_inflect_available,
+    is_invisible_watermark_available,
     is_k_diffusion_available,
     is_k_diffusion_version,
     is_librosa_available,
@@ -179,6 +180,14 @@ else:
         VQDiffusionPipeline,
     )
 
+try:
+    if not (is_torch_available() and is_transformers_available() and is_invisible_watermark_available()):
+        raise OptionalDependencyNotAvailable()
+except OptionalDependencyNotAvailable:
+    from .utils.dummy_torch_and_transformers_and_invisible_watermark_objects import *  # noqa F403
+else:
+    from .pipelines import StableDiffusionXLImg2ImgPipeline, StableDiffusionXLPipeline
+
 try:
     if not (is_torch_available() and is_transformers_available() and is_k_diffusion_available()):
         raise OptionalDependencyNotAvailable()

src/diffusers/dependency_versions_table.py

@@ -13,6 +13,7 @@ deps = {
     "huggingface-hub": "huggingface-hub>=0.13.2",
     "requests-mock": "requests-mock==1.10.0",
     "importlib_metadata": "importlib_metadata",
+    "invisible-watermark": "invisible-watermark",
     "isort": "isort>=5.5.4",
     "jax": "jax>=0.2.8,!=0.3.2",
     "jaxlib": "jaxlib>=0.1.65",

src/diffusers/models/attention_processor.py

@@ -1118,7 +1118,9 @@ class AttnProcessor2_0:
         value = attn.to_v(encoder_hidden_states)
 
         head_dim = inner_dim // attn.heads
+
         query = query.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)
+
         key = key.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)
         value = value.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)
 

src/diffusers/models/unet_2d_blocks.py

@@ -38,6 +38,7 @@ def get_down_block(
     add_downsample,
     resnet_eps,
     resnet_act_fn,
+    transformer_layers_per_block=1,
     num_attention_heads=None,
     resnet_groups=None,
     cross_attention_dim=None,
@@ -111,6 +112,7 @@ def get_down_block(
             raise ValueError("cross_attention_dim must be specified for CrossAttnDownBlock2D")
         return CrossAttnDownBlock2D(
             num_layers=num_layers,
+            transformer_layers_per_block=transformer_layers_per_block,
             in_channels=in_channels,
             out_channels=out_channels,
             temb_channels=temb_channels,
@@ -232,6 +234,7 @@ def get_up_block(
     add_upsample,
     resnet_eps,
     resnet_act_fn,
+    transformer_layers_per_block=1,
     num_attention_heads=None,
     resnet_groups=None,
     cross_attention_dim=None,
@@ -287,6 +290,7 @@ def get_up_block(
             raise ValueError("cross_attention_dim must be specified for CrossAttnUpBlock2D")
         return CrossAttnUpBlock2D(
             num_layers=num_layers,
+            transformer_layers_per_block=transformer_layers_per_block,
             in_channels=in_channels,
             out_channels=out_channels,
             prev_output_channel=prev_output_channel,
@@ -517,6 +521,7 @@ class UNetMidBlock2DCrossAttn(nn.Module):
         temb_channels: int,
         dropout: float = 0.0,
         num_layers: int = 1,
+        transformer_layers_per_block: int = 1,
         resnet_eps: float = 1e-6,
         resnet_time_scale_shift: str = "default",
         resnet_act_fn: str = "swish",
@@ -559,7 +564,7 @@ class UNetMidBlock2DCrossAttn(nn.Module):
                         num_attention_heads,
                         in_channels // num_attention_heads,
                         in_channels=in_channels,
-                        num_layers=1,
+                        num_layers=transformer_layers_per_block,
                         cross_attention_dim=cross_attention_dim,
                         norm_num_groups=resnet_groups,
                         use_linear_projection=use_linear_projection,
@@ -862,6 +867,7 @@ class CrossAttnDownBlock2D(nn.Module):
         temb_channels: int,
         dropout: float = 0.0,
         num_layers: int = 1,
+        transformer_layers_per_block: int = 1,
         resnet_eps: float = 1e-6,
         resnet_time_scale_shift: str = "default",
         resnet_act_fn: str = "swish",
@@ -906,7 +912,7 @@ class CrossAttnDownBlock2D(nn.Module):
                         num_attention_heads,
                         out_channels // num_attention_heads,
                         in_channels=out_channels,
-                        num_layers=1,
+                        num_layers=transformer_layers_per_block,
                         cross_attention_dim=cross_attention_dim,
                         norm_num_groups=resnet_groups,
                         use_linear_projection=use_linear_projection,
@@ -1995,6 +2001,7 @@ class CrossAttnUpBlock2D(nn.Module):
         temb_channels: int,
         dropout: float = 0.0,
         num_layers: int = 1,
+        transformer_layers_per_block: int = 1,
         resnet_eps: float = 1e-6,
         resnet_time_scale_shift: str = "default",
         resnet_act_fn: str = "swish",
@@ -2040,7 +2047,7 @@ class CrossAttnUpBlock2D(nn.Module):
                         num_attention_heads,
                         out_channels // num_attention_heads,
                         in_channels=out_channels,
-                        num_layers=1,
+                        num_layers=transformer_layers_per_block,
                         cross_attention_dim=cross_attention_dim,
                         norm_num_groups=resnet_groups,
                         use_linear_projection=use_linear_projection,

src/diffusers/models/unet_2d_condition.py

@@ -98,7 +98,11 @@ class UNet2DConditionModel(ModelMixin, ConfigMixin, UNet2DConditionLoadersMixin)
         norm_eps (`float`, *optional*, defaults to 1e-5): The epsilon to use for the normalization.
         cross_attention_dim (`int` or `Tuple[int]`, *optional*, defaults to 1280):
             The dimension of the cross attention features.
-        encoder_hid_dim (`int`, *optional*, defaults to `None`):
+        transformer_layers_per_block (`int` or `Tuple[int]`, *optional*, defaults to 1):
+            The number of transformer blocks of type [`~models.attention.BasicTransformerBlock`]. Only relevant for
+            [`~models.unet_2d_blocks.CrossAttnDownBlock2D`], [`~models.unet_2d_blocks.CrossAttnUpBlock2D`],
+            [`~models.unet_2d_blocks.UNetMidBlock2DCrossAttn`].
+        encoder_hid_dim (`int`, *optional*, defaults to None):
             If `encoder_hid_dim_type` is defined, `encoder_hidden_states` will be projected from `encoder_hid_dim`
             dimension to `cross_attention_dim`.
         encoder_hid_dim_type (`str`, *optional*, defaults to `None`):
@@ -115,6 +119,8 @@ class UNet2DConditionModel(ModelMixin, ConfigMixin, UNet2DConditionLoadersMixin)
         addition_embed_type (`str`, *optional*, defaults to `None`):
             Configures an optional embedding which will be summed with the time embeddings. Choose from `None` or
             "text". "text" will use the `TextTimeEmbedding` layer.
+        addition_time_embed_dim: (`int`, *optional*, defaults to `None`):
+            Dimension for the timestep embeddings.
         num_class_embeds (`int`, *optional*, defaults to `None`):
             Input dimension of the learnable embedding matrix to be projected to `time_embed_dim`, when performing
             class conditioning with `class_embed_type` equal to `None`.
@@ -170,6 +176,7 @@ class UNet2DConditionModel(ModelMixin, ConfigMixin, UNet2DConditionLoadersMixin)
         norm_num_groups: Optional[int] = 32,
         norm_eps: float = 1e-5,
         cross_attention_dim: Union[int, Tuple[int]] = 1280,
+        transformer_layers_per_block: Union[int, Tuple[int]] = 1,
         encoder_hid_dim: Optional[int] = None,
         encoder_hid_dim_type: Optional[str] = None,
         attention_head_dim: Union[int, Tuple[int]] = 8,
@@ -178,6 +185,7 @@ class UNet2DConditionModel(ModelMixin, ConfigMixin, UNet2DConditionLoadersMixin)
         use_linear_projection: bool = False,
         class_embed_type: Optional[str] = None,
         addition_embed_type: Optional[str] = None,
+        addition_time_embed_dim: Optional[int] = None,
         num_class_embeds: Optional[int] = None,
         upcast_attention: bool = False,
         resnet_time_scale_shift: str = "default",
@@ -351,6 +359,10 @@ class UNet2DConditionModel(ModelMixin, ConfigMixin, UNet2DConditionLoadersMixin)
             self.add_embedding = TextImageTimeEmbedding(
                 text_embed_dim=cross_attention_dim, image_embed_dim=cross_attention_dim, time_embed_dim=time_embed_dim
             )
+        elif addition_embed_type == "text_time":
+            self.add_time_proj = Timesteps(addition_time_embed_dim, flip_sin_to_cos, freq_shift)
+            self.add_embedding = TimestepEmbedding(projection_class_embeddings_input_dim, time_embed_dim)
+
         elif addition_embed_type is not None:
             raise ValueError(f"addition_embed_type: {addition_embed_type} must be None, 'text' or 'text_image'.")
 
@@ -383,6 +395,9 @@ class UNet2DConditionModel(ModelMixin, ConfigMixin, UNet2DConditionLoadersMixin)
         if isinstance(layers_per_block, int):
             layers_per_block = [layers_per_block] * len(down_block_types)
 
+        if isinstance(transformer_layers_per_block, int):
+            transformer_layers_per_block = [transformer_layers_per_block] * len(down_block_types)
+
         if class_embeddings_concat:
             # The time embeddings are concatenated with the class embeddings. The dimension of the
             # time embeddings passed to the down, middle, and up blocks is twice the dimension of the
@@ -401,6 +416,7 @@ class UNet2DConditionModel(ModelMixin, ConfigMixin, UNet2DConditionLoadersMixin)
             down_block = get_down_block(
                 down_block_type,
                 num_layers=layers_per_block[i],
+                transformer_layers_per_block=transformer_layers_per_block[i],
                 in_channels=input_channel,
                 out_channels=output_channel,
                 temb_channels=blocks_time_embed_dim,
@@ -426,6 +442,7 @@ class UNet2DConditionModel(ModelMixin, ConfigMixin, UNet2DConditionLoadersMixin)
         # mid
         if mid_block_type == "UNetMidBlock2DCrossAttn":
             self.mid_block = UNetMidBlock2DCrossAttn(
+                transformer_layers_per_block=transformer_layers_per_block[-1],
                 in_channels=block_out_channels[-1],
                 temb_channels=blocks_time_embed_dim,
                 resnet_eps=norm_eps,
@@ -467,6 +484,7 @@ class UNet2DConditionModel(ModelMixin, ConfigMixin, UNet2DConditionLoadersMixin)
         reversed_num_attention_heads = list(reversed(num_attention_heads))
         reversed_layers_per_block = list(reversed(layers_per_block))
         reversed_cross_attention_dim = list(reversed(cross_attention_dim))
+        reversed_transformer_layers_per_block = list(reversed(transformer_layers_per_block))
         only_cross_attention = list(reversed(only_cross_attention))
 
         output_channel = reversed_block_out_channels[0]
@@ -487,6 +505,7 @@ class UNet2DConditionModel(ModelMixin, ConfigMixin, UNet2DConditionLoadersMixin)
             up_block = get_up_block(
                 up_block_type,
                 num_layers=reversed_layers_per_block[i] + 1,
+                transformer_layers_per_block=reversed_transformer_layers_per_block[i],
                 in_channels=input_channel,
                 out_channels=output_channel,
                 prev_output_channel=prev_output_channel,
@@ -693,6 +712,9 @@ class UNet2DConditionModel(ModelMixin, ConfigMixin, UNet2DConditionLoadersMixin)
                 tuple.
             cross_attention_kwargs (`dict`, *optional*):
                 A kwargs dictionary that if specified is passed along to the [`AttnProcessor`].
+            added_cond_kwargs: (`dict`, *optional*):
+                A kwargs dictionary containin additional embeddings that if specified are added to the embeddings that
+                are passed along to the UNet blocks.
 
         Returns:
             [`~models.unet_2d_condition.UNet2DConditionOutput`] or `tuple`:
@@ -763,6 +785,7 @@ class UNet2DConditionModel(ModelMixin, ConfigMixin, UNet2DConditionLoadersMixin)
         t_emb = t_emb.to(dtype=sample.dtype)
 
         emb = self.time_embedding(t_emb, timestep_cond)
+        aug_emb = None
 
         if self.class_embedding is not None:
             if class_labels is None:
@@ -784,7 +807,6 @@ class UNet2DConditionModel(ModelMixin, ConfigMixin, UNet2DConditionLoadersMixin)
 
         if self.config.addition_embed_type == "text":
             aug_emb = self.add_embedding(encoder_hidden_states)
-            emb = emb + aug_emb
         elif self.config.addition_embed_type == "text_image":
             # Kadinsky 2.1 - style
             if "image_embeds" not in added_cond_kwargs:
@@ -796,7 +818,25 @@ class UNet2DConditionModel(ModelMixin, ConfigMixin, UNet2DConditionLoadersMixin)
             text_embs = added_cond_kwargs.get("text_embeds", encoder_hidden_states)
 
             aug_emb = self.add_embedding(text_embs, image_embs)
-            emb = emb + aug_emb
+        elif self.config.addition_embed_type == "text_time":
+            if "text_embeds" not in added_cond_kwargs:
+                raise ValueError(
+                    f"{self.__class__} has the config param `addition_embed_type` set to 'text_time' which requires the keyword argument `text_embeds` to be passed in `added_cond_kwargs`"
+                )
+            text_embeds = added_cond_kwargs.get("text_embeds")
+            if "time_ids" not in added_cond_kwargs:
+                raise ValueError(
+                    f"{self.__class__} has the config param `addition_embed_type` set to 'text_time' which requires the keyword argument `time_ids` to be passed in `added_cond_kwargs`"
+                )
+            time_ids = added_cond_kwargs.get("time_ids")
+            time_embeds = self.add_time_proj(time_ids.flatten())
+            time_embeds = time_embeds.reshape((text_embeds.shape[0], -1))
+
+            add_embeds = torch.concat([text_embeds, time_embeds], dim=-1)
+            add_embeds = add_embeds.to(emb.dtype)
+            aug_emb = self.add_embedding(add_embeds)
+
+        emb = emb + aug_emb if aug_emb is not None else emb
 
         if self.time_embed_act is not None:
             emb = self.time_embed_act(emb)

src/diffusers/pipelines/__init__.py

 from ..utils import (
     OptionalDependencyNotAvailable,
     is_flax_available,
+    is_invisible_watermark_available,
     is_k_diffusion_available,
     is_librosa_available,
     is_note_seq_available,
@@ -101,6 +102,15 @@ else:
     )
     from .vq_diffusion import VQDiffusionPipeline
 
+
+try:
+    if not (is_torch_available() and is_transformers_available() and is_invisible_watermark_available()):
+        raise OptionalDependencyNotAvailable()
+except OptionalDependencyNotAvailable:
+    from ..utils.dummy_torch_and_transformers_and_invisible_watermark_objects import *  # noqa F403
+else:
+    from .stable_diffusion_xl import StableDiffusionXLImg2ImgPipeline, StableDiffusionXLPipeline
+
 try:
     if not is_onnx_available():
         raise OptionalDependencyNotAvailable()

src/diffusers/pipelines/stable_diffusion/convert_from_ckpt.py

@@ -233,7 +233,10 @@ def create_unet_diffusers_config(original_config, image_size: int, controlnet=Fa
     if controlnet:
         unet_params = original_config.model.params.control_stage_config.params
     else:
+        if original_config.model.params.unet_config is not None:
             unet_params = original_config.model.params.unet_config.params
+        else:
+            unet_params = original_config.model.params.network_config.params
 
     vae_params = original_config.model.params.first_stage_config.params.ddconfig
 
@@ -253,6 +256,15 @@ def create_unet_diffusers_config(original_config, image_size: int, controlnet=Fa
         up_block_types.append(block_type)
         resolution //= 2
 
+    if unet_params.transformer_depth is not None:
+        transformer_layers_per_block = (
+            unet_params.transformer_depth
+            if isinstance(unet_params.transformer_depth, int)
+            else list(unet_params.transformer_depth)
+        )
+    else:
+        transformer_layers_per_block = 1
+
     vae_scale_factor = 2 ** (len(vae_params.ch_mult) - 1)
 
     head_dim = unet_params.num_heads if "num_heads" in unet_params else None
@@ -262,13 +274,27 @@ def create_unet_diffusers_config(original_config, image_size: int, controlnet=Fa
     if use_linear_projection:
         # stable diffusion 2-base-512 and 2-768
         if head_dim is None:
-            head_dim = [5, 10, 20, 20]
+            head_dim_mult = unet_params.model_channels // unet_params.num_head_channels
+            head_dim = [head_dim_mult * c for c in list(unet_params.channel_mult)]
 
     class_embed_type = None
+    addition_embed_type = None
+    addition_time_embed_dim = None
     projection_class_embeddings_input_dim = None
+    context_dim = None
+
+    if unet_params.context_dim is not None:
+        context_dim = (
+            unet_params.context_dim if isinstance(unet_params.context_dim, int) else unet_params.context_dim[0]
+        )
 
     if "num_classes" in unet_params:
         if unet_params.num_classes == "sequential":
+            if context_dim in [2048, 1280]:
+                # SDXL
+                addition_embed_type = "text_time"
+                addition_time_embed_dim = 256
+            else:
                 class_embed_type = "projection"
             assert "adm_in_channels" in unet_params
             projection_class_embeddings_input_dim = unet_params.adm_in_channels
@@ -281,11 +307,14 @@ def create_unet_diffusers_config(original_config, image_size: int, controlnet=Fa
         "down_block_types": tuple(down_block_types),
         "block_out_channels": tuple(block_out_channels),
         "layers_per_block": unet_params.num_res_blocks,
-        "cross_attention_dim": unet_params.context_dim,
+        "cross_attention_dim": context_dim,
         "attention_head_dim": head_dim,
         "use_linear_projection": use_linear_projection,
         "class_embed_type": class_embed_type,
+        "addition_embed_type": addition_embed_type,
+        "addition_time_embed_dim": addition_time_embed_dim,
         "projection_class_embeddings_input_dim": projection_class_embeddings_input_dim,
+        "transformer_layers_per_block": transformer_layers_per_block,
     }
 
     if controlnet:
@@ -400,6 +429,12 @@ def convert_ldm_unet_checkpoint(
     else:
         raise NotImplementedError(f"Not implemented `class_embed_type`: {config['class_embed_type']}")
 
+    if config["addition_embed_type"] == "text_time":
+        new_checkpoint["add_embedding.linear_1.weight"] = unet_state_dict["label_emb.0.0.weight"]
+        new_checkpoint["add_embedding.linear_1.bias"] = unet_state_dict["label_emb.0.0.bias"]
+        new_checkpoint["add_embedding.linear_2.weight"] = unet_state_dict["label_emb.0.2.weight"]
+        new_checkpoint["add_embedding.linear_2.bias"] = unet_state_dict["label_emb.0.2.bias"]
+
     new_checkpoint["conv_in.weight"] = unet_state_dict["input_blocks.0.0.weight"]
     new_checkpoint["conv_in.bias"] = unet_state_dict["input_blocks.0.0.bias"]
 
@@ -745,9 +780,12 @@ def convert_ldm_clip_checkpoint(checkpoint, local_files_only=False, text_encoder
 
     text_model_dict = {}
 
+    remove_prefixes = ["cond_stage_model.transformer", "conditioner.embedders.0.transformer"]
+
     for key in keys:
-        if key.startswith("cond_stage_model.transformer"):
-            text_model_dict[key[len("cond_stage_model.transformer.") :]] = checkpoint[key]
+        for prefix in remove_prefixes:
+            if key.startswith(prefix):
+                text_model_dict[key[len(prefix + ".") :]] = checkpoint[key]
 
     text_model.load_state_dict(text_model_dict)
 
@@ -755,10 +793,11 @@ def convert_ldm_clip_checkpoint(checkpoint, local_files_only=False, text_encoder
 
 
 textenc_conversion_lst = [
-    ("cond_stage_model.model.positional_embedding", "text_model.embeddings.position_embedding.weight"),
-    ("cond_stage_model.model.token_embedding.weight", "text_model.embeddings.token_embedding.weight"),
-    ("cond_stage_model.model.ln_final.weight", "text_model.final_layer_norm.weight"),
-    ("cond_stage_model.model.ln_final.bias", "text_model.final_layer_norm.bias"),
+    ("positional_embedding", "text_model.embeddings.position_embedding.weight"),
+    ("token_embedding.weight", "text_model.embeddings.token_embedding.weight"),
+    ("ln_final.weight", "text_model.final_layer_norm.weight"),
+    ("ln_final.bias", "text_model.final_layer_norm.bias"),
+    ("text_projection", "text_projection.weight"),
 ]
 textenc_conversion_map = {x[0]: x[1] for x in textenc_conversion_lst}
 
@@ -845,27 +884,36 @@ def convert_paint_by_example_checkpoint(checkpoint):
     return model
 
 
-def convert_open_clip_checkpoint(checkpoint):
-    text_model = CLIPTextModel.from_pretrained("stabilityai/stable-diffusion-2", subfolder="text_encoder")
+def convert_open_clip_checkpoint(checkpoint, prefix="cond_stage_model.model."):
+    # 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
+    )
 
     keys = list(checkpoint.keys())
 
     text_model_dict = {}
 
-    if "cond_stage_model.model.text_projection" in checkpoint:
-        d_model = int(checkpoint["cond_stage_model.model.text_projection"].shape[0])
+    if prefix + "text_projection" in checkpoint:
+        d_model = int(checkpoint[prefix + "text_projection"].shape[0])
     else:
         d_model = 1024
 
     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 textenc_conversion_map:
-            text_model_dict[textenc_conversion_map[key]] = checkpoint[key]
-        if key.startswith("cond_stage_model.model.transformer."):
-            new_key = key[len("cond_stage_model.model.transformer.") :]
+        # if "resblocks.23" in key:  # Diffusers drops the final layer and only uses the penultimate layer
+        #     continue
+        if key[len(prefix) :] in textenc_conversion_map:
+            if key.endswith("text_projection"):
+                value = checkpoint[key].T
+            else:
+                value = checkpoint[key]
+
+            text_model_dict[textenc_conversion_map[key[len(prefix) :]]] = value
+
+        if key.startswith(prefix + "transformer."):
+            new_key = key[len(prefix + "transformer.") :]
             if new_key.endswith(".in_proj_weight"):
                 new_key = new_key[: -len(".in_proj_weight")]
                 new_key = textenc_pattern.sub(lambda m: protected[re.escape(m.group(0))], new_key)
@@ -1029,6 +1077,7 @@ def download_from_original_stable_diffusion_ckpt(
     load_safety_checker: bool = True,
     pipeline_class: DiffusionPipeline = None,
     local_files_only=False,
+    vae_path=None,
     text_encoder=None,
     tokenizer=None,
 ) -> DiffusionPipeline:
@@ -1096,6 +1145,8 @@ def download_from_original_stable_diffusion_ckpt(
         PaintByExamplePipeline,
         StableDiffusionControlNetPipeline,
         StableDiffusionPipeline,
+        StableDiffusionXLImg2ImgPipeline,
+        StableDiffusionXLPipeline,
         StableUnCLIPImg2ImgPipeline,
         StableUnCLIPPipeline,
     )
@@ -1187,9 +1238,9 @@ def download_from_original_stable_diffusion_ckpt(
             checkpoint, original_config, checkpoint_path, image_size, upcast_attention, extract_ema
         )
 
-    num_train_timesteps = original_config.model.params.timesteps
-    beta_start = original_config.model.params.linear_start
-    beta_end = original_config.model.params.linear_end
+    num_train_timesteps = original_config.model.params.timesteps or 1000
+    beta_start = original_config.model.params.linear_start or 0.02
+    beta_end = original_config.model.params.linear_end or 0.085
 
     scheduler = DDIMScheduler(
         beta_end=beta_end,
@@ -1231,20 +1282,27 @@ def download_from_original_stable_diffusion_ckpt(
     converted_unet_checkpoint = convert_ldm_unet_checkpoint(
         checkpoint, unet_config, path=checkpoint_path, extract_ema=extract_ema
     )
-
     unet.load_state_dict(converted_unet_checkpoint)
 
     # 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)
+    else:
+        vae = AutoencoderKL.from_pretrained(vae_path)
 
     # Convert the text model.
-    if model_type is None:
+    if model_type is None and original_config.model.params.cond_stage_config is not None:
         model_type = original_config.model.params.cond_stage_config.target.split(".")[-1]
         logger.debug(f"no `model_type` given, `model_type` inferred as: {model_type}")
+    elif model_type is None and original_config.model.params.network_config is not None:
+        if original_config.model.params.network_config.params.context_dim == 2048:
+            model_type = "SDXL"
+        else:
+            model_type = "SDXL-Refiner"
 
     if model_type == "FrozenOpenCLIPEmbedder":
         text_model = convert_open_clip_checkpoint(checkpoint)
@@ -1375,6 +1433,40 @@ def download_from_original_stable_diffusion_ckpt(
                 safety_checker=safety_checker,
                 feature_extractor=feature_extractor,
             )
+    elif model_type in ["SDXL", "SDXL-Refiner"]:
+        if model_type == "SDXL":
+            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.")
+
+            pipe = StableDiffusionXLPipeline(
+                vae=vae,
+                text_encoder=text_encoder,
+                tokenizer=tokenizer,
+                text_encoder_2=text_encoder_2,
+                tokenizer_2=tokenizer_2,
+                unet=unet,
+                scheduler=scheduler,
+                force_zeros_for_empty_prompt=True,
+            )
+        else:
+            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.")
+
+            pipe = StableDiffusionXLImg2ImgPipeline(
+                vae=vae,
+                text_encoder=text_encoder,
+                tokenizer=tokenizer,
+                text_encoder_2=text_encoder_2,
+                tokenizer_2=tokenizer_2,
+                unet=unet,
+                scheduler=scheduler,
+                requires_aesthetics_score=True,
+                force_zeros_for_empty_prompt=False,
+            )
     else:
         text_config = create_ldm_bert_config(original_config)
         text_model = convert_ldm_bert_checkpoint(checkpoint, text_config)

src/diffusers/pipelines/stable_diffusion/pipeline_stable_diffusion_upscale.py

@@ -24,7 +24,12 @@ from transformers import CLIPImageProcessor, CLIPTextModel, CLIPTokenizer
 from ...image_processor import VaeImageProcessor
 from ...loaders import LoraLoaderMixin, TextualInversionLoaderMixin
 from ...models import AutoencoderKL, UNet2DConditionModel
-from ...models.attention_processor import AttnProcessor2_0, LoRAXFormersAttnProcessor, XFormersAttnProcessor
+from ...models.attention_processor import (
+    AttnProcessor2_0,
+    LoRAAttnProcessor2_0,
+    LoRAXFormersAttnProcessor,
+    XFormersAttnProcessor,
+)
 from ...schedulers import DDPMScheduler, KarrasDiffusionSchedulers
 from ...utils import deprecate, is_accelerate_available, is_accelerate_version, logging, randn_tensor
 from ..pipeline_utils import DiffusionPipeline
@@ -747,6 +752,7 @@ class StableDiffusionUpscalePipeline(DiffusionPipeline, TextualInversionLoaderMi
             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

src/diffusers/pipelines/stable_diffusion_xl/__init__.py

+from dataclasses import dataclass
+from typing import List, Optional, Union
+
+import numpy as np
+import PIL
+
+from ...utils import BaseOutput, is_invisible_watermark_available, is_torch_available, is_transformers_available
+
+
+@dataclass
+# Copied from diffusers.pipelines.stable_diffusion.__init__.StableDiffusionPipelineOutput with StableDiffusion->StableDiffusionXL
+class StableDiffusionXLPipelineOutput(BaseOutput):
+    """
+    Output class for Stable Diffusion pipelines.
+
+    Args:
+        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():
+    from .pipeline_stable_diffusion_xl import StableDiffusionXLPipeline
+    from .pipeline_stable_diffusion_xl_img2img import StableDiffusionXLImg2ImgPipeline

src/diffusers/pipelines/stable_diffusion_xl/pipeline_stable_diffusion_xl.py

+# Copyright 2023 The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import inspect
+from typing import Any, Callable, Dict, List, Optional, Tuple, Union
+
+import torch
+from transformers import CLIPTextModel, CLIPTextModelWithProjection, CLIPTokenizer
+
+from ...image_processor import VaeImageProcessor
+from ...loaders import LoraLoaderMixin, TextualInversionLoaderMixin
+from ...models import AutoencoderKL, UNet2DConditionModel
+from ...models.attention_processor import (
+    AttnProcessor2_0,
+    LoRAAttnProcessor2_0,
+    LoRAXFormersAttnProcessor,
+    XFormersAttnProcessor,
+)
+from ...schedulers import KarrasDiffusionSchedulers
+from ...utils import (
+    is_accelerate_available,
+    is_accelerate_version,
+    logging,
+    randn_tensor,
+    replace_example_docstring,
+)
+from ..pipeline_utils import DiffusionPipeline
+from . import StableDiffusionXLPipelineOutput
+from .watermark import StableDiffusionXLWatermarker
+
+
+logger = logging.get_logger(__name__)  # pylint: disable=invalid-name
+
+EXAMPLE_DOC_STRING = """
+    Examples:
+        ```py
+        >>> import torch
+        >>> from diffusers import StableDiffusionXLPipeline
+
+        >>> pipe = StableDiffusionXLPipeline.from_pretrained(
+        ...     "stabilityai/stable-diffusion-xl-base-0.9", torch_dtype=torch.float16
+        ... )
+        >>> pipe = pipe.to("cuda")
+
+        >>> prompt = "a photo of an astronaut riding a horse on mars"
+        >>> image = pipe(prompt).images[0]
+        ```
+"""
+
+
+def rescale_noise_cfg(noise_cfg, noise_pred_text, guidance_rescale=0.0):
+    """
+    Rescale `noise_cfg` according to `guidance_rescale`. Based on findings of [Common Diffusion Noise Schedules and
+    Sample Steps are Flawed](https://arxiv.org/pdf/2305.08891.pdf). See Section 3.4
+    """
+    std_text = noise_pred_text.std(dim=list(range(1, noise_pred_text.ndim)), keepdim=True)
+    std_cfg = noise_cfg.std(dim=list(range(1, noise_cfg.ndim)), keepdim=True)
+    # rescale the results from guidance (fixes overexposure)
+    noise_pred_rescaled = noise_cfg * (std_text / std_cfg)
+    # mix with the original results from guidance by factor guidance_rescale to avoid "plain looking" images
+    noise_cfg = guidance_rescale * noise_pred_rescaled + (1 - guidance_rescale) * noise_cfg
+    return noise_cfg
+
+
+class StableDiffusionXLPipeline(DiffusionPipeline):
+    r"""
+    Pipeline for text-to-image generation using Stable Diffusion.
+
+    This model inherits from [`DiffusionPipeline`]. Check the superclass documentation for the generic methods the
+    library implements for all the pipelines (such as downloading or saving, running on a particular device, etc.)
+
+    In addition the pipeline inherits the following loading methods:
+        - *Textual-Inversion*: [`loaders.TextualInversionLoaderMixin.load_textual_inversion`]
+        - *LoRA*: [`loaders.LoraLoaderMixin.load_lora_weights`]
+        - *Ckpt*: [`loaders.FromCkptMixin.from_ckpt`]
+
+    as well as the following saving methods:
+        - *LoRA*: [`loaders.LoraLoaderMixin.save_lora_weights`]
+
+    Args:
+        vae ([`AutoencoderKL`]):
+            Variational Auto-Encoder (VAE) Model to encode and decode images to and from latent representations.
+        text_encoder ([`CLIPTextModel`]):
+            Frozen text-encoder. Stable Diffusion uses the text portion of
+            [CLIP](https://huggingface.co/docs/transformers/model_doc/clip#transformers.CLIPTextModel), specifically
+            the [clip-vit-large-patch14](https://huggingface.co/openai/clip-vit-large-patch14) variant.
+        tokenizer (`CLIPTokenizer`):
+            Tokenizer of class
+            [CLIPTokenizer](https://huggingface.co/docs/transformers/v4.21.0/en/model_doc/clip#transformers.CLIPTokenizer).
+        unet ([`UNet2DConditionModel`]): Conditional U-Net architecture to denoise the encoded image latents.
+        scheduler ([`SchedulerMixin`]):
+            A scheduler to be used in combination with `unet` to denoise the encoded image latents. Can be one of
+            [`DDIMScheduler`], [`LMSDiscreteScheduler`], or [`PNDMScheduler`].
+    """
+
+    def __init__(
+        self,
+        vae: AutoencoderKL,
+        text_encoder: CLIPTextModel,
+        text_encoder_2: CLIPTextModelWithProjection,
+        tokenizer: CLIPTokenizer,
+        tokenizer_2: CLIPTokenizer,
+        unet: UNet2DConditionModel,
+        scheduler: KarrasDiffusionSchedulers,
+        force_zeros_for_empty_prompt: bool = True,
+    ):
+        super().__init__()
+
+        self.register_modules(
+            vae=vae,
+            text_encoder=text_encoder,
+            text_encoder_2=text_encoder_2,
+            tokenizer=tokenizer,
+            tokenizer_2=tokenizer_2,
+            unet=unet,
+            scheduler=scheduler,
+        )
+        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.watermark = StableDiffusionXLWatermarker()
+
+    def enable_vae_slicing(self):
+        r"""
+        Enable sliced VAE decoding.
+
+        When this option is enabled, the VAE will split the input tensor in slices to compute decoding in several
+        steps. This is useful to save some memory and allow larger batch sizes.
+        """
+        self.vae.enable_slicing()
+
+    def disable_vae_slicing(self):
+        r"""
+        Disable sliced VAE decoding. If `enable_vae_slicing` was previously invoked, this method will go back to
+        computing decoding in one step.
+        """
+        self.vae.disable_slicing()
+
+    def enable_vae_tiling(self):
+        r"""
+        Enable tiled VAE decoding.
+
+        When this option is enabled, the VAE will split the input tensor into tiles to compute decoding and encoding in
+        several steps. This is useful to save a large amount of memory and to allow the processing of larger images.
+        """
+        self.vae.enable_tiling()
+
+    def disable_vae_tiling(self):
+        r"""
+        Disable tiled VAE decoding. If `enable_vae_tiling` was previously invoked, this method will go back to
+        computing decoding in one step.
+        """
+        self.vae.disable_tiling()
+
+    def enable_sequential_cpu_offload(self, gpu_id=0):
+        r"""
+        Offloads all models to CPU using accelerate, significantly reducing memory usage. When called, unet,
+        text_encoder, vae and safety checker have their state dicts saved to CPU and then are moved to a
+        `torch.device('meta') and loaded to GPU only when their specific submodule has its `forward` method called.
+        Note that offloading happens on a submodule basis. Memory savings are higher than with
+        `enable_model_cpu_offload`, but performance is lower.
+        """
+        if is_accelerate_available() and is_accelerate_version(">=", "0.14.0"):
+            from accelerate import cpu_offload
+        else:
+            raise ImportError("`enable_sequential_cpu_offload` requires `accelerate v0.14.0` or higher")
+
+        device = torch.device(f"cuda:{gpu_id}")
+
+        if self.device.type != "cpu":
+            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]:
+            cpu_offload(cpu_offloaded_model, device)
+
+    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
+        to `enable_sequential_cpu_offload`, this method moves one whole model at a time to the GPU when its `forward`
+        method is called, and the model remains in GPU until the next model runs. Memory savings are lower than with
+        `enable_sequential_cpu_offload`, but performance is much better due to the iterative execution of the `unet`.
+        """
+        if is_accelerate_available() and is_accelerate_version(">=", "0.17.0.dev0"):
+            from accelerate import cpu_offload_with_hook
+        else:
+            raise ImportError("`enable_model_cpu_offload` requires `accelerate v0.17.0` or higher.")
+
+        device = torch.device(f"cuda:{gpu_id}")
+
+        if self.device.type != "cpu":
+            self.to("cpu", silence_dtype_warnings=True)
+            torch.cuda.empty_cache()  # otherwise we don't see the memory savings (but they probably exist)
+
+        model_sequence = (
+            [self.text_encoder, self.text_encoder_2] if self.text_encoder is not None else [self.text_encoder_2]
+        )
+        model_sequence.extend([self.unet, self.vae])
+
+        hook = None
+        for cpu_offloaded_model in model_sequence:
+            _, hook = cpu_offload_with_hook(cpu_offloaded_model, device, prev_module_hook=hook)
+
+        # We'll offload the last model manually.
+        self.final_offload_hook = hook
+
+    @property
+    def _execution_device(self):
+        r"""
+        Returns the device on which the pipeline's models will be executed. After calling
+        `pipeline.enable_sequential_cpu_offload()` the execution device can only be inferred from Accelerate's module
+        hooks.
+        """
+        if not hasattr(self.unet, "_hf_hook"):
+            return self.device
+        for module in self.unet.modules():
+            if (
+                hasattr(module, "_hf_hook")
+                and hasattr(module._hf_hook, "execution_device")
+                and module._hf_hook.execution_device is not None
+            ):
+                return torch.device(module._hf_hook.execution_device)
+        return self.device
+
+    def encode_prompt(
+        self,
+        prompt,
+        device,
+        num_images_per_prompt,
+        do_classifier_free_guidance,
+        negative_prompt=None,
+        prompt_embeds: Optional[torch.FloatTensor] = None,
+        negative_prompt_embeds: Optional[torch.FloatTensor] = None,
+        lora_scale: Optional[float] = None,
+    ):
+        r"""
+        Encodes the prompt into text encoder hidden states.
+
+        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
+            do_classifier_free_guidance (`bool`):
+                whether to use classifier free guidance or not
+            negative_prompt (`str` or `List[str]`, *optional*):
+                The prompt or prompts not to guide the image generation. If not defined, one has to pass
+                `negative_prompt_embeds` instead. Ignored when not using guidance (i.e., ignored if `guidance_scale` is
+                less than `1`).
+            prompt_embeds (`torch.FloatTensor`, *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.FloatTensor`, *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.
+            lora_scale (`float`, *optional*):
+                A lora scale that will be applied to all LoRA layers of the text encoder if LoRA layers are loaded.
+        """
+        # 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):
+            self._lora_scale = lora_scale
+
+        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]
+
+        # Define tokenizers and text encoders
+        tokenizers = [self.tokenizer, self.tokenizer_2] if self.tokenizer is not None else [self.tokenizer_2]
+        text_encoders = (
+            [self.text_encoder, self.text_encoder_2] if self.text_encoder is not None else [self.text_encoder_2]
+        )
+
+        if prompt_embeds is None:
+            # textual inversion: procecss multi-vector tokens if necessary
+            prompt_embeds_list = []
+            for tokenizer, text_encoder in zip(tokenizers, text_encoders):
+                if isinstance(self, TextualInversionLoaderMixin):
+                    prompt = self.maybe_convert_prompt(prompt, tokenizer)
+
+                text_inputs = tokenizer(
+                    prompt,
+                    padding="max_length",
+                    max_length=tokenizer.model_max_length,
+                    truncation=True,
+                    return_tensors="pt",
+                )
+                text_input_ids = text_inputs.input_ids
+                untruncated_ids = tokenizer(prompt, padding="longest", return_tensors="pt").input_ids
+
+                if untruncated_ids.shape[-1] >= text_input_ids.shape[-1] and not torch.equal(
+                    text_input_ids, untruncated_ids
+                ):
+                    removed_text = tokenizer.batch_decode(untruncated_ids[:, tokenizer.model_max_length - 1 : -1])
+                    logger.warning(
+                        "The following part of your input was truncated because CLIP can only handle sequences up to"
+                        f" {tokenizer.model_max_length} tokens: {removed_text}"
+                    )
+
+                prompt_embeds = text_encoder(
+                    text_input_ids.to(device),
+                    output_hidden_states=True,
+                )
+
+                # We are only ALWAYS interested in the pooled output of the final text encoder
+                pooled_prompt_embeds = prompt_embeds[0]
+                prompt_embeds = prompt_embeds.hidden_states[-2]
+
+                bs_embed, seq_len, _ = prompt_embeds.shape
+                # duplicate text embeddings for each generation per prompt, using mps friendly method
+                prompt_embeds = prompt_embeds.repeat(1, num_images_per_prompt, 1)
+                prompt_embeds = prompt_embeds.view(bs_embed * num_images_per_prompt, seq_len, -1)
+
+                prompt_embeds_list.append(prompt_embeds)
+
+            prompt_embeds = torch.concat(prompt_embeds_list, dim=-1)
+
+        # get unconditional embeddings for classifier free guidance
+        zero_out_negative_prompt = negative_prompt is None and self.config.force_zeros_for_empty_prompt
+        if do_classifier_free_guidance and negative_prompt_embeds is None and zero_out_negative_prompt:
+            negative_prompt_embeds = torch.zeros_like(prompt_embeds)
+            negative_pooled_prompt_embeds = torch.zeros_like(pooled_prompt_embeds)
+        elif do_classifier_free_guidance and negative_prompt_embeds is None:
+            negative_prompt = negative_prompt or ""
+            uncond_tokens: List[str]
+            if prompt is not None and type(prompt) is not type(negative_prompt):
+                raise TypeError(
+                    f"`negative_prompt` should be the same type to `prompt`, but got {type(negative_prompt)} !="
+                    f" {type(prompt)}."
+                )
+            elif isinstance(negative_prompt, str):
+                uncond_tokens = [negative_prompt]
+            elif batch_size != len(negative_prompt):
+                raise ValueError(
+                    f"`negative_prompt`: {negative_prompt} has batch size {len(negative_prompt)}, but `prompt`:"
+                    f" {prompt} has batch size {batch_size}. Please make sure that passed `negative_prompt` matches"
+                    " the batch size of `prompt`."
+                )
+            else:
+                uncond_tokens = negative_prompt
+
+            negative_prompt_embeds_list = []
+            for tokenizer, text_encoder in zip(tokenizers, text_encoders):
+                # textual inversion: procecss multi-vector tokens if necessary
+                if isinstance(self, TextualInversionLoaderMixin):
+                    uncond_tokens = self.maybe_convert_prompt(uncond_tokens, tokenizer)
+
+                max_length = prompt_embeds.shape[1]
+                uncond_input = tokenizer(
+                    uncond_tokens,
+                    padding="max_length",
+                    max_length=max_length,
+                    truncation=True,
+                    return_tensors="pt",
+                )
+
+                negative_prompt_embeds = text_encoder(
+                    uncond_input.input_ids.to(device),
+                    output_hidden_states=True,
+                )
+                # We are only ALWAYS interested in the pooled output of the final text encoder
+                negative_pooled_prompt_embeds = negative_prompt_embeds[0]
+                negative_prompt_embeds = negative_prompt_embeds.hidden_states[-2]
+
+                if do_classifier_free_guidance:
+                    # duplicate unconditional embeddings for each generation per prompt, using mps friendly method
+                    seq_len = negative_prompt_embeds.shape[1]
+
+                    negative_prompt_embeds = negative_prompt_embeds.to(dtype=text_encoder.dtype, device=device)
+
+                    negative_prompt_embeds = negative_prompt_embeds.repeat(1, num_images_per_prompt, 1)
+                    negative_prompt_embeds = negative_prompt_embeds.view(
+                        batch_size * num_images_per_prompt, seq_len, -1
+                    )
+
+                    # For classifier free guidance, we need to do two forward passes.
+                    # Here we concatenate the unconditional and text embeddings into a single batch
+                    # to avoid doing two forward passes
+
+                negative_prompt_embeds_list.append(negative_prompt_embeds)
+
+            negative_prompt_embeds = torch.concat(negative_prompt_embeds_list, dim=-1)
+
+        pooled_prompt_embeds = pooled_prompt_embeds.repeat(1, num_images_per_prompt).view(
+            bs_embed * num_images_per_prompt, -1
+        )
+        negative_pooled_prompt_embeds = negative_pooled_prompt_embeds.repeat(1, num_images_per_prompt).view(
+            bs_embed * num_images_per_prompt, -1
+        )
+
+        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
+
+    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.
+        # eta corresponds to η in DDIM paper: https://arxiv.org/abs/2010.02502
+        # and should be between [0, 1]
+
+        accepts_eta = "eta" in set(inspect.signature(self.scheduler.step).parameters.keys())
+        extra_step_kwargs = {}
+        if accepts_eta:
+            extra_step_kwargs["eta"] = eta
+
+        # check if the scheduler accepts generator
+        accepts_generator = "generator" in set(inspect.signature(self.scheduler.step).parameters.keys())
+        if accepts_generator:
+            extra_step_kwargs["generator"] = generator
+        return extra_step_kwargs
+
+    def check_inputs(
+        self,
+        prompt,
+        height,
+        width,
+        callback_steps,
+        negative_prompt=None,
+        prompt_embeds=None,
+        negative_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}.")
+
+        if (callback_steps is None) or (
+            callback_steps is not None and (not isinstance(callback_steps, int) or callback_steps <= 0)
+        ):
+            raise ValueError(
+                f"`callback_steps` has to be a positive integer but is {callback_steps} of type"
+                f" {type(callback_steps)}."
+            )
+
+        if prompt is not None and prompt_embeds is not None:
+            raise ValueError(
+                f"Cannot forward both `prompt`: {prompt} and `prompt_embeds`: {prompt_embeds}. Please make sure to"
+                " only forward one of the two."
+            )
+        elif prompt is None and prompt_embeds is None:
+            raise ValueError(
+                "Provide either `prompt` or `prompt_embeds`. Cannot leave both `prompt` and `prompt_embeds` undefined."
+            )
+        elif prompt is not None and (not isinstance(prompt, str) and not isinstance(prompt, list)):
+            raise ValueError(f"`prompt` has to be of type `str` or `list` but is {type(prompt)}")
+
+        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 negative_prompt_embeds is not None:
+            if prompt_embeds.shape != negative_prompt_embeds.shape:
+                raise ValueError(
+                    "`prompt_embeds` and `negative_prompt_embeds` must have the same shape when passed directly, but"
+                    f" got: `prompt_embeds` {prompt_embeds.shape} != `negative_prompt_embeds`"
+                    f" {negative_prompt_embeds.shape}."
+                )
+
+    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:
+            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)
+        else:
+            latents = latents.to(device)
+
+        # scale the initial noise by the standard deviation required by the scheduler
+        latents = latents * self.scheduler.init_noise_sigma
+        return latents
+
+    def _get_add_time_ids(self, original_size, crops_coords_top_left, target_size, dtype):
+        add_time_ids = list(original_size + crops_coords_top_left + target_size)
+
+        passed_add_embed_dim = (
+            self.unet.config.addition_time_embed_dim * len(add_time_ids) + self.text_encoder_2.config.projection_dim
+        )
+        expected_add_embed_dim = self.unet.add_embedding.linear_1.in_features
+
+        if expected_add_embed_dim != passed_add_embed_dim:
+            raise ValueError(
+                f"Model expects an added time embedding vector of length {expected_add_embed_dim}, but a vector of {passed_add_embed_dim} was created. The model has an incorrect config. Please check `unet.config.time_embedding_type` and `text_encoder_2.config.projection_dim`."
+            )
+
+        add_time_ids = torch.tensor([add_time_ids], dtype=dtype)
+        return add_time_ids
+
+    @torch.no_grad()
+    @replace_example_docstring(EXAMPLE_DOC_STRING)
+    def __call__(
+        self,
+        prompt: Union[str, List[str]] = None,
+        height: Optional[int] = None,
+        width: Optional[int] = None,
+        num_inference_steps: int = 50,
+        guidance_scale: float = 5.0,
+        negative_prompt: Optional[Union[str, List[str]]] = None,
+        num_images_per_prompt: Optional[int] = 1,
+        eta: float = 0.0,
+        generator: Optional[Union[torch.Generator, List[torch.Generator]]] = None,
+        latents: Optional[torch.FloatTensor] = None,
+        prompt_embeds: Optional[torch.FloatTensor] = None,
+        negative_prompt_embeds: Optional[torch.FloatTensor] = None,
+        output_type: Optional[str] = "pil",
+        return_dict: bool = True,
+        callback: Optional[Callable[[int, int, torch.FloatTensor], None]] = None,
+        callback_steps: int = 1,
+        cross_attention_kwargs: Optional[Dict[str, Any]] = None,
+        guidance_rescale: float = 0.0,
+        original_size: Tuple[int, int] = (1024, 1024),
+        crops_coords_top_left: Tuple[int, int] = (0, 0),
+        target_size: Tuple[int, int] = (1024, 1024),
+    ):
+        r"""
+        Function invoked when calling the pipeline for generation.
+
+        Args:
+            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.
+            height (`int`, *optional*, defaults to self.unet.config.sample_size * self.vae_scale_factor):
+                The height in pixels of the generated image.
+            width (`int`, *optional*, defaults to self.unet.config.sample_size * self.vae_scale_factor):
+                The width in pixels of the generated image.
+            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.
+            guidance_scale (`float`, *optional*, defaults to 7.5):
+                Guidance scale as defined in [Classifier-Free Diffusion Guidance](https://arxiv.org/abs/2207.12598).
+                `guidance_scale` is defined as `w` of equation 2. of [Imagen
+                Paper](https://arxiv.org/pdf/2205.11487.pdf). 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.
+            negative_prompt (`str` or `List[str]`, *optional*):
+                The prompt or prompts not to guide the image generation. If not defined, one has to pass
+                `negative_prompt_embeds` instead. Ignored when not using guidance (i.e., ignored if `guidance_scale` is
+                less than `1`).
+            num_images_per_prompt (`int`, *optional*, defaults to 1):
+                The number of images to generate per prompt.
+            eta (`float`, *optional*, defaults to 0.0):
+                Corresponds to parameter eta (η) in the DDIM paper: https://arxiv.org/abs/2010.02502. Only applies to
+                [`schedulers.DDIMScheduler`], will be ignored for others.
+            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.FloatTensor`, *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 ge generated by sampling using the supplied random `generator`.
+            prompt_embeds (`torch.FloatTensor`, *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.FloatTensor`, *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.stable_diffusion.StableDiffusionXLPipelineOutput`] instead of a
+                plain tuple.
+            callback (`Callable`, *optional*):
+                A function that will be called every `callback_steps` steps during inference. The function will be
+                called with the following arguments: `callback(step: int, timestep: int, latents: torch.FloatTensor)`.
+            callback_steps (`int`, *optional*, defaults to 1):
+                The frequency at which the `callback` function will be called. If not specified, the callback will be
+                called at every step.
+            cross_attention_kwargs (`dict`, *optional*):
+                A kwargs dictionary that if specified is passed along to the `AttentionProcessor` as defined under
+                `self.processor` in
+                [diffusers.cross_attention](https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/cross_attention.py).
+            guidance_rescale (`float`, *optional*, defaults to 0.7):
+                Guidance rescale factor proposed by [Common Diffusion Noise Schedules and Sample Steps are
+                Flawed](https://arxiv.org/pdf/2305.08891.pdf) `guidance_scale` is defined as `φ` in equation 16. of
+                [Common Diffusion Noise Schedules and Sample Steps are Flawed](https://arxiv.org/pdf/2305.08891.pdf).
+                Guidance rescale factor should fix overexposure when using zero terminal SNR.
+            original_size (`Tuple[int]`, *optional*, defaults to (1024, 1024)):
+                TODO
+            crops_coords_top_left (`Tuple[int]`, *optional*, defaults to (0, 0)):
+                TODO
+            target_size (`Tuple[int]`, *optional*, defaults to (1024, 1024)):
+                TODO
+
+        Examples:
+
+        Returns:
+            [`~pipelines.stable_diffusion.StableDiffusionXLPipelineOutput`] or `tuple`:
+            [`~pipelines.stable_diffusion.StableDiffusionXLPipelineOutput`] if `return_dict` is True, otherwise a
+            `tuple. When returning a tuple, the first element is a list with the generated images, and the second
+            element is a list of `bool`s denoting whether the corresponding generated image likely represents
+            "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
+
+        # 1. Check inputs. Raise error if not correct
+        self.check_inputs(
+            prompt, height, width, callback_steps, negative_prompt, prompt_embeds, negative_prompt_embeds
+        )
+
+        # 2. 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]
+
+        device = self._execution_device
+
+        # here `guidance_scale` is defined analog to the guidance weight `w` of equation (2)
+        # of the Imagen paper: https://arxiv.org/pdf/2205.11487.pdf . `guidance_scale = 1`
+        # corresponds to doing no classifier free guidance.
+        do_classifier_free_guidance = guidance_scale > 1.0
+
+        # 3. Encode input prompt
+        text_encoder_lora_scale = (
+            cross_attention_kwargs.get("scale", None) if cross_attention_kwargs is not None else None
+        )
+        (
+            prompt_embeds,
+            negative_prompt_embeds,
+            pooled_prompt_embeds,
+            negative_pooled_prompt_embeds,
+        ) = self.encode_prompt(
+            prompt,
+            device,
+            num_images_per_prompt,
+            do_classifier_free_guidance,
+            negative_prompt,
+            prompt_embeds=prompt_embeds,
+            negative_prompt_embeds=negative_prompt_embeds,
+            lora_scale=text_encoder_lora_scale,
+        )
+
+        # 4. Prepare timesteps
+        self.scheduler.set_timesteps(num_inference_steps, device=device)
+
+        timesteps = self.scheduler.timesteps
+
+        # 5. Prepare latent variables
+        num_channels_latents = self.unet.config.in_channels
+        latents = self.prepare_latents(
+            batch_size * num_images_per_prompt,
+            num_channels_latents,
+            height,
+            width,
+            prompt_embeds.dtype,
+            device,
+            generator,
+            latents,
+        )
+
+        # 6. Prepare extra step kwargs. TODO: Logic should ideally just be moved out of the pipeline
+        extra_step_kwargs = self.prepare_extra_step_kwargs(generator, eta)
+
+        # 7. Prepare added time ids & embeddings
+        add_text_embeds = pooled_prompt_embeds
+        add_time_ids = self._get_add_time_ids(
+            original_size, crops_coords_top_left, target_size, dtype=prompt_embeds.dtype
+        )
+
+        if do_classifier_free_guidance:
+            prompt_embeds = torch.cat([negative_prompt_embeds, prompt_embeds], dim=0)
+            add_text_embeds = torch.cat([negative_pooled_prompt_embeds, add_text_embeds], dim=0)
+            add_time_ids = torch.cat([add_time_ids, add_time_ids], dim=0)
+
+        prompt_embeds = prompt_embeds.to(device)
+        add_text_embeds = add_text_embeds.to(device)
+        add_time_ids = add_time_ids.to(device).repeat(batch_size * num_images_per_prompt, 1)
+
+        # 8. Denoising loop
+        num_warmup_steps = len(timesteps) - num_inference_steps * self.scheduler.order
+        with self.progress_bar(total=num_inference_steps) as progress_bar:
+            for i, t in enumerate(timesteps):
+                # expand the latents if we are doing classifier free guidance
+                latent_model_input = torch.cat([latents] * 2) if do_classifier_free_guidance else latents
+
+                latent_model_input = self.scheduler.scale_model_input(latent_model_input, t)
+
+                # predict the noise residual
+                added_cond_kwargs = {"text_embeds": add_text_embeds, "time_ids": add_time_ids}
+                noise_pred = self.unet(
+                    latent_model_input,
+                    t,
+                    encoder_hidden_states=prompt_embeds,
+                    cross_attention_kwargs=cross_attention_kwargs,
+                    added_cond_kwargs=added_cond_kwargs,
+                    return_dict=False,
+                )[0]
+
+                # perform guidance
+                if do_classifier_free_guidance:
+                    noise_pred_uncond, noise_pred_text = noise_pred.chunk(2)
+                    noise_pred = noise_pred_uncond + guidance_scale * (noise_pred_text - noise_pred_uncond)
+
+                if do_classifier_free_guidance and guidance_rescale > 0.0:
+                    # Based on 3.4. in https://arxiv.org/pdf/2305.08891.pdf
+                    noise_pred = rescale_noise_cfg(noise_pred, noise_pred_text, guidance_rescale=guidance_rescale)
+
+                # compute the previous noisy sample x_t -> x_t-1
+                latents = self.scheduler.step(noise_pred, t, latents, **extra_step_kwargs, return_dict=False)[0]
+
+                # 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 callback is not None and i % callback_steps == 0:
+                        callback(i, t, latents)
+
+        # 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,
+        ]
+        # 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:
+            self.vae.post_quant_conv.to(latents.dtype)
+            self.vae.decoder.conv_in.to(latents.dtype)
+            self.vae.decoder.mid_block.to(latents.dtype)
+        else:
+            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]
+
+        image = self.watermark.apply_watermark(image)
+        image = self.image_processor.postprocess(image, output_type=output_type, do_denormalize=do_denormalize)
+
+        # 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 StableDiffusionXLPipelineOutput(images=image, nsfw_content_detected=has_nsfw_concept)

src/diffusers/pipelines/stable_diffusion_xl/pipeline_stable_diffusion_xl_img2img.py

+# Copyright 2023 The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import inspect
+from typing import Any, Callable, Dict, List, Optional, Tuple, Union
+
+import numpy as np
+import PIL.Image
+import torch
+from transformers import CLIPTextModel, CLIPTextModelWithProjection, CLIPTokenizer
+
+from ...image_processor import VaeImageProcessor
+from ...loaders import LoraLoaderMixin, TextualInversionLoaderMixin
+from ...models import AutoencoderKL, UNet2DConditionModel
+from ...models.attention_processor import (
+    AttnProcessor2_0,
+    LoRAAttnProcessor2_0,
+    LoRAXFormersAttnProcessor,
+    XFormersAttnProcessor,
+)
+from ...schedulers import KarrasDiffusionSchedulers
+from ...utils import (
+    is_accelerate_available,
+    is_accelerate_version,
+    logging,
+    randn_tensor,
+    replace_example_docstring,
+)
+from ..pipeline_utils import DiffusionPipeline
+from . import StableDiffusionXLPipelineOutput
+from .watermark import StableDiffusionXLWatermarker
+
+
+logger = logging.get_logger(__name__)  # pylint: disable=invalid-name
+
+EXAMPLE_DOC_STRING = """
+    Examples:
+        ```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]
+        ```
+"""
+
+
+def rescale_noise_cfg(noise_cfg, noise_pred_text, guidance_rescale=0.0):
+    """
+    Rescale `noise_cfg` according to `guidance_rescale`. Based on findings of [Common Diffusion Noise Schedules and
+    Sample Steps are Flawed](https://arxiv.org/pdf/2305.08891.pdf). See Section 3.4
+    """
+    std_text = noise_pred_text.std(dim=list(range(1, noise_pred_text.ndim)), keepdim=True)
+    std_cfg = noise_cfg.std(dim=list(range(1, noise_cfg.ndim)), keepdim=True)
+    # rescale the results from guidance (fixes overexposure)
+    noise_pred_rescaled = noise_cfg * (std_text / std_cfg)
+    # mix with the original results from guidance by factor guidance_rescale to avoid "plain looking" images
+    noise_cfg = guidance_rescale * noise_pred_rescaled + (1 - guidance_rescale) * noise_cfg
+    return noise_cfg
+
+
+class StableDiffusionXLImg2ImgPipeline(DiffusionPipeline):
+    r"""
+    Pipeline for text-to-image generation using Stable Diffusion.
+
+    This model inherits from [`DiffusionPipeline`]. Check the superclass documentation for the generic methods the
+    library implements for all the pipelines (such as downloading or saving, running on a particular device, etc.)
+
+    In addition the pipeline inherits the following loading methods:
+        - *Textual-Inversion*: [`loaders.TextualInversionLoaderMixin.load_textual_inversion`]
+        - *LoRA*: [`loaders.LoraLoaderMixin.load_lora_weights`]
+        - *Ckpt*: [`loaders.FromCkptMixin.from_ckpt`]
+
+    as well as the following saving methods:
+        - *LoRA*: [`loaders.LoraLoaderMixin.save_lora_weights`]
+
+    Args:
+        vae ([`AutoencoderKL`]):
+            Variational Auto-Encoder (VAE) Model to encode and decode images to and from latent representations.
+        text_encoder ([`CLIPTextModel`]):
+            Frozen text-encoder. Stable Diffusion uses the text portion of
+            [CLIP](https://huggingface.co/docs/transformers/model_doc/clip#transformers.CLIPTextModel), specifically
+            the [clip-vit-large-patch14](https://huggingface.co/openai/clip-vit-large-patch14) variant.
+        tokenizer (`CLIPTokenizer`):
+            Tokenizer of class
+            [CLIPTokenizer](https://huggingface.co/docs/transformers/v4.21.0/en/model_doc/clip#transformers.CLIPTokenizer).
+        unet ([`UNet2DConditionModel`]): Conditional U-Net architecture to denoise the encoded image latents.
+        scheduler ([`SchedulerMixin`]):
+            A scheduler to be used in combination with `unet` to denoise the encoded image latents. Can be one of
+            [`DDIMScheduler`], [`LMSDiscreteScheduler`], or [`PNDMScheduler`].
+    """
+    _optional_components = ["tokenizer", "text_encoder"]
+
+    def __init__(
+        self,
+        vae: AutoencoderKL,
+        text_encoder: CLIPTextModel,
+        text_encoder_2: CLIPTextModelWithProjection,
+        tokenizer: CLIPTokenizer,
+        tokenizer_2: CLIPTokenizer,
+        unet: UNet2DConditionModel,
+        scheduler: KarrasDiffusionSchedulers,
+        requires_aesthetics_score: bool = False,
+        force_zeros_for_empty_prompt: bool = True,
+    ):
+        super().__init__()
+
+        self.register_modules(
+            vae=vae,
+            text_encoder=text_encoder,
+            text_encoder_2=text_encoder_2,
+            tokenizer=tokenizer,
+            tokenizer_2=tokenizer_2,
+            unet=unet,
+            scheduler=scheduler,
+        )
+        self.register_to_config(force_zeros_for_empty_prompt=force_zeros_for_empty_prompt)
+        self.register_to_config(requires_aesthetics_score=requires_aesthetics_score)
+        self.vae_scale_factor = 2 ** (len(self.vae.config.block_out_channels) - 1)
+        self.vae_scale_factor = 8
+        self.image_processor = VaeImageProcessor(vae_scale_factor=self.vae_scale_factor)
+
+        self.watermark = StableDiffusionXLWatermarker()
+
+    def enable_vae_slicing(self):
+        r"""
+        Enable sliced VAE decoding.
+
+        When this option is enabled, the VAE will split the input tensor in slices to compute decoding in several
+        steps. This is useful to save some memory and allow larger batch sizes.
+        """
+        self.vae.enable_slicing()
+
+    def disable_vae_slicing(self):
+        r"""
+        Disable sliced VAE decoding. If `enable_vae_slicing` was previously invoked, this method will go back to
+        computing decoding in one step.
+        """
+        self.vae.disable_slicing()
+
+    def enable_vae_tiling(self):
+        r"""
+        Enable tiled VAE decoding.
+
+        When this option is enabled, the VAE will split the input tensor into tiles to compute decoding and encoding in
+        several steps. This is useful to save a large amount of memory and to allow the processing of larger images.
+        """
+        self.vae.enable_tiling()
+
+    def disable_vae_tiling(self):
+        r"""
+        Disable tiled VAE decoding. If `enable_vae_tiling` was previously invoked, this method will go back to
+        computing decoding in one step.
+        """
+        self.vae.disable_tiling()
+
+    def enable_sequential_cpu_offload(self, gpu_id=0):
+        r"""
+        Offloads all models to CPU using accelerate, significantly reducing memory usage. When called, unet,
+        text_encoder, vae and safety checker have their state dicts saved to CPU and then are moved to a
+        `torch.device('meta') and loaded to GPU only when their specific submodule has its `forward` method called.
+        Note that offloading happens on a submodule basis. Memory savings are higher than with
+        `enable_model_cpu_offload`, but performance is lower.
+        """
+        if is_accelerate_available() and is_accelerate_version(">=", "0.14.0"):
+            from accelerate import cpu_offload
+        else:
+            raise ImportError("`enable_sequential_cpu_offload` requires `accelerate v0.14.0` or higher")
+
+        device = torch.device(f"cuda:{gpu_id}")
+
+        if self.device.type != "cpu":
+            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]:
+            cpu_offload(cpu_offloaded_model, device)
+
+    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
+        to `enable_sequential_cpu_offload`, this method moves one whole model at a time to the GPU when its `forward`
+        method is called, and the model remains in GPU until the next model runs. Memory savings are lower than with
+        `enable_sequential_cpu_offload`, but performance is much better due to the iterative execution of the `unet`.
+        """
+        if is_accelerate_available() and is_accelerate_version(">=", "0.17.0.dev0"):
+            from accelerate import cpu_offload_with_hook
+        else:
+            raise ImportError("`enable_model_cpu_offload` requires `accelerate v0.17.0` or higher.")
+
+        device = torch.device(f"cuda:{gpu_id}")
+
+        if self.device.type != "cpu":
+            self.to("cpu", silence_dtype_warnings=True)
+            torch.cuda.empty_cache()  # otherwise we don't see the memory savings (but they probably exist)
+
+        model_sequence = (
+            [self.text_encoder, self.text_encoder_2] if self.text_encoder is not None else [self.text_encoder_2]
+        )
+        model_sequence.extend([self.unet, self.vae])
+
+        hook = None
+        for cpu_offloaded_model in model_sequence:
+            _, hook = cpu_offload_with_hook(cpu_offloaded_model, device, prev_module_hook=hook)
+
+        # We'll offload the last model manually.
+        self.final_offload_hook = hook
+
+    @property
+    def _execution_device(self):
+        r"""
+        Returns the device on which the pipeline's models will be executed. After calling
+        `pipeline.enable_sequential_cpu_offload()` the execution device can only be inferred from Accelerate's module
+        hooks.
+        """
+        if not hasattr(self.unet, "_hf_hook"):
+            return self.device
+        for module in self.unet.modules():
+            if (
+                hasattr(module, "_hf_hook")
+                and hasattr(module._hf_hook, "execution_device")
+                and module._hf_hook.execution_device is not None
+            ):
+                return torch.device(module._hf_hook.execution_device)
+        return self.device
+
+    def encode_prompt(
+        self,
+        prompt,
+        device,
+        num_images_per_prompt,
+        do_classifier_free_guidance,
+        negative_prompt=None,
+        prompt_embeds: Optional[torch.FloatTensor] = None,
+        negative_prompt_embeds: Optional[torch.FloatTensor] = None,
+        lora_scale: Optional[float] = None,
+    ):
+        r"""
+        Encodes the prompt into text encoder hidden states.
+
+        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
+            do_classifier_free_guidance (`bool`):
+                whether to use classifier free guidance or not
+            negative_prompt (`str` or `List[str]`, *optional*):
+                The prompt or prompts not to guide the image generation. If not defined, one has to pass
+                `negative_prompt_embeds` instead. Ignored when not using guidance (i.e., ignored if `guidance_scale` is
+                less than `1`).
+            prompt_embeds (`torch.FloatTensor`, *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.FloatTensor`, *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.
+            lora_scale (`float`, *optional*):
+                A lora scale that will be applied to all LoRA layers of the text encoder if LoRA layers are loaded.
+        """
+        # 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):
+            self._lora_scale = lora_scale
+
+        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]
+
+        # Define tokenizers and text encoders
+        tokenizers = [self.tokenizer, self.tokenizer_2] if self.tokenizer is not None else [self.tokenizer_2]
+        text_encoders = (
+            [self.text_encoder, self.text_encoder_2] if self.text_encoder is not None else [self.text_encoder_2]
+        )
+
+        if prompt_embeds is None:
+            # textual inversion: procecss multi-vector tokens if necessary
+            prompt_embeds_list = []
+            for tokenizer, text_encoder in zip(tokenizers, text_encoders):
+                if isinstance(self, TextualInversionLoaderMixin):
+                    prompt = self.maybe_convert_prompt(prompt, tokenizer)
+
+                text_inputs = tokenizer(
+                    prompt,
+                    padding="max_length",
+                    max_length=tokenizer.model_max_length,
+                    truncation=True,
+                    return_tensors="pt",
+                )
+                text_input_ids = text_inputs.input_ids
+                untruncated_ids = tokenizer(prompt, padding="longest", return_tensors="pt").input_ids
+
+                if untruncated_ids.shape[-1] >= text_input_ids.shape[-1] and not torch.equal(
+                    text_input_ids, untruncated_ids
+                ):
+                    removed_text = tokenizer.batch_decode(untruncated_ids[:, tokenizer.model_max_length - 1 : -1])
+                    logger.warning(
+                        "The following part of your input was truncated because CLIP can only handle sequences up to"
+                        f" {tokenizer.model_max_length} tokens: {removed_text}"
+                    )
+
+                prompt_embeds = text_encoder(
+                    text_input_ids.to(device),
+                    output_hidden_states=True,
+                )
+                # We are only ALWAYS interested in the pooled output of the final text encoder
+                pooled_prompt_embeds = prompt_embeds[0]
+
+                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)
+                prompt_embeds = prompt_embeds.view(bs_embed * num_images_per_prompt, seq_len, -1)
+
+                prompt_embeds_list.append(prompt_embeds)
+
+            prompt_embeds = torch.concat(prompt_embeds_list, dim=-1)
+
+        # get unconditional embeddings for classifier free guidance
+        zero_out_negative_prompt = negative_prompt is None and self.config.force_zeros_for_empty_prompt
+        if do_classifier_free_guidance and negative_prompt_embeds is None and zero_out_negative_prompt:
+            negative_prompt_embeds = torch.zeros_like(prompt_embeds)
+            negative_pooled_prompt_embeds = torch.zeros_like(pooled_prompt_embeds)
+        elif do_classifier_free_guidance and negative_prompt_embeds is None:
+            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):
+                raise TypeError(
+                    f"`negative_prompt` should be the same type to `prompt`, but got {type(negative_prompt)} !="
+                    f" {type(prompt)}."
+                )
+            elif isinstance(negative_prompt, str):
+                uncond_tokens = [negative_prompt]
+            elif batch_size != len(negative_prompt):
+                raise ValueError(
+                    f"`negative_prompt`: {negative_prompt} has batch size {len(negative_prompt)}, but `prompt`:"
+                    f" {prompt} has batch size {batch_size}. Please make sure that passed `negative_prompt` matches"
+                    " the batch size of `prompt`."
+                )
+            else:
+                uncond_tokens = negative_prompt
+
+            negative_prompt_embeds_list = []
+            for tokenizer, text_encoder in zip(tokenizers, text_encoders):
+                # textual inversion: procecss multi-vector tokens if necessary
+                if isinstance(self, TextualInversionLoaderMixin):
+                    uncond_tokens = self.maybe_convert_prompt(uncond_tokens, tokenizer)
+
+                max_length = prompt_embeds.shape[1]
+                uncond_input = tokenizer(
+                    uncond_tokens,
+                    padding="max_length",
+                    max_length=max_length,
+                    truncation=True,
+                    return_tensors="pt",
+                )
+
+                negative_prompt_embeds = text_encoder(
+                    uncond_input.input_ids.to(device),
+                    output_hidden_states=True,
+                )
+                # We are only ALWAYS interested in the pooled output of the final text encoder
+                negative_pooled_prompt_embeds = negative_prompt_embeds[0]
+
+                negative_prompt_embeds = negative_prompt_embeds.hidden_states[-2]
+
+                if do_classifier_free_guidance:
+                    # duplicate unconditional embeddings for each generation per prompt, using mps friendly method
+                    seq_len = negative_prompt_embeds.shape[1]
+
+                    negative_prompt_embeds = negative_prompt_embeds.to(dtype=text_encoder.dtype, device=device)
+
+                    negative_prompt_embeds = negative_prompt_embeds.repeat(1, num_images_per_prompt, 1)
+                    negative_prompt_embeds = negative_prompt_embeds.view(
+                        batch_size * num_images_per_prompt, seq_len, -1
+                    )
+
+                    # For classifier free guidance, we need to do two forward passes.
+                    # Here we concatenate the unconditional and text embeddings into a single batch
+                    # to avoid doing two forward passes
+
+                negative_prompt_embeds_list.append(negative_prompt_embeds)
+
+            negative_prompt_embeds = torch.concat(negative_prompt_embeds_list, dim=-1)
+
+        pooled_prompt_embeds = pooled_prompt_embeds.repeat(1, num_images_per_prompt).view(
+            bs_embed * num_images_per_prompt, -1
+        )
+        negative_pooled_prompt_embeds = negative_pooled_prompt_embeds.repeat(1, num_images_per_prompt).view(
+            bs_embed * num_images_per_prompt, -1
+        )
+
+        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
+
+    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.
+        # eta corresponds to η in DDIM paper: https://arxiv.org/abs/2010.02502
+        # and should be between [0, 1]
+
+        accepts_eta = "eta" in set(inspect.signature(self.scheduler.step).parameters.keys())
+        extra_step_kwargs = {}
+        if accepts_eta:
+            extra_step_kwargs["eta"] = eta
+
+        # check if the scheduler accepts generator
+        accepts_generator = "generator" in set(inspect.signature(self.scheduler.step).parameters.keys())
+        if accepts_generator:
+            extra_step_kwargs["generator"] = generator
+        return extra_step_kwargs
+
+    def check_inputs(
+        self, prompt, strength, callback_steps, negative_prompt=None, prompt_embeds=None, negative_prompt_embeds=None
+    ):
+        if strength < 0 or strength > 1:
+            raise ValueError(f"The value of strength should in [0.0, 1.0] but is {strength}")
+
+        if (callback_steps is None) or (
+            callback_steps is not None and (not isinstance(callback_steps, int) or callback_steps <= 0)
+        ):
+            raise ValueError(
+                f"`callback_steps` has to be a positive integer but is {callback_steps} of type"
+                f" {type(callback_steps)}."
+            )
+
+        if prompt is not None and prompt_embeds is not None:
+            raise ValueError(
+                f"Cannot forward both `prompt`: {prompt} and `prompt_embeds`: {prompt_embeds}. Please make sure to"
+                " only forward one of the two."
+            )
+        elif prompt is None and prompt_embeds is None:
+            raise ValueError(
+                "Provide either `prompt` or `prompt_embeds`. Cannot leave both `prompt` and `prompt_embeds` undefined."
+            )
+        elif prompt is not None and (not isinstance(prompt, str) and not isinstance(prompt, list)):
+            raise ValueError(f"`prompt` has to be of type `str` or `list` but is {type(prompt)}")
+
+        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 negative_prompt_embeds is not None:
+            if prompt_embeds.shape != negative_prompt_embeds.shape:
+                raise ValueError(
+                    "`prompt_embeds` and `negative_prompt_embeds` must have the same shape when passed directly, but"
+                    f" got: `prompt_embeds` {prompt_embeds.shape} != `negative_prompt_embeds`"
+                    f" {negative_prompt_embeds.shape}."
+                )
+
+    def get_timesteps(self, num_inference_steps, strength, device):
+        # get the original timestep using init_timestep
+        init_timestep = min(int(num_inference_steps * strength), num_inference_steps)
+
+        t_start = max(num_inference_steps - init_timestep, 0)
+        timesteps = self.scheduler.timesteps[t_start * self.scheduler.order :]
+
+        return timesteps, num_inference_steps - t_start
+
+    def prepare_latents(self, image, timestep, batch_size, num_images_per_prompt, dtype, device, generator=None):
+        if not isinstance(image, (torch.Tensor, PIL.Image.Image, list)):
+            raise ValueError(
+                f"`image` has to be of type `torch.Tensor`, `PIL.Image.Image` or list but is {type(image)}"
+            )
+
+        # Offload text encoder if `enable_model_cpu_offload` was enabled
+        if hasattr(self, "final_offload_hook") and self.final_offload_hook is not None:
+            self.text_encoder_2.to("cpu")
+            torch.cuda.empty_cache()
+
+        image = image.to(device=device, dtype=dtype)
+
+        batch_size = batch_size * num_images_per_prompt
+
+        if image.shape[1] == 4:
+            init_latents = image
+
+        else:
+            # make sure the VAE is in float32 mode, as it overflows in float16
+            image = image.float()
+            self.vae.to(dtype=torch.float32)
+
+            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."
+                )
+
+            elif isinstance(generator, list):
+                init_latents = [
+                    self.vae.encode(image[i : i + 1]).latent_dist.sample(generator[i]) for i in range(batch_size)
+                ]
+                init_latents = torch.cat(init_latents, dim=0)
+            else:
+                init_latents = self.vae.encode(image).latent_dist.sample(generator)
+
+            self.vae.to(dtype)
+            init_latents = init_latents.to(dtype)
+
+            init_latents = self.vae.config.scaling_factor * init_latents
+
+        if batch_size > init_latents.shape[0] and batch_size % init_latents.shape[0] == 0:
+            # expand init_latents for batch_size
+            additional_image_per_prompt = batch_size // init_latents.shape[0]
+            init_latents = torch.cat([init_latents] * additional_image_per_prompt, dim=0)
+        elif batch_size > init_latents.shape[0] and batch_size % init_latents.shape[0] != 0:
+            raise ValueError(
+                f"Cannot duplicate `image` of batch size {init_latents.shape[0]} to {batch_size} text prompts."
+            )
+        else:
+            init_latents = torch.cat([init_latents], dim=0)
+
+        shape = init_latents.shape
+        noise = randn_tensor(shape, generator=generator, device=device, dtype=dtype)
+
+        # get latents
+        init_latents = self.scheduler.add_noise(init_latents, noise, timestep)
+        latents = init_latents
+
+        return latents
+
+    def _get_add_time_ids(
+        self, original_size, crops_coords_top_left, target_size, aesthetic_score, negative_aesthetic_score, dtype
+    ):
+        if self.config.requires_aesthetics_score:
+            add_time_ids = list(original_size + crops_coords_top_left + (aesthetic_score,))
+            add_neg_time_ids = list(original_size + crops_coords_top_left + (negative_aesthetic_score,))
+        else:
+            add_time_ids = list(original_size + crops_coords_top_left + target_size)
+            add_neg_time_ids = list(original_size + crops_coords_top_left + target_size)
+
+        passed_add_embed_dim = (
+            self.unet.config.addition_time_embed_dim * len(add_time_ids) + self.text_encoder_2.config.projection_dim
+        )
+        expected_add_embed_dim = self.unet.add_embedding.linear_1.in_features
+
+        if (
+            expected_add_embed_dim > passed_add_embed_dim
+            and (expected_add_embed_dim - passed_add_embed_dim) == self.unet.config.addition_time_embed_dim
+        ):
+            raise ValueError(
+                f"Model expects an added time embedding vector of length {expected_add_embed_dim}, but a vector of {passed_add_embed_dim} was created. Please make sure to enable `requires_aesthetics_score` with `pipe.register_to_config(requires_aesthetics_score=True)` to make sure `aesthetic_score` {aesthetic_score} and `negative_aesthetic_score` {negative_aesthetic_score} is correctly used by the model."
+            )
+        elif (
+            expected_add_embed_dim < passed_add_embed_dim
+            and (passed_add_embed_dim - expected_add_embed_dim) == self.unet.config.addition_time_embed_dim
+        ):
+            raise ValueError(
+                f"Model expects an added time embedding vector of length {expected_add_embed_dim}, but a vector of {passed_add_embed_dim} was created. Please make sure to disable `requires_aesthetics_score` with `pipe.register_to_config(requires_aesthetics_score=False)` to make sure `target_size` {target_size} is correctly used by the model."
+            )
+        elif expected_add_embed_dim != passed_add_embed_dim:
+            raise ValueError(
+                f"Model expects an added time embedding vector of length {expected_add_embed_dim}, but a vector of {passed_add_embed_dim} was created. The model has an incorrect config. Please check `unet.config.time_embedding_type` and `text_encoder_2.config.projection_dim`."
+            )
+
+        add_time_ids = torch.tensor([add_time_ids], dtype=dtype)
+        add_neg_time_ids = torch.tensor([add_neg_time_ids], dtype=dtype)
+
+        return add_time_ids, add_neg_time_ids
+
+    @torch.no_grad()
+    @replace_example_docstring(EXAMPLE_DOC_STRING)
+    def __call__(
+        self,
+        prompt: Union[str, List[str]] = None,
+        image: Union[
+            torch.FloatTensor,
+            PIL.Image.Image,
+            np.ndarray,
+            List[torch.FloatTensor],
+            List[PIL.Image.Image],
+            List[np.ndarray],
+        ] = None,
+        strength: float = 0.3,
+        num_inference_steps: int = 50,
+        guidance_scale: float = 5.0,
+        negative_prompt: Optional[Union[str, List[str]]] = None,
+        num_images_per_prompt: Optional[int] = 1,
+        eta: float = 0.0,
+        generator: Optional[Union[torch.Generator, List[torch.Generator]]] = None,
+        latents: Optional[torch.FloatTensor] = None,
+        prompt_embeds: Optional[torch.FloatTensor] = None,
+        negative_prompt_embeds: Optional[torch.FloatTensor] = None,
+        output_type: Optional[str] = "pil",
+        return_dict: bool = True,
+        callback: Optional[Callable[[int, int, torch.FloatTensor], None]] = None,
+        callback_steps: int = 1,
+        cross_attention_kwargs: Optional[Dict[str, Any]] = None,
+        guidance_rescale: float = 0.0,
+        original_size: Tuple[int, int] = (1024, 1024),
+        crops_coords_top_left: Tuple[int, int] = (0, 0),
+        target_size: Tuple[int, int] = (1024, 1024),
+        aesthetic_score: float = 6.0,
+        negative_aesthetic_score: float = 2.5,
+    ):
+        r"""
+        Function invoked when calling the pipeline for generation.
+
+        Args:
+            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.
+            image (`torch.FloatTensor` or `PIL.Image.Image` or `np.ndarray` or `List[torch.FloatTensor]` or `List[PIL.Image.Image]` or `List[np.ndarray]`):
+                The image(s) to modify with the pipeline.
+            strength (`float`, *optional*, defaults to 0.8):
+                Conceptually, indicates how much to transform the reference `image`. Must be between 0 and 1. `image`
+                will be used as a starting point, adding more noise to it the larger the `strength`. The number of
+                denoising steps depends on the amount of noise initially added. When `strength` is 1, added noise will
+                be maximum and the denoising process will run for the full number of iterations specified in
+                `num_inference_steps`. A value of 1, therefore, essentially ignores `image`.
+            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.
+            guidance_scale (`float`, *optional*, defaults to 7.5):
+                Guidance scale as defined in [Classifier-Free Diffusion Guidance](https://arxiv.org/abs/2207.12598).
+                `guidance_scale` is defined as `w` of equation 2. of [Imagen
+                Paper](https://arxiv.org/pdf/2205.11487.pdf). 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.
+            negative_prompt (`str` or `List[str]`, *optional*):
+                The prompt or prompts not to guide the image generation. If not defined, one has to pass
+                `negative_prompt_embeds` instead. Ignored when not using guidance (i.e., ignored if `guidance_scale` is
+                less than `1`).
+            num_images_per_prompt (`int`, *optional*, defaults to 1):
+                The number of images to generate per prompt.
+            eta (`float`, *optional*, defaults to 0.0):
+                Corresponds to parameter eta (η) in the DDIM paper: https://arxiv.org/abs/2010.02502. Only applies to
+                [`schedulers.DDIMScheduler`], will be ignored for others.
+            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.FloatTensor`, *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 ge generated by sampling using the supplied random `generator`.
+            prompt_embeds (`torch.FloatTensor`, *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.FloatTensor`, *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.stable_diffusion.StableDiffusionXLPipelineOutput`] instead of a
+                plain tuple.
+            callback (`Callable`, *optional*):
+                A function that will be called every `callback_steps` steps during inference. The function will be
+                called with the following arguments: `callback(step: int, timestep: int, latents: torch.FloatTensor)`.
+            callback_steps (`int`, *optional*, defaults to 1):
+                The frequency at which the `callback` function will be called. If not specified, the callback will be
+                called at every step.
+            cross_attention_kwargs (`dict`, *optional*):
+                A kwargs dictionary that if specified is passed along to the `AttentionProcessor` as defined under
+                `self.processor` in
+                [diffusers.cross_attention](https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/cross_attention.py).
+            guidance_rescale (`float`, *optional*, defaults to 0.7):
+                Guidance rescale factor proposed by [Common Diffusion Noise Schedules and Sample Steps are
+                Flawed](https://arxiv.org/pdf/2305.08891.pdf) `guidance_scale` is defined as `φ` in equation 16. of
+                [Common Diffusion Noise Schedules and Sample Steps are Flawed](https://arxiv.org/pdf/2305.08891.pdf).
+                Guidance rescale factor should fix overexposure when using zero terminal SNR.
+            original_size (`Tuple[int]`, *optional*, defaults to (1024, 1024)):
+                TODO
+            crops_coords_top_left (`Tuple[int]`, *optional*, defaults to (0, 0)):
+                TODO
+            target_size (`Tuple[int]`, *optional*, defaults to (1024, 1024)):
+                TODO
+            aesthetic_score (`float`, *optional*, defaults to 6.0):
+                TODO
+            negative_aesthetic_score (`float`, *optional*, defaults to 2.5):
+                TDOO
+
+        Examples:
+
+        Returns:
+            [`~pipelines.stable_diffusion.StableDiffusionXLPipelineOutput`] or `tuple`:
+            [`~pipelines.stable_diffusion.StableDiffusionXLPipelineOutput`] if `return_dict` is True, otherwise a
+            `tuple. When returning a tuple, the first element is a list with the generated images, and the second
+            element is a list of `bool`s denoting whether the corresponding generated image likely represents
+            "not-safe-for-work" (nsfw) content, according to the `safety_checker`.
+        """
+        # 1. Check inputs. Raise error if not correct
+        self.check_inputs(prompt, strength, callback_steps, negative_prompt, prompt_embeds, negative_prompt_embeds)
+
+        # 2. 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]
+
+        device = self._execution_device
+
+        # here `guidance_scale` is defined analog to the guidance weight `w` of equation (2)
+        # of the Imagen paper: https://arxiv.org/pdf/2205.11487.pdf . `guidance_scale = 1`
+        # corresponds to doing no classifier free guidance.
+        do_classifier_free_guidance = guidance_scale > 1.0
+
+        # 3. Encode input prompt
+        text_encoder_lora_scale = (
+            cross_attention_kwargs.get("scale", None) if cross_attention_kwargs is not None else None
+        )
+        (
+            prompt_embeds,
+            negative_prompt_embeds,
+            pooled_prompt_embeds,
+            negative_pooled_prompt_embeds,
+        ) = self.encode_prompt(
+            prompt,
+            device,
+            num_images_per_prompt,
+            do_classifier_free_guidance,
+            negative_prompt,
+            prompt_embeds=prompt_embeds,
+            negative_prompt_embeds=negative_prompt_embeds,
+            lora_scale=text_encoder_lora_scale,
+        )
+
+        # 4. Preprocess image
+        image = self.image_processor.preprocess(image)
+
+        # 5. Prepare timesteps
+        self.scheduler.set_timesteps(num_inference_steps, device=device)
+        timesteps, num_inference_steps = self.get_timesteps(num_inference_steps, strength, device)
+        latent_timestep = timesteps[:1].repeat(batch_size * num_images_per_prompt)
+
+        # 6. Prepare latent variables
+        latents = self.prepare_latents(
+            image, latent_timestep, batch_size, num_images_per_prompt, prompt_embeds.dtype, device, generator
+        )
+        # 7. Prepare extra step kwargs.
+        extra_step_kwargs = self.prepare_extra_step_kwargs(generator, eta)
+
+        # 8. Prepare added time ids & embeddings
+        add_text_embeds = pooled_prompt_embeds
+        add_time_ids, add_neg_time_ids = self._get_add_time_ids(
+            original_size,
+            crops_coords_top_left,
+            target_size,
+            aesthetic_score,
+            negative_aesthetic_score,
+            dtype=prompt_embeds.dtype,
+        )
+
+        if do_classifier_free_guidance:
+            prompt_embeds = torch.cat([negative_prompt_embeds, prompt_embeds], dim=0)
+            add_text_embeds = torch.cat([negative_pooled_prompt_embeds, add_text_embeds], dim=0)
+            add_time_ids = torch.cat([add_neg_time_ids, add_time_ids], dim=0)
+
+        prompt_embeds = prompt_embeds.to(device)
+        add_text_embeds = add_text_embeds.to(device)
+        add_time_ids = add_time_ids.to(device).repeat(batch_size * num_images_per_prompt, 1)
+
+        # 9. Denoising loop
+        num_warmup_steps = len(timesteps) - num_inference_steps * self.scheduler.order
+        with self.progress_bar(total=num_inference_steps) as progress_bar:
+            for i, t in enumerate(timesteps):
+                # expand the latents if we are doing classifier free guidance
+                latent_model_input = torch.cat([latents] * 2) if do_classifier_free_guidance else latents
+
+                latent_model_input = self.scheduler.scale_model_input(latent_model_input, t)
+
+                # predict the noise residual
+                added_cond_kwargs = {"text_embeds": add_text_embeds, "time_ids": add_time_ids}
+                noise_pred = self.unet(
+                    latent_model_input,
+                    t,
+                    encoder_hidden_states=prompt_embeds,
+                    cross_attention_kwargs=cross_attention_kwargs,
+                    added_cond_kwargs=added_cond_kwargs,
+                    return_dict=False,
+                )[0]
+
+                # perform guidance
+                if do_classifier_free_guidance:
+                    noise_pred_uncond, noise_pred_text = noise_pred.chunk(2)
+                    noise_pred = noise_pred_uncond + guidance_scale * (noise_pred_text - noise_pred_uncond)
+
+                if do_classifier_free_guidance and guidance_rescale > 0.0:
+                    # Based on 3.4. in https://arxiv.org/pdf/2305.08891.pdf
+                    noise_pred = rescale_noise_cfg(noise_pred, noise_pred_text, guidance_rescale=guidance_rescale)
+
+                # compute the previous noisy sample x_t -> x_t-1
+                latents = self.scheduler.step(noise_pred, t, latents, **extra_step_kwargs, return_dict=False)[0]
+
+                # 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 callback is not None and i % callback_steps == 0:
+                        callback(i, t, latents)
+
+        # 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,
+        ]
+        # 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:
+            self.vae.post_quant_conv.to(latents.dtype)
+            self.vae.decoder.conv_in.to(latents.dtype)
+            self.vae.decoder.mid_block.to(latents.dtype)
+        else:
+            latents = latents.float()
+
+        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]
+
+        image = self.watermark.apply_watermark(image)
+        image = self.image_processor.postprocess(image, output_type=output_type, do_denormalize=do_denormalize)
+
+        # 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 StableDiffusionXLPipelineOutput(images=image, nsfw_content_detected=has_nsfw_concept)

src/diffusers/pipelines/stable_diffusion_xl/watermark.py

+import numpy as np
+import torch
+from imwatermark import WatermarkEncoder
+
+
+# Copied from https://github.com/Stability-AI/generative-models/blob/613af104c6b85184091d42d374fef420eddb356d/scripts/demo/streamlit_helpers.py#L66
+WATERMARK_MESSAGE = 0b101100111110110010010000011110111011000110011110
+# bin(x)[2:] gives bits of x as str, use int to convert them to 0/1
+WATERMARK_BITS = [int(bit) for bit in bin(WATERMARK_MESSAGE)[2:]]
+
+
+class StableDiffusionXLWatermarker:
+    def __init__(self):
+        self.watermark = WATERMARK_BITS
+        self.encoder = WatermarkEncoder()
+
+        self.encoder.set_watermark("bits", self.watermark)
+
+    def apply_watermark(self, images: torch.FloatTensor):
+        # can't encode images that are smaller than 256
+        if images.shape[-1] < 256:
+            return images
+
+        images = (255 * (images / 2 + 0.5)).cpu().permute(0, 2, 3, 1).float().numpy()
+
+        images = [self.encoder.encode(image, "dwtDct") for image in images]
+
+        images = torch.from_numpy(np.array(images)).permute(0, 3, 1, 2)
+
+        images = torch.clamp(2 * (images / 255 - 0.5), min=-1.0, max=1.0)
+        return images