[SD-XL] Add inpainting (#4098)

* Add more

* more

* up

* Get ensemble of expert denoisers working

* Fix code

* add tests

* up

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

@@ -57,6 +57,50 @@ prompt = "Astronaut in a jungle, cold color palette, muted colors, detailed, 8k"
 image = pipe(prompt=prompt).images[0]
 ```
 
+### Image-to-image 
+
+You can use SDXL as follows for *image-to-image*:
+
+```py 
+import torch
+from diffusers import StableDiffusionXLImg2ImgPipeline
+from diffusers.utils import load_image
+
+pipe = StableDiffusionXLImg2ImgPipeline.from_pretrained(
+    "stabilityai/stable-diffusion-xl-refiner-0.9", torch_dtype=torch.float16, variant="fp16", use_safetensors=True
+)
+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]
+```
+
+### Inpainting
+
+You can use SDXL as follows for *inpainting*
+
+```py 
+import torch
+from diffusers import StableDiffusionXLInpaintPipeline
+from diffusers.utils import load_image
+
+pipe = StableDiffusionXLInpaintPipeline.from_pretrained(
+    "stabilityai/stable-diffusion-xl-base-0.9", torch_dtype=torch.float16, variant="fp16", use_safetensors=True
+)
+pipe.to("cuda")
+
+img_url = "https://raw.githubusercontent.com/CompVis/latent-diffusion/main/data/inpainting_examples/overture-creations-5sI6fQgYIuo.png"
+mask_url = "https://raw.githubusercontent.com/CompVis/latent-diffusion/main/data/inpainting_examples/overture-creations-5sI6fQgYIuo_mask.png"
+
+init_image = load_image(img_url).convert("RGB")
+mask_image = load_image(mask_url).convert("RGB")
+
+prompt = "A majestic tiger sitting on a bench"
+image = pipe(prompt=prompt, image=init_image, mask_image=mask_image, num_inference_steps=50, strength=0.80).images[0]
+```
+
 ### Refining the image output
 
 In addition to the [base model checkpoint](https://huggingface.co/stabilityai/stable-diffusion-xl-base-0.9), 
@@ -183,24 +227,65 @@ image = refiner(prompt=prompt, image=image[None, :]).images[0]
 |---|---|
 | ![](https://huggingface.co/datasets/diffusers/docs-images/resolve/main/sd_xl/init_image.png) | ![](https://huggingface.co/datasets/diffusers/docs-images/resolve/main/sd_xl/refined_image.png) |
 
-### Image-to-image 
+<Tip>
 
-```py 
-import torch
-from diffusers import StableDiffusionXLImg2ImgPipeline
+The refiner can also very well be used in an in-painting setting. To do so just make
+  sure you use the [`StableDiffusionXLInpaintPipeline`] classes as shown below
+
+</Tip>
+
+To use the refiner for inpainting in the Ensemble of Expert Denoisers setting you can do the following:
+
+```py
+from diffusers import StableDiffusionXLInpaintPipeline
 from diffusers.utils import load_image
 
-pipe = StableDiffusionXLImg2ImgPipeline.from_pretrained(
-    "stabilityai/stable-diffusion-xl-refiner-0.9", torch_dtype=torch.float16
+pipe = StableDiffusionXLInpaintPipeline.from_pretrained(
+    "stabilityai/stable-diffusion-xl-base-0.9", torch_dtype=torch.float16, variant="fp16", use_safetensors=True
 )
-pipe = pipe.to("cuda")
-url = "https://huggingface.co/datasets/patrickvonplaten/images/resolve/main/aa_xl/000000009.png"
+pipe.to("cuda")
 
-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]
+refiner = StableDiffusionXLInpaintPipeline.from_pretrained(
+    "stabilityai/stable-diffusion-xl-refiner-0.9",
+    text_encoder_2=pipe.text_encoder_2,
+    vae=pipe.vae,
+    torch_dtype=torch.float16,
+    use_safetensors=True,
+    variant="fp16",
+)
+refiner.to("cuda")
+
+img_url = "https://raw.githubusercontent.com/CompVis/latent-diffusion/main/data/inpainting_examples/overture-creations-5sI6fQgYIuo.png"
+mask_url = "https://raw.githubusercontent.com/CompVis/latent-diffusion/main/data/inpainting_examples/overture-creations-5sI6fQgYIuo_mask.png"
+
+init_image = load_image(img_url).convert("RGB")
+mask_image = load_image(mask_url).convert("RGB")
+
+prompt = "A majestic tiger sitting on a bench"
+num_inference_steps = 75
+high_noise_frac = 0.7
+
+image = pipe(
+    prompt=prompt,
+    image=init_image,
+    mask_image=mask_image,
+    num_inference_steps=num_inference_steps,
+    strength=0.80,
+    denoising_start=high_noise_frac,
+    output_type="latent",
+).images
+image = refiner(
+    prompt=prompt,
+    image=image,
+    mask_image=mask_image,
+    num_inference_steps=num_inference_steps,
+    denoising_start=high_noise_frac,
+).images[0]
 ```
 
+To use the refiner for inpainting in the standard SDE-style setting, simply remove `denoising_end` and `denoising_start` and choose a smaller
+number of inference steps for the refiner.
+
 ### Loading single file checkpoints / original file format
 
 By making use of [`~diffusers.loaders.FromSingleFileMixin.from_single_file`] you can also load the 
@@ -271,3 +356,9 @@ pip install xformers
 [[autodoc]] StableDiffusionXLImg2ImgPipeline
 	- all
 	- __call__
+
+## StableDiffusionXLInpaintPipeline
+
+[[autodoc]] StableDiffusionXLInpaintPipeline
+	- all
+	- __call__

src/diffusers/__init__.py

@@ -195,7 +195,11 @@ try:
 except OptionalDependencyNotAvailable:
     from .utils.dummy_torch_and_transformers_and_invisible_watermark_objects import *  # noqa F403
 else:
-    from .pipelines import StableDiffusionXLImg2ImgPipeline, StableDiffusionXLPipeline
+    from .pipelines import (
+        StableDiffusionXLImg2ImgPipeline,
+        StableDiffusionXLInpaintPipeline,
+        StableDiffusionXLPipeline,
+    )
 
 try:
     if not (is_torch_available() and is_transformers_available() and is_k_diffusion_available()):

src/diffusers/pipelines/__init__.py

@@ -119,7 +119,11 @@ try:
 except OptionalDependencyNotAvailable:
     from ..utils.dummy_torch_and_transformers_and_invisible_watermark_objects import *  # noqa F403
 else:
-    from .stable_diffusion_xl import StableDiffusionXLImg2ImgPipeline, StableDiffusionXLPipeline
+    from .stable_diffusion_xl import (
+        StableDiffusionXLImg2ImgPipeline,
+        StableDiffusionXLInpaintPipeline,
+        StableDiffusionXLPipeline,
+    )
 
 try:
     if not is_onnx_available():

src/diffusers/pipelines/stable_diffusion/pipeline_stable_diffusion_inpaint.py

@@ -981,8 +981,6 @@ class StableDiffusionInpaintPipeline(
             generator,
             do_classifier_free_guidance,
         )
-        init_image = init_image.to(device=device, dtype=masked_image_latents.dtype)
-        init_image = self._encode_vae_image(init_image, generator=generator)
 
         # 8. Check that sizes of mask, masked image and latents match
         if num_channels_unet == 9:

src/diffusers/pipelines/stable_diffusion_xl/__init__.py

@@ -24,3 +24,4 @@ class StableDiffusionXLPipelineOutput(BaseOutput):
 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
+    from .pipeline_stable_diffusion_xl_inpaint import StableDiffusionXLInpaintPipeline

src/diffusers/pipelines/stable_diffusion_xl/pipeline_stable_diffusion_xl.py

@@ -59,6 +59,7 @@ EXAMPLE_DOC_STRING = """
 """
 
 
+# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.rescale_noise_cfg
 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
@@ -75,7 +76,7 @@ def rescale_noise_cfg(noise_cfg, noise_pred_text, guidance_rescale=0.0):
 
 class StableDiffusionXLPipeline(DiffusionPipeline, FromSingleFileMixin, LoraLoaderMixin):
     r"""
-    Pipeline for text-to-image generation using Stable Diffusion.
+    Pipeline for text-to-image generation using Stable Diffusion XL.
 
     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.)
@@ -92,12 +93,21 @@ class StableDiffusionXLPipeline(DiffusionPipeline, FromSingleFileMixin, LoraLoad
         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
+            Frozen text-encoder. Stable Diffusion XL 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.
+        text_encoder_2 ([` CLIPTextModelWithProjection`]):
+            Second frozen text-encoder. Stable Diffusion XL uses the text and pool portion of
+            [CLIP](https://huggingface.co/docs/transformers/model_doc/clip#transformers.CLIPTextModelWithProjection),
+            specifically the
+            [laion/CLIP-ViT-bigG-14-laion2B-39B-b160k](https://huggingface.co/laion/CLIP-ViT-bigG-14-laion2B-39B-b160k)
+            variant.
         tokenizer (`CLIPTokenizer`):
             Tokenizer of class
             [CLIPTokenizer](https://huggingface.co/docs/transformers/v4.21.0/en/model_doc/clip#transformers.CLIPTokenizer).
+        tokenizer_2 (`CLIPTokenizer`):
+            Second 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

src/diffusers/pipelines/stable_diffusion_xl/pipeline_stable_diffusion_xl_img2img.py

@@ -64,6 +64,7 @@ EXAMPLE_DOC_STRING = """
 """
 
 
+# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.rescale_noise_cfg
 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
@@ -80,7 +81,7 @@ def rescale_noise_cfg(noise_cfg, noise_pred_text, guidance_rescale=0.0):
 
 class StableDiffusionXLImg2ImgPipeline(DiffusionPipeline, FromSingleFileMixin, LoraLoaderMixin):
     r"""
-    Pipeline for text-to-image generation using Stable Diffusion.
+    Pipeline for text-to-image generation using Stable Diffusion XL.
 
     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.)
@@ -97,12 +98,21 @@ class StableDiffusionXLImg2ImgPipeline(DiffusionPipeline, FromSingleFileMixin, L
         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
+            Frozen text-encoder. Stable Diffusion XL 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.
+        text_encoder_2 ([` CLIPTextModelWithProjection`]):
+            Second frozen text-encoder. Stable Diffusion XL uses the text and pool portion of
+            [CLIP](https://huggingface.co/docs/transformers/model_doc/clip#transformers.CLIPTextModelWithProjection),
+            specifically the
+            [laion/CLIP-ViT-bigG-14-laion2B-39B-b160k](https://huggingface.co/laion/CLIP-ViT-bigG-14-laion2B-39B-b160k)
+            variant.
         tokenizer (`CLIPTokenizer`):
             Tokenizer of class
             [CLIPTokenizer](https://huggingface.co/docs/transformers/v4.21.0/en/model_doc/clip#transformers.CLIPTokenizer).
+        tokenizer_2 (`CLIPTokenizer`):
+            Second 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

src/diffusers/pipelines/stable_diffusion_xl/pipeline_stable_diffusion_xl_inpaint.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
+import torch
+from transformers import CLIPTextModel, CLIPTextModelWithProjection, CLIPTokenizer
+
+from ...image_processor import VaeImageProcessor
+from ...loaders import FromSingleFileMixin, 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
+from ..pipeline_utils import DiffusionPipeline
+from . import StableDiffusionXLPipelineOutput
+from .watermark import StableDiffusionXLWatermarker
+
+
+logger = logging.get_logger(__name__)  # pylint: disable=invalid-name
+
+
+# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.rescale_noise_cfg
+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
+
+
+def mask_pil_to_torch(mask, height, width):
+    # preprocess mask
+    if isinstance(mask, (PIL.Image.Image, np.ndarray)):
+        mask = [mask]
+
+    if isinstance(mask, list) and isinstance(mask[0], PIL.Image.Image):
+        mask = [i.resize((width, height), resample=PIL.Image.LANCZOS) for i in mask]
+        mask = np.concatenate([np.array(m.convert("L"))[None, None, :] for m in mask], axis=0)
+        mask = mask.astype(np.float32) / 255.0
+    elif isinstance(mask, list) and isinstance(mask[0], np.ndarray):
+        mask = np.concatenate([m[None, None, :] for m in mask], axis=0)
+
+    mask = torch.from_numpy(mask)
+    return mask
+
+
+def prepare_mask_and_masked_image(image, mask, height, width, return_image: bool = False):
+    """
+    Prepares a pair (image, mask) to be consumed by the Stable Diffusion pipeline. This means that those inputs will be
+    converted to ``torch.Tensor`` with shapes ``batch x channels x height x width`` where ``channels`` is ``3`` for the
+    ``image`` and ``1`` for the ``mask``.
+
+    The ``image`` will be converted to ``torch.float32`` and normalized to be in ``[-1, 1]``. The ``mask`` will be
+    binarized (``mask > 0.5``) and cast to ``torch.float32`` too.
+
+    Args:
+        image (Union[np.array, PIL.Image, torch.Tensor]): The image to inpaint.
+            It can be a ``PIL.Image``, or a ``height x width x 3`` ``np.array`` or a ``channels x height x width``
+            ``torch.Tensor`` or a ``batch x channels x height x width`` ``torch.Tensor``.
+        mask (_type_): The mask to apply to the image, i.e. regions to inpaint.
+            It can be a ``PIL.Image``, or a ``height x width`` ``np.array`` or a ``1 x height x width``
+            ``torch.Tensor`` or a ``batch x 1 x height x width`` ``torch.Tensor``.
+
+
+    Raises:
+        ValueError: ``torch.Tensor`` images should be in the ``[-1, 1]`` range. ValueError: ``torch.Tensor`` mask
+        should be in the ``[0, 1]`` range. ValueError: ``mask`` and ``image`` should have the same spatial dimensions.
+        TypeError: ``mask`` is a ``torch.Tensor`` but ``image`` is not
+            (ot the other way around).
+
+    Returns:
+        tuple[torch.Tensor]: The pair (mask, masked_image) as ``torch.Tensor`` with 4
+            dimensions: ``batch x channels x height x width``.
+    """
+
+    # checkpoint. TOD(Yiyi) - need to clean this up later
+    if image is None:
+        raise ValueError("`image` input cannot be undefined.")
+
+    if mask is None:
+        raise ValueError("`mask_image` input cannot be undefined.")
+
+    if isinstance(image, torch.Tensor):
+        if not isinstance(mask, torch.Tensor):
+            mask = mask_pil_to_torch(mask, height, width)
+
+        if image.ndim == 3:
+            image = image.unsqueeze(0)
+
+        # Batch and add channel dim for single mask
+        if mask.ndim == 2:
+            mask = mask.unsqueeze(0).unsqueeze(0)
+
+        # Batch single mask or add channel dim
+        if mask.ndim == 3:
+            # Single batched mask, no channel dim or single mask not batched but channel dim
+            if mask.shape[0] == 1:
+                mask = mask.unsqueeze(0)
+
+            # Batched masks no channel dim
+            else:
+                mask = mask.unsqueeze(1)
+
+        assert image.ndim == 4 and mask.ndim == 4, "Image and Mask must have 4 dimensions"
+        # assert image.shape[-2:] == mask.shape[-2:], "Image and Mask must have the same spatial dimensions"
+        assert image.shape[0] == mask.shape[0], "Image and Mask must have the same batch size"
+
+        # Check image is in [-1, 1]
+        # if image.min() < -1 or image.max() > 1:
+        #    raise ValueError("Image should be in [-1, 1] range")
+
+        # Check mask is in [0, 1]
+        if mask.min() < 0 or mask.max() > 1:
+            raise ValueError("Mask should be in [0, 1] range")
+
+        # Binarize mask
+        mask[mask < 0.5] = 0
+        mask[mask >= 0.5] = 1
+
+        # Image as float32
+        image = image.to(dtype=torch.float32)
+    elif isinstance(mask, torch.Tensor):
+        raise TypeError(f"`mask` is a torch.Tensor but `image` (type: {type(image)} is not")
+    else:
+        # preprocess image
+        if isinstance(image, (PIL.Image.Image, np.ndarray)):
+            image = [image]
+        if isinstance(image, list) and isinstance(image[0], PIL.Image.Image):
+            # resize all images w.r.t passed height an width
+            image = [i.resize((width, height), resample=PIL.Image.LANCZOS) for i in image]
+            image = [np.array(i.convert("RGB"))[None, :] for i in image]
+            image = np.concatenate(image, axis=0)
+        elif isinstance(image, list) and isinstance(image[0], np.ndarray):
+            image = np.concatenate([i[None, :] for i in image], axis=0)
+
+        image = image.transpose(0, 3, 1, 2)
+        image = torch.from_numpy(image).to(dtype=torch.float32) / 127.5 - 1.0
+
+        mask = mask_pil_to_torch(mask, height, width)
+        mask[mask < 0.5] = 0
+        mask[mask >= 0.5] = 1
+
+    if image.shape[1] == 4:
+        # images are in latent space and thus can't
+        # be masked set masked_image to None
+        # we assume that the checkpoint is not an inpainting
+        # checkpoint. TOD(Yiyi) - need to clean this up later
+        masked_image = None
+    else:
+        masked_image = image * (mask < 0.5)
+
+    # n.b. ensure backwards compatibility as old function does not return image
+    if return_image:
+        return mask, masked_image, image
+
+    return mask, masked_image
+
+
+class StableDiffusionXLInpaintPipeline(
+    DiffusionPipeline, TextualInversionLoaderMixin, LoraLoaderMixin, FromSingleFileMixin
+):
+    r"""
+    Pipeline for text-to-image generation using Stable Diffusion XL.
+
+    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.FromSingleFileMixin.from_single_file`]
+
+    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 XL 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.
+        text_encoder_2 ([` CLIPTextModelWithProjection`]):
+            Second frozen text-encoder. Stable Diffusion XL uses the text and pool portion of
+            [CLIP](https://huggingface.co/docs/transformers/model_doc/clip#transformers.CLIPTextModelWithProjection),
+            specifically the
+            [laion/CLIP-ViT-bigG-14-laion2B-39B-b160k](https://huggingface.co/laion/CLIP-ViT-bigG-14-laion2B-39B-b160k)
+            variant.
+        tokenizer (`CLIPTokenizer`):
+            Tokenizer of class
+            [CLIPTokenizer](https://huggingface.co/docs/transformers/v4.21.0/en/model_doc/clip#transformers.CLIPTokenizer).
+        tokenizer_2 (`CLIPTokenizer`):
+            Second 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.image_processor = VaeImageProcessor(vae_scale_factor=self.vae_scale_factor)
+
+        self.watermark = StableDiffusionXLWatermarker()
+
+    # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.enable_vae_slicing
+    def enable_vae_slicing(self):
+        r"""
+        Enable sliced VAE decoding.
+
+        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()
+
+    # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.disable_vae_slicing
+    def disable_vae_slicing(self):
+        r"""
+        Disable sliced VAE decoding. If `enable_vae_slicing` was previously invoked, this method will go back to
+        computing decoding in one step.
+        """
+        self.vae.disable_slicing()
+
+    # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.enable_vae_tiling
+    def enable_vae_tiling(self):
+        r"""
+        Enable tiled VAE decoding.
+
+        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()
+
+    # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.disable_vae_tiling
+    def disable_vae_tiling(self):
+        r"""
+        Disable tiled VAE decoding. If `enable_vae_tiling` was previously invoked, this method will go back to
+        computing decoding in one step.
+        """
+        self.vae.disable_tiling()
+
+    # Copied from diffusers.pipelines.stable_diffusion_xl.pipeline_stable_diffusion_xl_img2img.StableDiffusionXLImg2ImgPipeline.enable_sequential_cpu_offload
+    def enable_sequential_cpu_offload(self, gpu_id=0):
+        r"""
+        Offloads all models to CPU using accelerate, significantly reducing memory usage. When called, unet,
+        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_2, self.vae]:
+            cpu_offload(cpu_offloaded_model, device)
+
+        if self.text_encoder is not None:
+            cpu_offload(self.text_encoder, device)
+
+    # Copied from diffusers.pipelines.stable_diffusion_xl.pipeline_stable_diffusion_xl_img2img.StableDiffusionXLImg2ImgPipeline.enable_model_cpu_offload
+    def enable_model_cpu_offload(self, gpu_id=0):
+        r"""
+        Offloads all models to CPU using accelerate, reducing memory usage with a low impact on performance. Compared
+        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
+
+    # Copied from diffusers.pipelines.stable_diffusion_xl.pipeline_stable_diffusion_xl.StableDiffusionXLPipeline.encode_prompt
+    def encode_prompt(
+        self,
+        prompt,
+        device: Optional[torch.device] = None,
+        num_images_per_prompt: int = 1,
+        do_classifier_free_guidance: bool = True,
+        negative_prompt=None,
+        prompt_embeds: Optional[torch.FloatTensor] = None,
+        negative_prompt_embeds: Optional[torch.FloatTensor] = None,
+        pooled_prompt_embeds: Optional[torch.FloatTensor] = None,
+        negative_pooled_prompt_embeds: Optional[torch.FloatTensor] = None,
+        lora_scale: Optional[float] = None,
+    ):
+        r"""
+        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.
+            pooled_prompt_embeds (`torch.FloatTensor`, *optional*):
+                Pre-generated pooled text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting.
+                If not provided, pooled text embeddings will be generated from `prompt` input argument.
+            negative_pooled_prompt_embeds (`torch.FloatTensor`, *optional*):
+                Pre-generated negative pooled text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt
+                weighting. If not provided, pooled negative_prompt_embeds will be generated from `negative_prompt`
+                input argument.
+            lora_scale (`float`, *optional*):
+                A lora scale that will be applied to all LoRA layers of the text encoder if LoRA layers are loaded.
+        """
+        device = device or self._execution_device
+
+        # set lora scale so that monkey patched LoRA
+        # function of text encoder can correctly access it
+        if lora_scale is not None and isinstance(self, LoraLoaderMixin):
+            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)
+
+        bs_embed = pooled_prompt_embeds.shape[0]
+        pooled_prompt_embeds = pooled_prompt_embeds.repeat(1, num_images_per_prompt).view(
+            bs_embed * num_images_per_prompt, -1
+        )
+        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
+
+    @property
+    # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline._execution_device
+    def _execution_device(self):
+        r"""
+        Returns the device on which the pipeline's models will be executed. After calling
+        `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
+
+    # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.prepare_extra_step_kwargs
+    def prepare_extra_step_kwargs(self, generator, eta):
+        # prepare extra kwargs for the scheduler step, since not all schedulers have the same signature
+        # eta (η) is only used with the DDIMScheduler, it will be ignored for other schedulers.
+        # 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
+
+    # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion_inpaint.StableDiffusionInpaintPipeline.check_inputs
+    def check_inputs(
+        self,
+        prompt,
+        height,
+        width,
+        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 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,
+        image=None,
+        timestep=None,
+        is_strength_max=True,
+        add_noise=True,
+        return_noise=False,
+        return_image_latents=False,
+    ):
+        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 (image is None or timestep is None) and not is_strength_max:
+            raise ValueError(
+                "Since strength < 1. initial latents are to be initialised as a combination of Image + Noise."
+                "However, either the image or the noise timestep has not been provided."
+            )
+
+        if image.shape[1] == 4:
+            image_latents = image.to(device=device, dtype=dtype)
+        elif return_image_latents or (latents is None and not is_strength_max):
+            image = image.to(device=device, dtype=dtype)
+            image_latents = self._encode_vae_image(image=image, generator=generator)
+
+        if latents is None and add_noise:
+            noise = randn_tensor(shape, generator=generator, device=device, dtype=dtype)
+            # if strength is 1. then initialise the latents to noise, else initial to image + noise
+            latents = noise if is_strength_max else self.scheduler.add_noise(image_latents, noise, timestep)
+            # if pure noise then scale the initial latents by the  Scheduler's init sigma
+            latents = latents * self.scheduler.init_noise_sigma if is_strength_max else latents
+        elif add_noise:
+            noise = latents.to(device)
+            latents = noise * self.scheduler.init_noise_sigma
+        else:
+            noise = randn_tensor(shape, generator=generator, device=device, dtype=dtype)
+            latents = image_latents.to(device)
+
+        outputs = (latents,)
+
+        if return_noise:
+            outputs += (noise,)
+
+        if return_image_latents:
+            outputs += (image_latents,)
+
+        return outputs
+
+    def _encode_vae_image(self, image: torch.Tensor, generator: torch.Generator):
+        if self.vae.config.force_upcast:
+            dtype = image.dtype
+            image = image.float()
+            self.vae.to(dtype=torch.float32)
+
+        if isinstance(generator, list):
+            image_latents = [
+                self.vae.encode(image[i : i + 1]).latent_dist.sample(generator=generator[i])
+                for i in range(image.shape[0])
+            ]
+            image_latents = torch.cat(image_latents, dim=0)
+        else:
+            image_latents = self.vae.encode(image).latent_dist.sample(generator=generator)
+
+        if self.vae.config.force_upcast:
+            self.vae.to(dtype)
+
+        image_latents = image_latents.to(dtype)
+        image_latents = self.vae.config.scaling_factor * image_latents
+
+        return image_latents
+
+    def prepare_mask_latents(
+        self, mask, masked_image, batch_size, height, width, dtype, device, generator, do_classifier_free_guidance
+    ):
+        # resize the mask to latents shape as we concatenate the mask to the latents
+        # we do that before converting to dtype to avoid breaking in case we're using cpu_offload
+        # and half precision
+        mask = torch.nn.functional.interpolate(
+            mask, size=(height // self.vae_scale_factor, width // self.vae_scale_factor)
+        )
+        mask = mask.to(device=device, dtype=dtype)
+
+        # duplicate mask and masked_image_latents for each generation per prompt, using mps friendly method
+        if mask.shape[0] < batch_size:
+            if not batch_size % mask.shape[0] == 0:
+                raise ValueError(
+                    "The passed mask and the required batch size don't match. Masks are supposed to be duplicated to"
+                    f" a total batch size of {batch_size}, but {mask.shape[0]} masks were passed. Make sure the number"
+                    " of masks that you pass is divisible by the total requested batch size."
+                )
+            mask = mask.repeat(batch_size // mask.shape[0], 1, 1, 1)
+
+        mask = torch.cat([mask] * 2) if do_classifier_free_guidance else mask
+
+        masked_image_latents = None
+        if masked_image is not None:
+            masked_image = masked_image.to(device=device, dtype=dtype)
+            masked_image_latents = self._encode_vae_image(masked_image, generator=generator)
+            if masked_image_latents.shape[0] < batch_size:
+                if not batch_size % masked_image_latents.shape[0] == 0:
+                    raise ValueError(
+                        "The passed images and the required batch size don't match. Images are supposed to be duplicated"
+                        f" to a total batch size of {batch_size}, but {masked_image_latents.shape[0]} images were passed."
+                        " Make sure the number of images that you pass is divisible by the total requested batch size."
+                    )
+                masked_image_latents = masked_image_latents.repeat(
+                    batch_size // masked_image_latents.shape[0], 1, 1, 1
+                )
+
+            masked_image_latents = (
+                torch.cat([masked_image_latents] * 2) if do_classifier_free_guidance else masked_image_latents
+            )
+
+            # aligning device to prevent device errors when concating it with the latent model input
+            masked_image_latents = masked_image_latents.to(device=device, dtype=dtype)
+
+        return mask, masked_image_latents
+
+    # Copied from diffusers.pipelines.stable_diffusion_xl.pipeline_stable_diffusion_xl_img2img.StableDiffusionXLImg2ImgPipeline.get_timesteps
+    def get_timesteps(self, num_inference_steps, strength, device, denoising_start=None):
+        # get the original timestep using init_timestep
+        if denoising_start is None:
+            init_timestep = min(int(num_inference_steps * strength), num_inference_steps)
+            t_start = max(num_inference_steps - init_timestep, 0)
+        else:
+            t_start = int(round(denoising_start * num_inference_steps))
+
+        timesteps = self.scheduler.timesteps[t_start * self.scheduler.order :]
+
+        return timesteps, num_inference_steps - t_start
+
+    # Copied from diffusers.pipelines.stable_diffusion_xl.pipeline_stable_diffusion_xl_img2img.StableDiffusionXLImg2ImgPipeline._get_add_time_ids
+    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
+
+    # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion_upscale.StableDiffusionUpscalePipeline.upcast_vae
+    def upcast_vae(self):
+        dtype = self.vae.dtype
+        self.vae.to(dtype=torch.float32)
+        use_torch_2_0_or_xformers = isinstance(
+            self.vae.decoder.mid_block.attentions[0].processor,
+            (
+                AttnProcessor2_0,
+                XFormersAttnProcessor,
+                LoRAXFormersAttnProcessor,
+                LoRAAttnProcessor2_0,
+            ),
+        )
+        # if xformers or torch_2_0 is used attention block does not need
+        # to be in float32 which can save lots of memory
+        if use_torch_2_0_or_xformers:
+            self.vae.post_quant_conv.to(dtype)
+            self.vae.decoder.conv_in.to(dtype)
+            self.vae.decoder.mid_block.to(dtype)
+
+    @torch.no_grad()
+    def __call__(
+        self,
+        prompt: Union[str, List[str]] = None,
+        image: Union[torch.FloatTensor, PIL.Image.Image] = None,
+        mask_image: Union[torch.FloatTensor, PIL.Image.Image] = None,
+        height: Optional[int] = None,
+        width: Optional[int] = None,
+        strength: float = 1.0,
+        num_inference_steps: int = 50,
+        denoising_start: Optional[float] = None,
+        denoising_end: Optional[float] = None,
+        guidance_scale: float = 7.5,
+        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,
+        pooled_prompt_embeds: Optional[torch.FloatTensor] = None,
+        negative_pooled_prompt_embeds: Optional[torch.FloatTensor] = None,
+        output_type: Optional[str] = "pil",
+        return_dict: bool = True,
+        callback: Optional[Callable[[int, int, torch.FloatTensor], None]] = None,
+        callback_steps: int = 1,
+        cross_attention_kwargs: Optional[Dict[str, Any]] = None,
+        guidance_rescale: float = 0.0,
+        original_size: Tuple[int, int] = None,
+        crops_coords_top_left: Tuple[int, int] = (0, 0),
+        target_size: Tuple[int, int] = None,
+        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 (`PIL.Image.Image`):
+                `Image`, or tensor representing an image batch which will be inpainted, *i.e.* parts of the image will
+                be masked out with `mask_image` and repainted according to `prompt`.
+            mask_image (`PIL.Image.Image`):
+                `Image`, or tensor representing an image batch, to mask `image`. White pixels in the mask will be
+                repainted, while black pixels will be preserved. If `mask_image` is a PIL image, it will be converted
+                to a single channel (luminance) before use. If it's a tensor, it should contain one color channel (L)
+                instead of 3, so the expected shape would be `(B, H, W, 1)`.
+            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.
+            strength (`float`, *optional*, defaults to 1.):
+                Conceptually, indicates how much to transform the masked portion of 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 the masked
+                portion of the reference `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.
+            denoising_start (`float`, *optional*):
+                When specified, indicates the fraction (between 0.0 and 1.0) of the total denoising process to be
+                bypassed before it is initiated. For example, if `denoising_start` is set to 0.7 and
+                num_inference_steps is fixed at 50, the process will begin only from the 35th (i.e., 0.7 * 50)
+                denoising step. Consequently, the initial part of the denoising process is skipped and it is assumed
+                that the passed `image` is a partly denoised image. The `denoising_start` parameter is particularly
+                beneficial when this pipeline is integrated into a "Mixture of Denoisers" multi-pipeline setup, as
+                detailed in [**Refining the Image
+                Output**](https://huggingface.co/docs/diffusers/api/pipelines/stable_diffusion/stable_diffusion_xl#refining-the-image-output).
+            denoising_end (`float`, *optional*):
+                When specified, determines the fraction (between 0.0 and 1.0) of the total denoising process to be
+                completed before it is intentionally prematurely terminated. For instance, if denoising_end is set to
+                0.7 and `num_inference_steps` is fixed at 50, the process will execute only 35 (i.e., 0.7 * 50)
+                denoising steps. As a result, the returned sample will still retain a substantial amount of noise (ca.
+                30%) and should be denoised by a successor pipeline that has `denoising_start` set to 0.7 so that it
+                only denoised the final 30%. The denoising_end parameter should ideally be utilized when this pipeline
+                forms a part of a "Mixture of Denoisers" multi-pipeline setup, as elaborated in [**Refining the Image
+                Output**](https://huggingface.co/docs/diffusers/api/pipelines/stable_diffusion/stable_diffusion_xl#refining-the-image-output).
+            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.
+            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.
+            pooled_prompt_embeds (`torch.FloatTensor`, *optional*):
+                Pre-generated pooled text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting.
+                If not provided, pooled text embeddings will be generated from `prompt` input argument.
+            negative_pooled_prompt_embeds (`torch.FloatTensor`, *optional*):
+                Pre-generated negative pooled text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt
+                weighting. If not provided, pooled negative_prompt_embeds will be generated from `negative_prompt`
+                input argument.
+            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`, *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.StableDiffusionPipelineOutput`] 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).
+        Examples:
+
+        ```py
+        >>> import PIL
+        >>> import requests
+        >>> import torch
+        >>> from io import BytesIO
+
+        >>> from diffusers import StableDiffusionInpaintPipeline
+
+
+        >>> def download_image(url):
+        ...     response = requests.get(url)
+        ...     return PIL.Image.open(BytesIO(response.content)).convert("RGB")
+
+
+        >>> img_url = "https://raw.githubusercontent.com/CompVis/latent-diffusion/main/data/inpainting_examples/overture-creations-5sI6fQgYIuo.png"
+        >>> mask_url = "https://raw.githubusercontent.com/CompVis/latent-diffusion/main/data/inpainting_examples/overture-creations-5sI6fQgYIuo_mask.png"
+
+        >>> init_image = download_image(img_url).resize((512, 512))
+        >>> mask_image = download_image(mask_url).resize((512, 512))
+
+        >>> pipe = StableDiffusionInpaintPipeline.from_pretrained(
+        ...     "runwayml/stable-diffusion-inpainting", torch_dtype=torch.float16
+        ... )
+        >>> pipe = pipe.to("cuda")
+
+        >>> prompt = "Face of a yellow cat, high resolution, sitting on a park bench"
+        >>> image = pipe(prompt=prompt, image=init_image, mask_image=mask_image).images[0]
+        ```
+
+        Returns:
+            [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] or `tuple`:
+            [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] 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
+        self.check_inputs(
+            prompt,
+            height,
+            width,
+            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,
+            pooled_prompt_embeds=pooled_prompt_embeds,
+            negative_pooled_prompt_embeds=negative_pooled_prompt_embeds,
+            lora_scale=text_encoder_lora_scale,
+        )
+
+        # 4. set timesteps
+        original_num_steps = num_inference_steps  # save for denoising_start/end later
+        self.scheduler.set_timesteps(num_inference_steps, device=device)
+        timesteps, num_inference_steps = self.get_timesteps(
+            num_inference_steps, strength, device, denoising_start=denoising_start
+        )
+        # check that number of inference steps is not < 1 - as this doesn't make sense
+        if num_inference_steps < 1:
+            raise ValueError(
+                f"After adjusting the num_inference_steps by strength parameter: {strength}, the number of pipeline"
+                f"steps is {num_inference_steps} which is < 1 and not appropriate for this pipeline."
+            )
+        # at which timestep to set the initial noise (n.b. 50% if strength is 0.5)
+        latent_timestep = timesteps[:1].repeat(batch_size * num_images_per_prompt)
+        # create a boolean to check if the strength is set to 1. if so then initialise the latents with pure noise
+        is_strength_max = strength == 1.0
+
+        # 5. Preprocess mask and image
+        mask, masked_image, init_image = prepare_mask_and_masked_image(
+            image, mask_image, height, width, return_image=True
+        )
+
+        # 6. Prepare latent variables
+        num_channels_latents = self.vae.config.latent_channels
+        num_channels_unet = self.unet.config.in_channels
+        return_image_latents = num_channels_unet == 4
+
+        add_noise = True if denoising_start is None else False
+        latents_outputs = self.prepare_latents(
+            batch_size * num_images_per_prompt,
+            num_channels_latents,
+            height,
+            width,
+            prompt_embeds.dtype,
+            device,
+            generator,
+            latents,
+            image=init_image,
+            timestep=latent_timestep,
+            is_strength_max=is_strength_max,
+            add_noise=add_noise,
+            return_noise=True,
+            return_image_latents=return_image_latents,
+        )
+
+        if return_image_latents:
+            latents, noise, image_latents = latents_outputs
+        else:
+            latents, noise = latents_outputs
+
+        # 7. Prepare mask latent variables
+        mask, masked_image_latents = self.prepare_mask_latents(
+            mask,
+            masked_image,
+            batch_size * num_images_per_prompt,
+            height,
+            width,
+            prompt_embeds.dtype,
+            device,
+            generator,
+            do_classifier_free_guidance,
+        )
+
+        # 8. Check that sizes of mask, masked image and latents match
+        if num_channels_unet == 9:
+            # default case for runwayml/stable-diffusion-inpainting
+            num_channels_mask = mask.shape[1]
+            num_channels_masked_image = masked_image_latents.shape[1]
+            if num_channels_latents + num_channels_mask + num_channels_masked_image != self.unet.config.in_channels:
+                raise ValueError(
+                    f"Incorrect configuration settings! The config of `pipeline.unet`: {self.unet.config} expects"
+                    f" {self.unet.config.in_channels} but received `num_channels_latents`: {num_channels_latents} +"
+                    f" `num_channels_mask`: {num_channels_mask} + `num_channels_masked_image`: {num_channels_masked_image}"
+                    f" = {num_channels_latents+num_channels_masked_image+num_channels_mask}. Please verify the config of"
+                    " `pipeline.unet` or your `mask_image` or `image` input."
+                )
+        elif num_channels_unet != 4:
+            raise ValueError(
+                f"The unet {self.unet.__class__} should have either 4 or 9 input channels, not {self.unet.config.in_channels}."
+            )
+        # 8.1 Prepare extra step kwargs.
+        extra_step_kwargs = self.prepare_extra_step_kwargs(generator, eta)
+
+        # 9. Prepare extra step kwargs. TODO: Logic should ideally just be moved out of the pipeline
+        height, width = latents.shape[-2:]
+        height = height * self.vae_scale_factor
+        width = width * self.vae_scale_factor
+
+        original_size = original_size or (height, width)
+        target_size = target_size or (height, width)
+
+        # 10. 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)
+
+        # 11. Denoising loop
+        num_warmup_steps = max(len(timesteps) - num_inference_steps * self.scheduler.order, 0)
+
+        if denoising_end is not None and denoising_start is not None:
+            if denoising_start >= denoising_end:
+                raise ValueError(
+                    f"`denoising_end`: {denoising_end} cannot be larger than `denoising_start`: {denoising_start}."
+                )
+
+            skipped_final_steps = int(round((1 - denoising_end) * original_num_steps))
+            num_inference_steps = num_inference_steps - skipped_final_steps
+            timesteps = timesteps[: num_warmup_steps + self.scheduler.order * num_inference_steps]
+        elif denoising_end is not None:
+            num_inference_steps = int(round(denoising_end * num_inference_steps))
+            timesteps = timesteps[: num_warmup_steps + self.scheduler.order * num_inference_steps]
+
+        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
+
+                # concat latents, mask, masked_image_latents in the channel dimension
+                latent_model_input = self.scheduler.scale_model_input(latent_model_input, t)
+
+                if num_channels_unet == 9:
+                    latent_model_input = torch.cat([latent_model_input, mask, masked_image_latents], dim=1)
+
+                # 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]
+
+                if num_channels_unet == 4:
+                    init_latents_proper = image_latents[:1]
+                    init_mask = mask[:1]
+
+                    if i < len(timesteps) - 1:
+                        noise_timestep = timesteps[i + 1]
+                        init_latents_proper = self.scheduler.add_noise(
+                            init_latents_proper, noise, torch.tensor([noise_timestep])
+                        )
+
+                    latents = (1 - init_mask) * init_latents_proper + init_mask * latents
+
+                # 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
+        if self.vae.dtype == torch.float16 and self.vae.config.force_upcast:
+            self.upcast_vae()
+            latents = latents.to(next(iter(self.vae.post_quant_conv.parameters())).dtype)
+
+        if not output_type == "latent":
+            image = self.vae.decode(latents / self.vae.config.scaling_factor, return_dict=False)[0]
+        else:
+            return StableDiffusionXLPipelineOutput(images=latents)
+
+        image = self.image_processor.postprocess(image, output_type=output_type)
+
+        # Offload last model to CPU
+        if hasattr(self, "final_offload_hook") and self.final_offload_hook is not None:
+            self.final_offload_hook.offload()
+
+        if not return_dict:
+            return (image,)
+
+        return StableDiffusionXLPipelineOutput(images=image)

src/diffusers/utils/dummy_torch_and_transformers_and_invisible_watermark_objects.py

@@ -17,6 +17,21 @@ class StableDiffusionXLImg2ImgPipeline(metaclass=DummyObject):
         requires_backends(cls, ["torch", "transformers", "invisible_watermark"])
 
 
+class StableDiffusionXLInpaintPipeline(metaclass=DummyObject):
+    _backends = ["torch", "transformers", "invisible_watermark"]
+
+    def __init__(self, *args, **kwargs):
+        requires_backends(self, ["torch", "transformers", "invisible_watermark"])
+
+    @classmethod
+    def from_config(cls, *args, **kwargs):
+        requires_backends(cls, ["torch", "transformers", "invisible_watermark"])
+
+    @classmethod
+    def from_pretrained(cls, *args, **kwargs):
+        requires_backends(cls, ["torch", "transformers", "invisible_watermark"])
+
+
 class StableDiffusionXLPipeline(metaclass=DummyObject):
     _backends = ["torch", "transformers", "invisible_watermark"]
 

tests/pipelines/stable_diffusion_xl/test_stable_diffusion_xl_inpaint.py

+# coding=utf-8
+# Copyright 2023 HuggingFace Inc.
+#
+# 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 copy
+import random
+import unittest
+
+import numpy as np
+import torch
+from PIL import Image
+from transformers import CLIPTextConfig, CLIPTextModel, CLIPTextModelWithProjection, CLIPTokenizer
+
+from diffusers import (
+    AutoencoderKL,
+    DDIMScheduler,
+    DPMSolverMultistepScheduler,
+    EulerDiscreteScheduler,
+    HeunDiscreteScheduler,
+    StableDiffusionXLInpaintPipeline,
+    UNet2DConditionModel,
+    UniPCMultistepScheduler,
+)
+from diffusers.utils import floats_tensor, torch_device
+from diffusers.utils.testing_utils import enable_full_determinism, require_torch_gpu
+
+from ..pipeline_params import TEXT_GUIDED_IMAGE_INPAINTING_BATCH_PARAMS, TEXT_GUIDED_IMAGE_INPAINTING_PARAMS
+from ..test_pipelines_common import PipelineLatentTesterMixin, PipelineTesterMixin
+
+
+enable_full_determinism()
+
+
+class StableDiffusionXLInpaintPipelineFastTests(PipelineLatentTesterMixin, PipelineTesterMixin, unittest.TestCase):
+    pipeline_class = StableDiffusionXLInpaintPipeline
+    params = TEXT_GUIDED_IMAGE_INPAINTING_PARAMS
+    batch_params = TEXT_GUIDED_IMAGE_INPAINTING_BATCH_PARAMS
+    image_params = frozenset([])
+    # TO-DO: update image_params once pipeline is refactored with VaeImageProcessor.preprocess
+    image_latents_params = frozenset([])
+
+    def get_dummy_components(self):
+        torch.manual_seed(0)
+        unet = UNet2DConditionModel(
+            block_out_channels=(32, 64),
+            layers_per_block=2,
+            sample_size=32,
+            in_channels=4,
+            out_channels=4,
+            down_block_types=("DownBlock2D", "CrossAttnDownBlock2D"),
+            up_block_types=("CrossAttnUpBlock2D", "UpBlock2D"),
+            # SD2-specific config below
+            attention_head_dim=(2, 4),
+            use_linear_projection=True,
+            addition_embed_type="text_time",
+            addition_time_embed_dim=8,
+            transformer_layers_per_block=(1, 2),
+            projection_class_embeddings_input_dim=80,  # 6 * 8 + 32
+            cross_attention_dim=64,
+        )
+        scheduler = EulerDiscreteScheduler(
+            beta_start=0.00085,
+            beta_end=0.012,
+            steps_offset=1,
+            beta_schedule="scaled_linear",
+            timestep_spacing="leading",
+        )
+        torch.manual_seed(0)
+        vae = AutoencoderKL(
+            block_out_channels=[32, 64],
+            in_channels=3,
+            out_channels=3,
+            down_block_types=["DownEncoderBlock2D", "DownEncoderBlock2D"],
+            up_block_types=["UpDecoderBlock2D", "UpDecoderBlock2D"],
+            latent_channels=4,
+            sample_size=128,
+        )
+        torch.manual_seed(0)
+        text_encoder_config = CLIPTextConfig(
+            bos_token_id=0,
+            eos_token_id=2,
+            hidden_size=32,
+            intermediate_size=37,
+            layer_norm_eps=1e-05,
+            num_attention_heads=4,
+            num_hidden_layers=5,
+            pad_token_id=1,
+            vocab_size=1000,
+            # SD2-specific config below
+            hidden_act="gelu",
+            projection_dim=32,
+        )
+        text_encoder = CLIPTextModel(text_encoder_config)
+        tokenizer = CLIPTokenizer.from_pretrained("hf-internal-testing/tiny-random-clip", local_files_only=True)
+
+        text_encoder_2 = CLIPTextModelWithProjection(text_encoder_config)
+        tokenizer_2 = CLIPTokenizer.from_pretrained("hf-internal-testing/tiny-random-clip", local_files_only=True)
+
+        components = {
+            "unet": unet,
+            "scheduler": scheduler,
+            "vae": vae,
+            "text_encoder": text_encoder,
+            "tokenizer": tokenizer,
+            "text_encoder_2": text_encoder_2,
+            "tokenizer_2": tokenizer_2,
+        }
+        return components
+
+    def get_dummy_inputs(self, device, seed=0):
+        # TODO: use tensor inputs instead of PIL, this is here just to leave the old expected_slices untouched
+        image = floats_tensor((1, 3, 32, 32), rng=random.Random(seed)).to(device)
+        image = image.cpu().permute(0, 2, 3, 1)[0]
+        init_image = Image.fromarray(np.uint8(image)).convert("RGB").resize((64, 64))
+        mask_image = Image.fromarray(np.uint8(image + 4)).convert("RGB").resize((64, 64))
+        if str(device).startswith("mps"):
+            generator = torch.manual_seed(seed)
+        else:
+            generator = torch.Generator(device=device).manual_seed(seed)
+        inputs = {
+            "prompt": "A painting of a squirrel eating a burger",
+            "image": init_image,
+            "mask_image": mask_image,
+            "generator": generator,
+            "num_inference_steps": 2,
+            "guidance_scale": 6.0,
+            "output_type": "numpy",
+        }
+        return inputs
+
+    def test_stable_diffusion_xl_inpaint_euler(self):
+        device = "cpu"  # ensure determinism for the device-dependent torch.Generator
+        components = self.get_dummy_components()
+        sd_pipe = StableDiffusionXLInpaintPipeline(**components)
+        sd_pipe = sd_pipe.to(device)
+        sd_pipe.set_progress_bar_config(disable=None)
+
+        inputs = self.get_dummy_inputs(device)
+        image = sd_pipe(**inputs).images
+        image_slice = image[0, -3:, -3:, -1]
+
+        assert image.shape == (1, 64, 64, 3)
+
+        expected_slice = np.array([0.4924, 0.4966, 0.4100, 0.5233, 0.5322, 0.4532, 0.5804, 0.5876, 0.4150])
+
+        assert np.abs(image_slice.flatten() - expected_slice).max() < 1e-2
+
+    def test_attention_slicing_forward_pass(self):
+        super().test_attention_slicing_forward_pass(expected_max_diff=3e-3)
+
+    def test_inference_batch_single_identical(self):
+        super().test_inference_batch_single_identical(expected_max_diff=3e-3)
+
+    # TODO(Patrick, Sayak) - skip for now as this requires more refiner tests
+    def test_save_load_optional_components(self):
+        pass
+
+    def test_stable_diffusion_xl_inpaint_negative_prompt_embeds(self):
+        components = self.get_dummy_components()
+        sd_pipe = StableDiffusionXLInpaintPipeline(**components)
+        sd_pipe = sd_pipe.to(torch_device)
+        sd_pipe = sd_pipe.to(torch_device)
+        sd_pipe.set_progress_bar_config(disable=None)
+
+        # forward without prompt embeds
+        inputs = self.get_dummy_inputs(torch_device)
+        negative_prompt = 3 * ["this is a negative prompt"]
+        inputs["negative_prompt"] = negative_prompt
+        inputs["prompt"] = 3 * [inputs["prompt"]]
+
+        output = sd_pipe(**inputs)
+        image_slice_1 = output.images[0, -3:, -3:, -1]
+
+        # forward with prompt embeds
+        inputs = self.get_dummy_inputs(torch_device)
+        negative_prompt = 3 * ["this is a negative prompt"]
+        prompt = 3 * [inputs.pop("prompt")]
+
+        (
+            prompt_embeds,
+            negative_prompt_embeds,
+            pooled_prompt_embeds,
+            negative_pooled_prompt_embeds,
+        ) = sd_pipe.encode_prompt(prompt, negative_prompt=negative_prompt)
+
+        output = sd_pipe(
+            **inputs,
+            prompt_embeds=prompt_embeds,
+            negative_prompt_embeds=negative_prompt_embeds,
+            pooled_prompt_embeds=pooled_prompt_embeds,
+            negative_pooled_prompt_embeds=negative_pooled_prompt_embeds,
+        )
+        image_slice_2 = output.images[0, -3:, -3:, -1]
+
+        # make sure that it's equal
+        assert np.abs(image_slice_1.flatten() - image_slice_2.flatten()).max() < 1e-4
+
+    @require_torch_gpu
+    def test_stable_diffusion_xl_offloads(self):
+        pipes = []
+        components = self.get_dummy_components()
+        sd_pipe = StableDiffusionXLInpaintPipeline(**components).to(torch_device)
+        pipes.append(sd_pipe)
+
+        components = self.get_dummy_components()
+        sd_pipe = StableDiffusionXLInpaintPipeline(**components)
+        sd_pipe.enable_model_cpu_offload()
+        pipes.append(sd_pipe)
+
+        components = self.get_dummy_components()
+        sd_pipe = StableDiffusionXLInpaintPipeline(**components)
+        sd_pipe.enable_sequential_cpu_offload()
+        pipes.append(sd_pipe)
+
+        image_slices = []
+        for pipe in pipes:
+            pipe.unet.set_default_attn_processor()
+
+            inputs = self.get_dummy_inputs(torch_device)
+            image = pipe(**inputs).images
+
+            image_slices.append(image[0, -3:, -3:, -1].flatten())
+
+        assert np.abs(image_slices[0] - image_slices[1]).max() < 1e-3
+        assert np.abs(image_slices[0] - image_slices[2]).max() < 1e-3
+
+    def test_stable_diffusion_two_xl_mixture_of_denoiser(self):
+        components = self.get_dummy_components()
+        pipe_1 = StableDiffusionXLInpaintPipeline(**components).to(torch_device)
+        pipe_1.unet.set_default_attn_processor()
+        pipe_2 = StableDiffusionXLInpaintPipeline(**components).to(torch_device)
+        pipe_2.unet.set_default_attn_processor()
+
+        def assert_run_mixture(num_steps, split, scheduler_cls_orig):
+            inputs = self.get_dummy_inputs(torch_device)
+            inputs["num_inference_steps"] = num_steps
+
+            class scheduler_cls(scheduler_cls_orig):
+                pass
+
+            pipe_1.scheduler = scheduler_cls.from_config(pipe_1.scheduler.config)
+            pipe_2.scheduler = scheduler_cls.from_config(pipe_2.scheduler.config)
+
+            # Let's retrieve the number of timesteps we want to use
+            pipe_1.scheduler.set_timesteps(num_steps)
+            expected_steps = pipe_1.scheduler.timesteps.tolist()
+
+            split_id = int(round(split * num_steps)) * pipe_1.scheduler.order
+            expected_steps_1 = expected_steps[:split_id]
+            expected_steps_2 = expected_steps[split_id:]
+
+            # now we monkey patch step `done_steps`
+            # list into the step function for testing
+            done_steps = []
+            old_step = copy.copy(scheduler_cls.step)
+
+            def new_step(self, *args, **kwargs):
+                done_steps.append(args[1].cpu().item())  # args[1] is always the passed `t`
+                return old_step(self, *args, **kwargs)
+
+            scheduler_cls.step = new_step
+
+            inputs_1 = {**inputs, **{"denoising_end": split, "output_type": "latent"}}
+            latents = pipe_1(**inputs_1).images[0]
+
+            assert expected_steps_1 == done_steps, f"Failure with {scheduler_cls.__name__} and {num_steps} and {split}"
+
+            inputs_2 = {**inputs, **{"denoising_start": split, "image": latents}}
+            pipe_2(**inputs_2).images[0]
+
+            assert expected_steps_2 == done_steps[len(expected_steps_1) :]
+            assert expected_steps == done_steps, f"Failure with {scheduler_cls.__name__} and {num_steps} and {split}"
+
+        for steps in [5, 8]:
+            for split in [0.33, 0.49, 0.71]:
+                for scheduler_cls in [
+                    DDIMScheduler,
+                    EulerDiscreteScheduler,
+                    DPMSolverMultistepScheduler,
+                    UniPCMultistepScheduler,
+                    HeunDiscreteScheduler,
+                ]:
+                    assert_run_mixture(steps, split, scheduler_cls)
+
+    def test_stable_diffusion_three_xl_mixture_of_denoiser(self):
+        components = self.get_dummy_components()
+        pipe_1 = StableDiffusionXLInpaintPipeline(**components).to(torch_device)
+        pipe_1.unet.set_default_attn_processor()
+        pipe_2 = StableDiffusionXLInpaintPipeline(**components).to(torch_device)
+        pipe_2.unet.set_default_attn_processor()
+        pipe_3 = StableDiffusionXLInpaintPipeline(**components).to(torch_device)
+        pipe_3.unet.set_default_attn_processor()
+
+        def assert_run_mixture(num_steps, split_1, split_2, scheduler_cls_orig):
+            inputs = self.get_dummy_inputs(torch_device)
+            inputs["num_inference_steps"] = num_steps
+
+            class scheduler_cls(scheduler_cls_orig):
+                pass
+
+            pipe_1.scheduler = scheduler_cls.from_config(pipe_1.scheduler.config)
+            pipe_2.scheduler = scheduler_cls.from_config(pipe_2.scheduler.config)
+            pipe_3.scheduler = scheduler_cls.from_config(pipe_3.scheduler.config)
+
+            # Let's retrieve the number of timesteps we want to use
+            pipe_1.scheduler.set_timesteps(num_steps)
+            expected_steps = pipe_1.scheduler.timesteps.tolist()
+
+            split_id_1 = int(round(split_1 * num_steps)) * pipe_1.scheduler.order
+            split_id_2 = int(round(split_2 * num_steps)) * pipe_1.scheduler.order
+            expected_steps_1 = expected_steps[:split_id_1]
+            expected_steps_2 = expected_steps[split_id_1:split_id_2]
+            expected_steps_3 = expected_steps[split_id_2:]
+
+            # now we monkey patch step `done_steps`
+            # list into the step function for testing
+            done_steps = []
+            old_step = copy.copy(scheduler_cls.step)
+
+            def new_step(self, *args, **kwargs):
+                done_steps.append(args[1].cpu().item())  # args[1] is always the passed `t`
+                return old_step(self, *args, **kwargs)
+
+            scheduler_cls.step = new_step
+
+            inputs_1 = {**inputs, **{"denoising_end": split_1, "output_type": "latent"}}
+            latents = pipe_1(**inputs_1).images[0]
+
+            assert (
+                expected_steps_1 == done_steps
+            ), f"Failure with {scheduler_cls.__name__} and {num_steps} and {split_1} and {split_2}"
+
+            inputs_2 = {
+                **inputs,
+                **{"denoising_start": split_1, "denoising_end": split_2, "image": latents, "output_type": "latent"},
+            }
+            pipe_2(**inputs_2).images[0]
+
+            assert expected_steps_2 == done_steps[len(expected_steps_1) :]
+
+            inputs_3 = {**inputs, **{"denoising_start": split_2, "image": latents}}
+            pipe_3(**inputs_3).images[0]
+
+            assert expected_steps_3 == done_steps[len(expected_steps_1) + len(expected_steps_2) :]
+            assert (
+                expected_steps == done_steps
+            ), f"Failure with {scheduler_cls.__name__} and {num_steps} and {split_1} and {split_2}"
+
+        for steps in [7, 11]:
+            for split_1, split_2 in zip([0.19, 0.32], [0.81, 0.68]):
+                for scheduler_cls in [
+                    DDIMScheduler,
+                    EulerDiscreteScheduler,
+                    DPMSolverMultistepScheduler,
+                    UniPCMultistepScheduler,
+                    HeunDiscreteScheduler,
+                ]:
+                    assert_run_mixture(steps, split_1, split_2, scheduler_cls)