diffusers 0.24.0__py3-none-any.whl → 0.25.0__py3-none-any.whl

diffusers/pipelines/stable_diffusion_sag/__init__.py

@@ -0,0 +1,48 @@
+from typing import TYPE_CHECKING
+
+from ...utils import (
+    DIFFUSERS_SLOW_IMPORT,
+    OptionalDependencyNotAvailable,
+    _LazyModule,
+    get_objects_from_module,
+    is_torch_available,
+    is_transformers_available,
+)
+
+
+_dummy_objects = {}
+_import_structure = {}
+
+
+try:
+    if not (is_transformers_available() and is_torch_available()):
+        raise OptionalDependencyNotAvailable()
+except OptionalDependencyNotAvailable:
+    from ...utils import dummy_torch_and_transformers_objects  # noqa F403
+
+    _dummy_objects.update(get_objects_from_module(dummy_torch_and_transformers_objects))
+else:
+    _import_structure["pipeline_stable_diffusion_sag"] = ["StableDiffusionSAGPipeline"]
+
+if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT:
+    try:
+        if not (is_transformers_available() and is_torch_available()):
+            raise OptionalDependencyNotAvailable()
+
+    except OptionalDependencyNotAvailable:
+        from ...utils.dummy_torch_and_transformers_objects import *
+    else:
+        from .pipeline_stable_diffusion_sag import StableDiffusionSAGPipeline
+
+else:
+    import sys
+
+    sys.modules[__name__] = _LazyModule(
+        __name__,
+        globals()["__file__"],
+        _import_structure,
+        module_spec=__spec__,
+    )
+
+    for name, value in _dummy_objects.items():
+        setattr(sys.modules[__name__], name, value)

diffusers/pipelines/{stable_diffusion → stable_diffusion_sag}/pipeline_stable_diffusion_sag.py

@@ -17,11 +17,11 @@ from typing import Any, Callable, Dict, List, Optional, Union
 
 import torch
 import torch.nn.functional as F
-from transformers import CLIPImageProcessor, CLIPTextModel, CLIPTokenizer
+from transformers import CLIPImageProcessor, CLIPTextModel, CLIPTokenizer, CLIPVisionModelWithProjection
 
-from ...image_processor import VaeImageProcessor
-from ...loaders import LoraLoaderMixin, TextualInversionLoaderMixin
-from ...models import AutoencoderKL, UNet2DConditionModel
+from ...image_processor import PipelineImageInput, VaeImageProcessor
+from ...loaders import IPAdapterMixin, LoraLoaderMixin, TextualInversionLoaderMixin
+from ...models import AutoencoderKL, ImageProjection, UNet2DConditionModel
 from ...models.lora import adjust_lora_scale_text_encoder
 from ...schedulers import KarrasDiffusionSchedulers
 from ...utils import (
@@ -34,8 +34,8 @@ from ...utils import (
 )
 from ...utils.torch_utils import randn_tensor
 from ..pipeline_utils import DiffusionPipeline
-from . import StableDiffusionPipelineOutput
-from .safety_checker import StableDiffusionSafetyChecker
+from ..stable_diffusion import StableDiffusionPipelineOutput
+from ..stable_diffusion.safety_checker import StableDiffusionSafetyChecker
 
 
 logger = logging.get_logger(__name__)  # pylint: disable=invalid-name
@@ -98,13 +98,17 @@ class CrossAttnStoreProcessor:
 
 
 # Modified to get self-attention guidance scale in this paper (https://arxiv.org/pdf/2210.00939.pdf) as an input
-class StableDiffusionSAGPipeline(DiffusionPipeline, TextualInversionLoaderMixin):
+class StableDiffusionSAGPipeline(DiffusionPipeline, TextualInversionLoaderMixin, IPAdapterMixin):
     r"""
     Pipeline for text-to-image generation using Stable Diffusion.
 
     This model inherits from [`DiffusionPipeline`]. Check the superclass documentation for the generic methods
     implemented for all pipelines (downloading, saving, running on a particular device, etc.).
 
+    The pipeline also inherits the following loading methods:
+        - [`~loaders.TextualInversionLoaderMixin.load_textual_inversion`] for loading textual inversion embeddings
+        - [`~loaders.IPAdapterMixin.load_ip_adapter`] for loading IP Adapters
+
     Args:
         vae ([`AutoencoderKL`]):
             Variational Auto-Encoder (VAE) model to encode and decode images to and from latent representations.
@@ -126,7 +130,7 @@ class StableDiffusionSAGPipeline(DiffusionPipeline, TextualInversionLoaderMixin)
     """
 
     model_cpu_offload_seq = "text_encoder->unet->vae"
-    _optional_components = ["safety_checker", "feature_extractor"]
+    _optional_components = ["safety_checker", "feature_extractor", "image_encoder"]
     _exclude_from_cpu_offload = ["safety_checker"]
 
     def __init__(
@@ -138,6 +142,7 @@ class StableDiffusionSAGPipeline(DiffusionPipeline, TextualInversionLoaderMixin)
         scheduler: KarrasDiffusionSchedulers,
         safety_checker: StableDiffusionSafetyChecker,
         feature_extractor: CLIPImageProcessor,
+        image_encoder: Optional[CLIPVisionModelWithProjection] = None,
         requires_safety_checker: bool = True,
     ):
         super().__init__()
@@ -150,6 +155,7 @@ class StableDiffusionSAGPipeline(DiffusionPipeline, TextualInversionLoaderMixin)
             scheduler=scheduler,
             safety_checker=safety_checker,
             feature_extractor=feature_extractor,
+            image_encoder=image_encoder,
         )
         self.vae_scale_factor = 2 ** (len(self.vae.config.block_out_channels) - 1)
         self.image_processor = VaeImageProcessor(vae_scale_factor=self.vae_scale_factor)
@@ -386,6 +392,31 @@ class StableDiffusionSAGPipeline(DiffusionPipeline, TextualInversionLoaderMixin)
 
         return prompt_embeds, negative_prompt_embeds
 
+    # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.encode_image
+    def encode_image(self, image, device, num_images_per_prompt, output_hidden_states=None):
+        dtype = next(self.image_encoder.parameters()).dtype
+
+        if not isinstance(image, torch.Tensor):
+            image = self.feature_extractor(image, return_tensors="pt").pixel_values
+
+        image = image.to(device=device, dtype=dtype)
+        if output_hidden_states:
+            image_enc_hidden_states = self.image_encoder(image, output_hidden_states=True).hidden_states[-2]
+            image_enc_hidden_states = image_enc_hidden_states.repeat_interleave(num_images_per_prompt, dim=0)
+            uncond_image_enc_hidden_states = self.image_encoder(
+                torch.zeros_like(image), output_hidden_states=True
+            ).hidden_states[-2]
+            uncond_image_enc_hidden_states = uncond_image_enc_hidden_states.repeat_interleave(
+                num_images_per_prompt, dim=0
+            )
+            return image_enc_hidden_states, uncond_image_enc_hidden_states
+        else:
+            image_embeds = self.image_encoder(image).image_embeds
+            image_embeds = image_embeds.repeat_interleave(num_images_per_prompt, dim=0)
+            uncond_image_embeds = torch.zeros_like(image_embeds)
+
+            return image_embeds, uncond_image_embeds
+
     # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.run_safety_checker
     def run_safety_checker(self, image, device, dtype):
         if self.safety_checker is None:
@@ -519,6 +550,7 @@ class StableDiffusionSAGPipeline(DiffusionPipeline, TextualInversionLoaderMixin)
         latents: Optional[torch.FloatTensor] = None,
         prompt_embeds: Optional[torch.FloatTensor] = None,
         negative_prompt_embeds: Optional[torch.FloatTensor] = None,
+        ip_adapter_image: Optional[PipelineImageInput] = None,
         output_type: Optional[str] = "pil",
         return_dict: bool = True,
         callback: Optional[Callable[[int, int, torch.FloatTensor], None]] = None,
@@ -565,6 +597,8 @@ class StableDiffusionSAGPipeline(DiffusionPipeline, TextualInversionLoaderMixin)
             negative_prompt_embeds (`torch.FloatTensor`, *optional*):
                 Pre-generated negative text embeddings. Can be used to easily tweak text inputs (prompt weighting). If
                 not provided, `negative_prompt_embeds` are generated from the `negative_prompt` input argument.
+            ip_adapter_image: (`PipelineImageInput`, *optional*):
+                Optional image input to work with IP Adapters.
             output_type (`str`, *optional*, defaults to `"pil"`):
                 The output format of the generated image. Choose between `PIL.Image` or `np.array`.
             return_dict (`bool`, *optional*, defaults to `True`):
@@ -618,6 +652,14 @@ class StableDiffusionSAGPipeline(DiffusionPipeline, TextualInversionLoaderMixin)
         # `sag_scale = 0` means no self-attention guidance
         do_self_attention_guidance = sag_scale > 0.0
 
+        if ip_adapter_image is not None:
+            output_hidden_state = False if isinstance(self.unet.encoder_hid_proj, ImageProjection) else True
+            image_embeds, negative_image_embeds = self.encode_image(
+                ip_adapter_image, device, num_images_per_prompt, output_hidden_state
+            )
+            if do_classifier_free_guidance:
+                image_embeds = torch.cat([negative_image_embeds, image_embeds])
+
         # 3. Encode input prompt
         prompt_embeds, negative_prompt_embeds = self.encode_prompt(
             prompt,
@@ -655,6 +697,10 @@ class StableDiffusionSAGPipeline(DiffusionPipeline, TextualInversionLoaderMixin)
         # 6. Prepare extra step kwargs. TODO: Logic should ideally just be moved out of the pipeline
         extra_step_kwargs = self.prepare_extra_step_kwargs(generator, eta)
 
+        # 6.1 Add image embeds for IP-Adapter
+        added_cond_kwargs = {"image_embeds": image_embeds} if ip_adapter_image is not None else None
+        added_uncond_kwargs = {"image_embeds": negative_image_embeds} if ip_adapter_image is not None else None
+
         # 7. Denoising loop
         store_processor = CrossAttnStoreProcessor()
         self.unet.mid_block.attentions[0].transformer_blocks[0].attn1.processor = store_processor
@@ -680,6 +726,7 @@ class StableDiffusionSAGPipeline(DiffusionPipeline, TextualInversionLoaderMixin)
                         t,
                         encoder_hidden_states=prompt_embeds,
                         cross_attention_kwargs=cross_attention_kwargs,
+                        added_cond_kwargs=added_cond_kwargs,
                     ).sample
 
                     # perform guidance
@@ -703,7 +750,12 @@ class StableDiffusionSAGPipeline(DiffusionPipeline, TextualInversionLoaderMixin)
                             )
                             uncond_emb, _ = prompt_embeds.chunk(2)
                             # forward and give guidance
-                            degraded_pred = self.unet(degraded_latents, t, encoder_hidden_states=uncond_emb).sample
+                            degraded_pred = self.unet(
+                                degraded_latents,
+                                t,
+                                encoder_hidden_states=uncond_emb,
+                                added_cond_kwargs=added_uncond_kwargs,
+                            ).sample
                             noise_pred += sag_scale * (noise_pred_uncond - degraded_pred)
                         else:
                             # DDIM-like prediction of x0
@@ -715,7 +767,12 @@ class StableDiffusionSAGPipeline(DiffusionPipeline, TextualInversionLoaderMixin)
                                 pred_x0, cond_attn, map_size, t, self.pred_epsilon(latents, noise_pred, t)
                             )
                             # forward and give guidance
-                            degraded_pred = self.unet(degraded_latents, t, encoder_hidden_states=prompt_embeds).sample
+                            degraded_pred = self.unet(
+                                degraded_latents,
+                                t,
+                                encoder_hidden_states=prompt_embeds,
+                                added_cond_kwargs=added_cond_kwargs,
+                            ).sample
                             noise_pred += sag_scale * (noise_pred - degraded_pred)
 
                     # compute the previous noisy sample x_t -> x_t-1

diffusers/pipelines/stable_diffusion_xl/pipeline_stable_diffusion_xl.py

@@ -31,9 +31,10 @@ from ...loaders import (
     StableDiffusionXLLoraLoaderMixin,
     TextualInversionLoaderMixin,
 )
-from ...models import AutoencoderKL, UNet2DConditionModel
+from ...models import AutoencoderKL, ImageProjection, UNet2DConditionModel
 from ...models.attention_processor import (
     AttnProcessor2_0,
+    FusedAttnProcessor2_0,
     LoRAAttnProcessor2_0,
     LoRAXFormersAttnProcessor,
     XFormersAttnProcessor,
@@ -158,12 +159,12 @@ class StableDiffusionXLPipeline(
     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:
-        - *LoRA*: [`loaders.StableDiffusionXLLoraLoaderMixin.load_lora_weights`]
-        - *Ckpt*: [`loaders.FromSingleFileMixin.from_single_file`]
-
-    as well as the following saving methods:
-        - *LoRA*: [`loaders.StableDiffusionXLLoraLoaderMixin.save_lora_weights`]
+    The pipeline also inherits the following loading methods:
+        - [`~loaders.TextualInversionLoaderMixin.load_textual_inversion`] for loading textual inversion embeddings
+        - [`~loaders.FromSingleFileMixin.from_single_file`] for loading `.ckpt` files
+        - [`~loaders.StableDiffusionXLLoraLoaderMixin.load_lora_weights`] for loading LoRA weights
+        - [`~loaders.StableDiffusionXLLoraLoaderMixin.save_lora_weights`] for saving LoRA weights
+        - [`~loaders.IPAdapterMixin.load_ip_adapter`] for loading IP Adapters
 
     Args:
         vae ([`AutoencoderKL`]):
@@ -197,7 +198,7 @@ class StableDiffusionXLPipeline(
             watermarker will be used.
     """
 
-    model_cpu_offload_seq = "text_encoder->text_encoder_2->unet->vae"
+    model_cpu_offload_seq = "text_encoder->text_encoder_2->image_encoder->unet->vae"
     _optional_components = [
         "tokenizer",
         "tokenizer_2",
@@ -524,18 +525,29 @@ class StableDiffusionXLPipeline(
         return prompt_embeds, negative_prompt_embeds, pooled_prompt_embeds, negative_pooled_prompt_embeds
 
     # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.encode_image
-    def encode_image(self, image, device, num_images_per_prompt):
+    def encode_image(self, image, device, num_images_per_prompt, output_hidden_states=None):
         dtype = next(self.image_encoder.parameters()).dtype
 
         if not isinstance(image, torch.Tensor):
             image = self.feature_extractor(image, return_tensors="pt").pixel_values
 
         image = image.to(device=device, dtype=dtype)
-        image_embeds = self.image_encoder(image).image_embeds
-        image_embeds = image_embeds.repeat_interleave(num_images_per_prompt, dim=0)
+        if output_hidden_states:
+            image_enc_hidden_states = self.image_encoder(image, output_hidden_states=True).hidden_states[-2]
+            image_enc_hidden_states = image_enc_hidden_states.repeat_interleave(num_images_per_prompt, dim=0)
+            uncond_image_enc_hidden_states = self.image_encoder(
+                torch.zeros_like(image), output_hidden_states=True
+            ).hidden_states[-2]
+            uncond_image_enc_hidden_states = uncond_image_enc_hidden_states.repeat_interleave(
+                num_images_per_prompt, dim=0
+            )
+            return image_enc_hidden_states, uncond_image_enc_hidden_states
+        else:
+            image_embeds = self.image_encoder(image).image_embeds
+            image_embeds = image_embeds.repeat_interleave(num_images_per_prompt, dim=0)
+            uncond_image_embeds = torch.zeros_like(image_embeds)
 
-        uncond_image_embeds = torch.zeros_like(image_embeds)
-        return image_embeds, uncond_image_embeds
+            return image_embeds, uncond_image_embeds
 
     # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.prepare_extra_step_kwargs
     def prepare_extra_step_kwargs(self, generator, eta):
@@ -670,7 +682,6 @@ class StableDiffusionXLPipeline(
         add_time_ids = torch.tensor([add_time_ids], dtype=dtype)
         return add_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)
@@ -681,6 +692,7 @@ class StableDiffusionXLPipeline(
                 XFormersAttnProcessor,
                 LoRAXFormersAttnProcessor,
                 LoRAAttnProcessor2_0,
+                FusedAttnProcessor2_0,
             ),
         )
         # if xformers or torch_2_0 is used attention block does not need
@@ -718,6 +730,65 @@ class StableDiffusionXLPipeline(
         """Disables the FreeU mechanism if enabled."""
         self.unet.disable_freeu()
 
+    def fuse_qkv_projections(self, unet: bool = True, vae: bool = True):
+        """
+        Enables fused QKV projections. For self-attention modules, all projection matrices (i.e., query,
+        key, value) are fused. For cross-attention modules, key and value projection matrices are fused.
+
+        <Tip warning={true}>
+
+        This API is 🧪 experimental.
+
+        </Tip>
+
+        Args:
+            unet (`bool`, defaults to `True`): To apply fusion on the UNet.
+            vae (`bool`, defaults to `True`): To apply fusion on the VAE.
+        """
+        self.fusing_unet = False
+        self.fusing_vae = False
+
+        if unet:
+            self.fusing_unet = True
+            self.unet.fuse_qkv_projections()
+            self.unet.set_attn_processor(FusedAttnProcessor2_0())
+
+        if vae:
+            if not isinstance(self.vae, AutoencoderKL):
+                raise ValueError("`fuse_qkv_projections()` is only supported for the VAE of type `AutoencoderKL`.")
+
+            self.fusing_vae = True
+            self.vae.fuse_qkv_projections()
+            self.vae.set_attn_processor(FusedAttnProcessor2_0())
+
+    def unfuse_qkv_projections(self, unet: bool = True, vae: bool = True):
+        """Disable QKV projection fusion if enabled.
+
+        <Tip warning={true}>
+
+        This API is 🧪 experimental.
+
+        </Tip>
+
+        Args:
+            unet (`bool`, defaults to `True`): To apply fusion on the UNet.
+            vae (`bool`, defaults to `True`): To apply fusion on the VAE.
+
+        """
+        if unet:
+            if not self.fusing_unet:
+                logger.warning("The UNet was not initially fused for QKV projections. Doing nothing.")
+            else:
+                self.unet.unfuse_qkv_projections()
+                self.fusing_unet = False
+
+        if vae:
+            if not self.fusing_vae:
+                logger.warning("The VAE was not initially fused for QKV projections. Doing nothing.")
+            else:
+                self.vae.unfuse_qkv_projections()
+                self.fusing_vae = False
+
     # Copied from diffusers.pipelines.latent_consistency_models.pipeline_latent_consistency_text2img.LatentConsistencyModelPipeline.get_guidance_scale_embedding
     def get_guidance_scale_embedding(self, w, embedding_dim=512, dtype=torch.float32):
         """
@@ -778,6 +849,10 @@ class StableDiffusionXLPipeline(
     def num_timesteps(self):
         return self._num_timesteps
 
+    @property
+    def interrupt(self):
+        return self._interrupt
+
     @torch.no_grad()
     @replace_example_docstring(EXAMPLE_DOC_STRING)
     def __call__(
@@ -996,6 +1071,7 @@ class StableDiffusionXLPipeline(
         self._clip_skip = clip_skip
         self._cross_attention_kwargs = cross_attention_kwargs
         self._denoising_end = denoising_end
+        self._interrupt = False
 
         # 2. Define call parameters
         if prompt is not None and isinstance(prompt, str):
@@ -1087,7 +1163,10 @@ class StableDiffusionXLPipeline(
         add_time_ids = add_time_ids.to(device).repeat(batch_size * num_images_per_prompt, 1)
 
         if ip_adapter_image is not None:
-            image_embeds, negative_image_embeds = self.encode_image(ip_adapter_image, device, num_images_per_prompt)
+            output_hidden_state = False if isinstance(self.unet.encoder_hid_proj, ImageProjection) else True
+            image_embeds, negative_image_embeds = self.encode_image(
+                ip_adapter_image, device, num_images_per_prompt, output_hidden_state
+            )
             if self.do_classifier_free_guidance:
                 image_embeds = torch.cat([negative_image_embeds, image_embeds])
                 image_embeds = image_embeds.to(device)
@@ -1122,6 +1201,9 @@ class StableDiffusionXLPipeline(
         self._num_timesteps = len(timesteps)
         with self.progress_bar(total=num_inference_steps) as progress_bar:
             for i, t in enumerate(timesteps):
+                if self.interrupt:
+                    continue
+
                 # expand the latents if we are doing classifier free guidance
                 latent_model_input = torch.cat([latents] * 2) if self.do_classifier_free_guidance else latents
 

diffusers/pipelines/stable_diffusion_xl/pipeline_stable_diffusion_xl_img2img.py

@@ -32,9 +32,10 @@ from ...loaders import (
     StableDiffusionXLLoraLoaderMixin,
     TextualInversionLoaderMixin,
 )
-from ...models import AutoencoderKL, UNet2DConditionModel
+from ...models import AutoencoderKL, ImageProjection, UNet2DConditionModel
 from ...models.attention_processor import (
     AttnProcessor2_0,
+    FusedAttnProcessor2_0,
     LoRAAttnProcessor2_0,
     LoRAXFormersAttnProcessor,
     XFormersAttnProcessor,
@@ -176,12 +177,12 @@ class StableDiffusionXLImg2ImgPipeline(
     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:
-        - *LoRA*: [`loaders.StableDiffusionXLLoraLoaderMixin.load_lora_weights`]
-        - *Ckpt*: [`loaders.FromSingleFileMixin.from_single_file`]
-
-    as well as the following saving methods:
-        - *LoRA*: [`loaders.StableDiffusionXLLoraLoaderMixin.save_lora_weights`]
+    The pipeline also inherits the following loading methods:
+        - [`~loaders.TextualInversionLoaderMixin.load_textual_inversion`] for loading textual inversion embeddings
+        - [`~loaders.FromSingleFileMixin.from_single_file`] for loading `.ckpt` files
+        - [`~loaders.StableDiffusionXLLoraLoaderMixin.load_lora_weights`] for loading LoRA weights
+        - [`~loaders.StableDiffusionXLLoraLoaderMixin.save_lora_weights`] for saving LoRA weights
+        - [`~loaders.IPAdapterMixin.load_ip_adapter`] for loading IP Adapters
 
     Args:
         vae ([`AutoencoderKL`]):
@@ -218,7 +219,7 @@ class StableDiffusionXLImg2ImgPipeline(
             watermarker will be used.
     """
 
-    model_cpu_offload_seq = "text_encoder->text_encoder_2->unet->vae"
+    model_cpu_offload_seq = "text_encoder->text_encoder_2->image_encoder->unet->vae"
     _optional_components = [
         "tokenizer",
         "tokenizer_2",
@@ -741,18 +742,29 @@ class StableDiffusionXLImg2ImgPipeline(
         return latents
 
     # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.encode_image
-    def encode_image(self, image, device, num_images_per_prompt):
+    def encode_image(self, image, device, num_images_per_prompt, output_hidden_states=None):
         dtype = next(self.image_encoder.parameters()).dtype
 
         if not isinstance(image, torch.Tensor):
             image = self.feature_extractor(image, return_tensors="pt").pixel_values
 
         image = image.to(device=device, dtype=dtype)
-        image_embeds = self.image_encoder(image).image_embeds
-        image_embeds = image_embeds.repeat_interleave(num_images_per_prompt, dim=0)
+        if output_hidden_states:
+            image_enc_hidden_states = self.image_encoder(image, output_hidden_states=True).hidden_states[-2]
+            image_enc_hidden_states = image_enc_hidden_states.repeat_interleave(num_images_per_prompt, dim=0)
+            uncond_image_enc_hidden_states = self.image_encoder(
+                torch.zeros_like(image), output_hidden_states=True
+            ).hidden_states[-2]
+            uncond_image_enc_hidden_states = uncond_image_enc_hidden_states.repeat_interleave(
+                num_images_per_prompt, dim=0
+            )
+            return image_enc_hidden_states, uncond_image_enc_hidden_states
+        else:
+            image_embeds = self.image_encoder(image).image_embeds
+            image_embeds = image_embeds.repeat_interleave(num_images_per_prompt, dim=0)
+            uncond_image_embeds = torch.zeros_like(image_embeds)
 
-        uncond_image_embeds = torch.zeros_like(image_embeds)
-        return image_embeds, uncond_image_embeds
+            return image_embeds, uncond_image_embeds
 
     def _get_add_time_ids(
         self,
@@ -853,6 +865,67 @@ class StableDiffusionXLImg2ImgPipeline(
         """Disables the FreeU mechanism if enabled."""
         self.unet.disable_freeu()
 
+    # Copied from diffusers.pipelines.stable_diffusion_xl.pipeline_stable_diffusion_xl.StableDiffusionXLPipeline.fuse_qkv_projections
+    def fuse_qkv_projections(self, unet: bool = True, vae: bool = True):
+        """
+        Enables fused QKV projections. For self-attention modules, all projection matrices (i.e., query,
+        key, value) are fused. For cross-attention modules, key and value projection matrices are fused.
+
+        <Tip warning={true}>
+
+        This API is 🧪 experimental.
+
+        </Tip>
+
+        Args:
+            unet (`bool`, defaults to `True`): To apply fusion on the UNet.
+            vae (`bool`, defaults to `True`): To apply fusion on the VAE.
+        """
+        self.fusing_unet = False
+        self.fusing_vae = False
+
+        if unet:
+            self.fusing_unet = True
+            self.unet.fuse_qkv_projections()
+            self.unet.set_attn_processor(FusedAttnProcessor2_0())
+
+        if vae:
+            if not isinstance(self.vae, AutoencoderKL):
+                raise ValueError("`fuse_qkv_projections()` is only supported for the VAE of type `AutoencoderKL`.")
+
+            self.fusing_vae = True
+            self.vae.fuse_qkv_projections()
+            self.vae.set_attn_processor(FusedAttnProcessor2_0())
+
+    # Copied from diffusers.pipelines.stable_diffusion_xl.pipeline_stable_diffusion_xl.StableDiffusionXLPipeline.unfuse_qkv_projections
+    def unfuse_qkv_projections(self, unet: bool = True, vae: bool = True):
+        """Disable QKV projection fusion if enabled.
+
+        <Tip warning={true}>
+
+        This API is 🧪 experimental.
+
+        </Tip>
+
+        Args:
+            unet (`bool`, defaults to `True`): To apply fusion on the UNet.
+            vae (`bool`, defaults to `True`): To apply fusion on the VAE.
+
+        """
+        if unet:
+            if not self.fusing_unet:
+                logger.warning("The UNet was not initially fused for QKV projections. Doing nothing.")
+            else:
+                self.unet.unfuse_qkv_projections()
+                self.fusing_unet = False
+
+        if vae:
+            if not self.fusing_vae:
+                logger.warning("The VAE was not initially fused for QKV projections. Doing nothing.")
+            else:
+                self.vae.unfuse_qkv_projections()
+                self.fusing_vae = False
+
     # Copied from diffusers.pipelines.latent_consistency_models.pipeline_latent_consistency_text2img.LatentConsistencyModelPipeline.get_guidance_scale_embedding
     def get_guidance_scale_embedding(self, w, embedding_dim=512, dtype=torch.float32):
         """
@@ -917,6 +990,10 @@ class StableDiffusionXLImg2ImgPipeline(
     def num_timesteps(self):
         return self._num_timesteps
 
+    @property
+    def interrupt(self):
+        return self._interrupt
+
     @torch.no_grad()
     @replace_example_docstring(EXAMPLE_DOC_STRING)
     def __call__(
@@ -1148,6 +1225,7 @@ class StableDiffusionXLImg2ImgPipeline(
         self._cross_attention_kwargs = cross_attention_kwargs
         self._denoising_end = denoising_end
         self._denoising_start = denoising_start
+        self._interrupt = False
 
         # 2. Define call parameters
         if prompt is not None and isinstance(prompt, str):
@@ -1259,7 +1337,10 @@ class StableDiffusionXLImg2ImgPipeline(
         add_time_ids = add_time_ids.to(device)
 
         if ip_adapter_image is not None:
-            image_embeds, negative_image_embeds = self.encode_image(ip_adapter_image, device, num_images_per_prompt)
+            output_hidden_state = False if isinstance(self.unet.encoder_hid_proj, ImageProjection) else True
+            image_embeds, negative_image_embeds = self.encode_image(
+                ip_adapter_image, device, num_images_per_prompt, output_hidden_state
+            )
             if self.do_classifier_free_guidance:
                 image_embeds = torch.cat([negative_image_embeds, image_embeds])
                 image_embeds = image_embeds.to(device)
@@ -1300,6 +1381,9 @@ class StableDiffusionXLImg2ImgPipeline(
         self._num_timesteps = len(timesteps)
         with self.progress_bar(total=num_inference_steps) as progress_bar:
             for i, t in enumerate(timesteps):
+                if self.interrupt:
+                    continue
+
                 # expand the latents if we are doing classifier free guidance
                 latent_model_input = torch.cat([latents] * 2) if self.do_classifier_free_guidance else latents