diffusers 0.26.3__py3-none-any.whl → 0.27.0__py3-none-any.whl

diffusers/pipelines/stable_diffusion/pipeline_stable_diffusion_img2img.py

@@ -1,4 +1,4 @@
-# Copyright 2023 The HuggingFace Team. All rights reserved.
+# Copyright 2024 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.
@@ -25,7 +25,6 @@ from ...configuration_utils import FrozenDict
 from ...image_processor import PipelineImageInput, VaeImageProcessor
 from ...loaders import FromSingleFileMixin, IPAdapterMixin, LoraLoaderMixin, TextualInversionLoaderMixin
 from ...models import AutoencoderKL, ImageProjection, UNet2DConditionModel
-from ...models.attention_processor import FusedAttnProcessor2_0
 from ...models.lora import adjust_lora_scale_text_encoder
 from ...schedulers import KarrasDiffusionSchedulers
 from ...utils import (
@@ -38,7 +37,7 @@ from ...utils import (
     unscale_lora_layers,
 )
 from ...utils.torch_utils import randn_tensor
-from ..pipeline_utils import DiffusionPipeline
+from ..pipeline_utils import DiffusionPipeline, StableDiffusionMixin
 from . import StableDiffusionPipelineOutput
 from .safety_checker import StableDiffusionSafetyChecker
 
@@ -156,7 +155,12 @@ def retrieve_timesteps(
 
 
 class StableDiffusionImg2ImgPipeline(
-    DiffusionPipeline, TextualInversionLoaderMixin, IPAdapterMixin, LoraLoaderMixin, FromSingleFileMixin
+    DiffusionPipeline,
+    StableDiffusionMixin,
+    TextualInversionLoaderMixin,
+    IPAdapterMixin,
+    LoraLoaderMixin,
+    FromSingleFileMixin,
 ):
     r"""
     Pipeline for text-guided image-to-image generation using Stable Diffusion.
@@ -382,7 +386,7 @@ class StableDiffusionImg2ImgPipeline(
             batch_size = prompt_embeds.shape[0]
 
         if prompt_embeds is None:
-            # textual inversion: procecss multi-vector tokens if necessary
+            # textual inversion: process multi-vector tokens if necessary
             if isinstance(self, TextualInversionLoaderMixin):
                 prompt = self.maybe_convert_prompt(prompt, self.tokenizer)
 
@@ -464,7 +468,7 @@ class StableDiffusionImg2ImgPipeline(
             else:
                 uncond_tokens = negative_prompt
 
-            # textual inversion: procecss multi-vector tokens if necessary
+            # textual inversion: process multi-vector tokens if necessary
             if isinstance(self, TextualInversionLoaderMixin):
                 uncond_tokens = self.maybe_convert_prompt(uncond_tokens, self.tokenizer)
 
@@ -529,31 +533,54 @@ class StableDiffusionImg2ImgPipeline(
             return image_embeds, uncond_image_embeds
 
     # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.prepare_ip_adapter_image_embeds
-    def prepare_ip_adapter_image_embeds(self, ip_adapter_image, device, num_images_per_prompt):
-        if not isinstance(ip_adapter_image, list):
-            ip_adapter_image = [ip_adapter_image]
+    def prepare_ip_adapter_image_embeds(
+        self, ip_adapter_image, ip_adapter_image_embeds, device, num_images_per_prompt, do_classifier_free_guidance
+    ):
+        if ip_adapter_image_embeds is None:
+            if not isinstance(ip_adapter_image, list):
+                ip_adapter_image = [ip_adapter_image]
 
-        if len(ip_adapter_image) != len(self.unet.encoder_hid_proj.image_projection_layers):
-            raise ValueError(
-                f"`ip_adapter_image` must have same length as the number of IP Adapters. Got {len(ip_adapter_image)} images and {len(self.unet.encoder_hid_proj.image_projection_layers)} IP Adapters."
-            )
+            if len(ip_adapter_image) != len(self.unet.encoder_hid_proj.image_projection_layers):
+                raise ValueError(
+                    f"`ip_adapter_image` must have same length as the number of IP Adapters. Got {len(ip_adapter_image)} images and {len(self.unet.encoder_hid_proj.image_projection_layers)} IP Adapters."
+                )
 
-        image_embeds = []
-        for single_ip_adapter_image, image_proj_layer in zip(
-            ip_adapter_image, self.unet.encoder_hid_proj.image_projection_layers
-        ):
-            output_hidden_state = not isinstance(image_proj_layer, ImageProjection)
-            single_image_embeds, single_negative_image_embeds = self.encode_image(
-                single_ip_adapter_image, device, 1, output_hidden_state
-            )
-            single_image_embeds = torch.stack([single_image_embeds] * num_images_per_prompt, dim=0)
-            single_negative_image_embeds = torch.stack([single_negative_image_embeds] * num_images_per_prompt, dim=0)
+            image_embeds = []
+            for single_ip_adapter_image, image_proj_layer in zip(
+                ip_adapter_image, self.unet.encoder_hid_proj.image_projection_layers
+            ):
+                output_hidden_state = not isinstance(image_proj_layer, ImageProjection)
+                single_image_embeds, single_negative_image_embeds = self.encode_image(
+                    single_ip_adapter_image, device, 1, output_hidden_state
+                )
+                single_image_embeds = torch.stack([single_image_embeds] * num_images_per_prompt, dim=0)
+                single_negative_image_embeds = torch.stack(
+                    [single_negative_image_embeds] * num_images_per_prompt, dim=0
+                )
 
-            if self.do_classifier_free_guidance:
-                single_image_embeds = torch.cat([single_negative_image_embeds, single_image_embeds])
-                single_image_embeds = single_image_embeds.to(device)
+                if do_classifier_free_guidance:
+                    single_image_embeds = torch.cat([single_negative_image_embeds, single_image_embeds])
+                    single_image_embeds = single_image_embeds.to(device)
 
-            image_embeds.append(single_image_embeds)
+                image_embeds.append(single_image_embeds)
+        else:
+            repeat_dims = [1]
+            image_embeds = []
+            for single_image_embeds in ip_adapter_image_embeds:
+                if do_classifier_free_guidance:
+                    single_negative_image_embeds, single_image_embeds = single_image_embeds.chunk(2)
+                    single_image_embeds = single_image_embeds.repeat(
+                        num_images_per_prompt, *(repeat_dims * len(single_image_embeds.shape[1:]))
+                    )
+                    single_negative_image_embeds = single_negative_image_embeds.repeat(
+                        num_images_per_prompt, *(repeat_dims * len(single_negative_image_embeds.shape[1:]))
+                    )
+                    single_image_embeds = torch.cat([single_negative_image_embeds, single_image_embeds])
+                else:
+                    single_image_embeds = single_image_embeds.repeat(
+                        num_images_per_prompt, *(repeat_dims * len(single_image_embeds.shape[1:]))
+                    )
+                image_embeds.append(single_image_embeds)
 
         return image_embeds
 
@@ -610,6 +637,8 @@ class StableDiffusionImg2ImgPipeline(
         negative_prompt=None,
         prompt_embeds=None,
         negative_prompt_embeds=None,
+        ip_adapter_image=None,
+        ip_adapter_image_embeds=None,
         callback_on_step_end_tensor_inputs=None,
     ):
         if strength < 0 or strength > 1:
@@ -653,12 +682,29 @@ class StableDiffusionImg2ImgPipeline(
                     f" {negative_prompt_embeds.shape}."
                 )
 
+        if ip_adapter_image is not None and ip_adapter_image_embeds is not None:
+            raise ValueError(
+                "Provide either `ip_adapter_image` or `ip_adapter_image_embeds`. Cannot leave both `ip_adapter_image` and `ip_adapter_image_embeds` defined."
+            )
+
+        if ip_adapter_image_embeds is not None:
+            if not isinstance(ip_adapter_image_embeds, list):
+                raise ValueError(
+                    f"`ip_adapter_image_embeds` has to be of type `list` but is {type(ip_adapter_image_embeds)}"
+                )
+            elif ip_adapter_image_embeds[0].ndim not in [3, 4]:
+                raise ValueError(
+                    f"`ip_adapter_image_embeds` has to be a list of 3D or 4D tensors but is {ip_adapter_image_embeds[0].ndim}D"
+                )
+
     def get_timesteps(self, num_inference_steps, strength, device):
         # get the original timestep using init_timestep
         init_timestep = min(int(num_inference_steps * strength), num_inference_steps)
 
         t_start = max(num_inference_steps - init_timestep, 0)
         timesteps = self.scheduler.timesteps[t_start * self.scheduler.order :]
+        if hasattr(self.scheduler, "set_begin_index"):
+            self.scheduler.set_begin_index(t_start * self.scheduler.order)
 
         return timesteps, num_inference_steps - t_start
 
@@ -720,95 +766,6 @@ class StableDiffusionImg2ImgPipeline(
 
         return latents
 
-    # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.enable_freeu
-    def enable_freeu(self, s1: float, s2: float, b1: float, b2: float):
-        r"""Enables the FreeU mechanism as in https://arxiv.org/abs/2309.11497.
-
-        The suffixes after the scaling factors represent the stages where they are being applied.
-
-        Please refer to the [official repository](https://github.com/ChenyangSi/FreeU) for combinations of the values
-        that are known to work well for different pipelines such as Stable Diffusion v1, v2, and Stable Diffusion XL.
-
-        Args:
-            s1 (`float`):
-                Scaling factor for stage 1 to attenuate the contributions of the skip features. This is done to
-                mitigate "oversmoothing effect" in the enhanced denoising process.
-            s2 (`float`):
-                Scaling factor for stage 2 to attenuate the contributions of the skip features. This is done to
-                mitigate "oversmoothing effect" in the enhanced denoising process.
-            b1 (`float`): Scaling factor for stage 1 to amplify the contributions of backbone features.
-            b2 (`float`): Scaling factor for stage 2 to amplify the contributions of backbone features.
-        """
-        if not hasattr(self, "unet"):
-            raise ValueError("The pipeline must have `unet` for using FreeU.")
-        self.unet.enable_freeu(s1=s1, s2=s2, b1=b1, b2=b2)
-
-    # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.disable_freeu
-    def disable_freeu(self):
-        """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):
         """
@@ -882,6 +839,7 @@ class StableDiffusionImg2ImgPipeline(
         prompt_embeds: Optional[torch.FloatTensor] = None,
         negative_prompt_embeds: Optional[torch.FloatTensor] = None,
         ip_adapter_image: Optional[PipelineImageInput] = None,
+        ip_adapter_image_embeds: Optional[List[torch.FloatTensor]] = None,
         output_type: Optional[str] = "pil",
         return_dict: bool = True,
         cross_attention_kwargs: Optional[Dict[str, Any]] = None,
@@ -936,6 +894,11 @@ class StableDiffusionImg2ImgPipeline(
                 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.
+            ip_adapter_image_embeds (`List[torch.FloatTensor]`, *optional*):
+                Pre-generated image embeddings for IP-Adapter. It should be a list of length same as number of IP-adapters.
+                Each element should be a tensor of shape `(batch_size, num_images, emb_dim)`. It should contain the negative image embedding
+                if `do_classifier_free_guidance` is set to `True`.
+                If not provided, embeddings are computed from the `ip_adapter_image` input argument.
             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`):
@@ -990,6 +953,8 @@ class StableDiffusionImg2ImgPipeline(
             negative_prompt,
             prompt_embeds,
             negative_prompt_embeds,
+            ip_adapter_image,
+            ip_adapter_image_embeds,
             callback_on_step_end_tensor_inputs,
         )
 
@@ -1029,9 +994,13 @@ class StableDiffusionImg2ImgPipeline(
         if self.do_classifier_free_guidance:
             prompt_embeds = torch.cat([negative_prompt_embeds, prompt_embeds])
 
-        if ip_adapter_image is not None:
+        if ip_adapter_image is not None or ip_adapter_image_embeds is not None:
             image_embeds = self.prepare_ip_adapter_image_embeds(
-                ip_adapter_image, device, batch_size * num_images_per_prompt
+                ip_adapter_image,
+                ip_adapter_image_embeds,
+                device,
+                batch_size * num_images_per_prompt,
+                self.do_classifier_free_guidance,
             )
 
         # 4. Preprocess image
@@ -1057,7 +1026,11 @@ class StableDiffusionImg2ImgPipeline(
         extra_step_kwargs = self.prepare_extra_step_kwargs(generator, eta)
 
         # 7.1 Add image embeds for IP-Adapter
-        added_cond_kwargs = {"image_embeds": image_embeds} if ip_adapter_image is not None else None
+        added_cond_kwargs = (
+            {"image_embeds": image_embeds}
+            if ip_adapter_image is not None or ip_adapter_image_embeds is not None
+            else None
+        )
 
         # 7.2 Optionally get Guidance Scale Embedding
         timestep_cond = None

diffusers/pipelines/stable_diffusion/pipeline_stable_diffusion_inpaint.py

@@ -1,4 +1,4 @@
-# Copyright 2023 The HuggingFace Team. All rights reserved.
+# Copyright 2024 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.
@@ -25,12 +25,11 @@ from ...configuration_utils import FrozenDict
 from ...image_processor import PipelineImageInput, VaeImageProcessor
 from ...loaders import FromSingleFileMixin, IPAdapterMixin, LoraLoaderMixin, TextualInversionLoaderMixin
 from ...models import AsymmetricAutoencoderKL, AutoencoderKL, ImageProjection, UNet2DConditionModel
-from ...models.attention_processor import FusedAttnProcessor2_0
 from ...models.lora import adjust_lora_scale_text_encoder
 from ...schedulers import KarrasDiffusionSchedulers
 from ...utils import USE_PEFT_BACKEND, deprecate, logging, scale_lora_layers, unscale_lora_layers
 from ...utils.torch_utils import randn_tensor
-from ..pipeline_utils import DiffusionPipeline
+from ..pipeline_utils import DiffusionPipeline, StableDiffusionMixin
 from . import StableDiffusionPipelineOutput
 from .safety_checker import StableDiffusionSafetyChecker
 
@@ -220,7 +219,12 @@ def retrieve_timesteps(
 
 
 class StableDiffusionInpaintPipeline(
-    DiffusionPipeline, TextualInversionLoaderMixin, IPAdapterMixin, LoraLoaderMixin, FromSingleFileMixin
+    DiffusionPipeline,
+    StableDiffusionMixin,
+    TextualInversionLoaderMixin,
+    IPAdapterMixin,
+    LoraLoaderMixin,
+    FromSingleFileMixin,
 ):
     r"""
     Pipeline for text-guided image inpainting using Stable Diffusion.
@@ -454,7 +458,7 @@ class StableDiffusionInpaintPipeline(
             batch_size = prompt_embeds.shape[0]
 
         if prompt_embeds is None:
-            # textual inversion: procecss multi-vector tokens if necessary
+            # textual inversion: process multi-vector tokens if necessary
             if isinstance(self, TextualInversionLoaderMixin):
                 prompt = self.maybe_convert_prompt(prompt, self.tokenizer)
 
@@ -536,7 +540,7 @@ class StableDiffusionInpaintPipeline(
             else:
                 uncond_tokens = negative_prompt
 
-            # textual inversion: procecss multi-vector tokens if necessary
+            # textual inversion: process multi-vector tokens if necessary
             if isinstance(self, TextualInversionLoaderMixin):
                 uncond_tokens = self.maybe_convert_prompt(uncond_tokens, self.tokenizer)
 
@@ -601,31 +605,54 @@ class StableDiffusionInpaintPipeline(
             return image_embeds, uncond_image_embeds
 
     # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.prepare_ip_adapter_image_embeds
-    def prepare_ip_adapter_image_embeds(self, ip_adapter_image, device, num_images_per_prompt):
-        if not isinstance(ip_adapter_image, list):
-            ip_adapter_image = [ip_adapter_image]
+    def prepare_ip_adapter_image_embeds(
+        self, ip_adapter_image, ip_adapter_image_embeds, device, num_images_per_prompt, do_classifier_free_guidance
+    ):
+        if ip_adapter_image_embeds is None:
+            if not isinstance(ip_adapter_image, list):
+                ip_adapter_image = [ip_adapter_image]
 
-        if len(ip_adapter_image) != len(self.unet.encoder_hid_proj.image_projection_layers):
-            raise ValueError(
-                f"`ip_adapter_image` must have same length as the number of IP Adapters. Got {len(ip_adapter_image)} images and {len(self.unet.encoder_hid_proj.image_projection_layers)} IP Adapters."
-            )
+            if len(ip_adapter_image) != len(self.unet.encoder_hid_proj.image_projection_layers):
+                raise ValueError(
+                    f"`ip_adapter_image` must have same length as the number of IP Adapters. Got {len(ip_adapter_image)} images and {len(self.unet.encoder_hid_proj.image_projection_layers)} IP Adapters."
+                )
 
-        image_embeds = []
-        for single_ip_adapter_image, image_proj_layer in zip(
-            ip_adapter_image, self.unet.encoder_hid_proj.image_projection_layers
-        ):
-            output_hidden_state = not isinstance(image_proj_layer, ImageProjection)
-            single_image_embeds, single_negative_image_embeds = self.encode_image(
-                single_ip_adapter_image, device, 1, output_hidden_state
-            )
-            single_image_embeds = torch.stack([single_image_embeds] * num_images_per_prompt, dim=0)
-            single_negative_image_embeds = torch.stack([single_negative_image_embeds] * num_images_per_prompt, dim=0)
+            image_embeds = []
+            for single_ip_adapter_image, image_proj_layer in zip(
+                ip_adapter_image, self.unet.encoder_hid_proj.image_projection_layers
+            ):
+                output_hidden_state = not isinstance(image_proj_layer, ImageProjection)
+                single_image_embeds, single_negative_image_embeds = self.encode_image(
+                    single_ip_adapter_image, device, 1, output_hidden_state
+                )
+                single_image_embeds = torch.stack([single_image_embeds] * num_images_per_prompt, dim=0)
+                single_negative_image_embeds = torch.stack(
+                    [single_negative_image_embeds] * num_images_per_prompt, dim=0
+                )
 
-            if self.do_classifier_free_guidance:
-                single_image_embeds = torch.cat([single_negative_image_embeds, single_image_embeds])
-                single_image_embeds = single_image_embeds.to(device)
+                if do_classifier_free_guidance:
+                    single_image_embeds = torch.cat([single_negative_image_embeds, single_image_embeds])
+                    single_image_embeds = single_image_embeds.to(device)
 
-            image_embeds.append(single_image_embeds)
+                image_embeds.append(single_image_embeds)
+        else:
+            repeat_dims = [1]
+            image_embeds = []
+            for single_image_embeds in ip_adapter_image_embeds:
+                if do_classifier_free_guidance:
+                    single_negative_image_embeds, single_image_embeds = single_image_embeds.chunk(2)
+                    single_image_embeds = single_image_embeds.repeat(
+                        num_images_per_prompt, *(repeat_dims * len(single_image_embeds.shape[1:]))
+                    )
+                    single_negative_image_embeds = single_negative_image_embeds.repeat(
+                        num_images_per_prompt, *(repeat_dims * len(single_negative_image_embeds.shape[1:]))
+                    )
+                    single_image_embeds = torch.cat([single_negative_image_embeds, single_image_embeds])
+                else:
+                    single_image_embeds = single_image_embeds.repeat(
+                        num_images_per_prompt, *(repeat_dims * len(single_image_embeds.shape[1:]))
+                    )
+                image_embeds.append(single_image_embeds)
 
         return image_embeds
 
@@ -675,6 +702,8 @@ class StableDiffusionInpaintPipeline(
         negative_prompt=None,
         prompt_embeds=None,
         negative_prompt_embeds=None,
+        ip_adapter_image=None,
+        ip_adapter_image_embeds=None,
         callback_on_step_end_tensor_inputs=None,
         padding_mask_crop=None,
     ):
@@ -735,6 +764,21 @@ class StableDiffusionInpaintPipeline(
             if output_type != "pil":
                 raise ValueError(f"The output type should be PIL when inpainting mask crop, but is" f" {output_type}.")
 
+        if ip_adapter_image is not None and ip_adapter_image_embeds is not None:
+            raise ValueError(
+                "Provide either `ip_adapter_image` or `ip_adapter_image_embeds`. Cannot leave both `ip_adapter_image` and `ip_adapter_image_embeds` defined."
+            )
+
+        if ip_adapter_image_embeds is not None:
+            if not isinstance(ip_adapter_image_embeds, list):
+                raise ValueError(
+                    f"`ip_adapter_image_embeds` has to be of type `list` but is {type(ip_adapter_image_embeds)}"
+                )
+            elif ip_adapter_image_embeds[0].ndim not in [3, 4]:
+                raise ValueError(
+                    f"`ip_adapter_image_embeds` has to be a list of 3D or 4D tensors but is {ip_adapter_image_embeds[0].ndim}D"
+                )
+
     def prepare_latents(
         self,
         batch_size,
@@ -859,98 +903,11 @@ class StableDiffusionInpaintPipeline(
 
         t_start = max(num_inference_steps - init_timestep, 0)
         timesteps = self.scheduler.timesteps[t_start * self.scheduler.order :]
+        if hasattr(self.scheduler, "set_begin_index"):
+            self.scheduler.set_begin_index(t_start * self.scheduler.order)
 
         return timesteps, num_inference_steps - t_start
 
-    # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.enable_freeu
-    def enable_freeu(self, s1: float, s2: float, b1: float, b2: float):
-        r"""Enables the FreeU mechanism as in https://arxiv.org/abs/2309.11497.
-
-        The suffixes after the scaling factors represent the stages where they are being applied.
-
-        Please refer to the [official repository](https://github.com/ChenyangSi/FreeU) for combinations of the values
-        that are known to work well for different pipelines such as Stable Diffusion v1, v2, and Stable Diffusion XL.
-
-        Args:
-            s1 (`float`):
-                Scaling factor for stage 1 to attenuate the contributions of the skip features. This is done to
-                mitigate "oversmoothing effect" in the enhanced denoising process.
-            s2 (`float`):
-                Scaling factor for stage 2 to attenuate the contributions of the skip features. This is done to
-                mitigate "oversmoothing effect" in the enhanced denoising process.
-            b1 (`float`): Scaling factor for stage 1 to amplify the contributions of backbone features.
-            b2 (`float`): Scaling factor for stage 2 to amplify the contributions of backbone features.
-        """
-        if not hasattr(self, "unet"):
-            raise ValueError("The pipeline must have `unet` for using FreeU.")
-        self.unet.enable_freeu(s1=s1, s2=s2, b1=b1, b2=b2)
-
-    # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.disable_freeu
-    def disable_freeu(self):
-        """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):
         """
@@ -1029,6 +986,7 @@ class StableDiffusionInpaintPipeline(
         prompt_embeds: Optional[torch.FloatTensor] = None,
         negative_prompt_embeds: Optional[torch.FloatTensor] = None,
         ip_adapter_image: Optional[PipelineImageInput] = None,
+        ip_adapter_image_embeds: Optional[List[torch.FloatTensor]] = None,
         output_type: Optional[str] = "pil",
         return_dict: bool = True,
         cross_attention_kwargs: Optional[Dict[str, Any]] = None,
@@ -1105,6 +1063,11 @@ class StableDiffusionInpaintPipeline(
                 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.
+            ip_adapter_image_embeds (`List[torch.FloatTensor]`, *optional*):
+                Pre-generated image embeddings for IP-Adapter. It should be a list of length same as number of IP-adapters.
+                Each element should be a tensor of shape `(batch_size, num_images, emb_dim)`. It should contain the negative image embedding
+                if `do_classifier_free_guidance` is set to `True`.
+                If not provided, embeddings are computed from the `ip_adapter_image` input argument.
             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`):
@@ -1197,6 +1160,8 @@ class StableDiffusionInpaintPipeline(
             negative_prompt,
             prompt_embeds,
             negative_prompt_embeds,
+            ip_adapter_image,
+            ip_adapter_image_embeds,
             callback_on_step_end_tensor_inputs,
             padding_mask_crop,
         )
@@ -1237,9 +1202,13 @@ class StableDiffusionInpaintPipeline(
         if self.do_classifier_free_guidance:
             prompt_embeds = torch.cat([negative_prompt_embeds, prompt_embeds])
 
-        if ip_adapter_image is not None:
+        if ip_adapter_image is not None or ip_adapter_image_embeds is not None:
             image_embeds = self.prepare_ip_adapter_image_embeds(
-                ip_adapter_image, device, batch_size * num_images_per_prompt
+                ip_adapter_image,
+                ip_adapter_image_embeds,
+                device,
+                batch_size * num_images_per_prompt,
+                self.do_classifier_free_guidance,
             )
 
         # 4. set timesteps
@@ -1343,7 +1312,11 @@ class StableDiffusionInpaintPipeline(
         extra_step_kwargs = self.prepare_extra_step_kwargs(generator, eta)
 
         # 9.1 Add image embeds for IP-Adapter
-        added_cond_kwargs = {"image_embeds": image_embeds} if ip_adapter_image is not None else None
+        added_cond_kwargs = (
+            {"image_embeds": image_embeds}
+            if ip_adapter_image is not None or ip_adapter_image_embeds is not None
+            else None
+        )
 
         # 9.2 Optionally get Guidance Scale Embedding
         timestep_cond = None