diffusers 0.33.1__py3-none-any.whl → 0.34.0__py3-none-any.whl

diffusers/pipelines/hunyuan_video/pipeline_hunyuan_video_image2video.py

@@ -1,4 +1,4 @@
-# Copyright 2024 The HunyuanVideo Team and The HuggingFace Team. All rights reserved.
+# Copyright 2025 The HunyuanVideo Team and 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.
@@ -100,6 +100,50 @@ DEFAULT_PROMPT_TEMPLATE = {
 }
 
 
+def _expand_input_ids_with_image_tokens(
+    text_input_ids,
+    prompt_attention_mask,
+    max_sequence_length,
+    image_token_index,
+    image_emb_len,
+    image_emb_start,
+    image_emb_end,
+    pad_token_id,
+):
+    special_image_token_mask = text_input_ids == image_token_index
+    num_special_image_tokens = torch.sum(special_image_token_mask, dim=-1)
+    batch_indices, non_image_indices = torch.where(text_input_ids != image_token_index)
+
+    max_expanded_length = max_sequence_length + (num_special_image_tokens.max() * (image_emb_len - 1))
+    new_token_positions = torch.cumsum((special_image_token_mask * (image_emb_len - 1) + 1), -1) - 1
+    text_to_overwrite = new_token_positions[batch_indices, non_image_indices]
+
+    expanded_input_ids = torch.full(
+        (text_input_ids.shape[0], max_expanded_length),
+        pad_token_id,
+        dtype=text_input_ids.dtype,
+        device=text_input_ids.device,
+    )
+    expanded_input_ids[batch_indices, text_to_overwrite] = text_input_ids[batch_indices, non_image_indices]
+    expanded_input_ids[batch_indices, image_emb_start:image_emb_end] = image_token_index
+
+    expanded_attention_mask = torch.zeros(
+        (text_input_ids.shape[0], max_expanded_length),
+        dtype=prompt_attention_mask.dtype,
+        device=prompt_attention_mask.device,
+    )
+    attn_batch_indices, attention_indices = torch.where(expanded_input_ids != pad_token_id)
+    expanded_attention_mask[attn_batch_indices, attention_indices] = 1.0
+    expanded_attention_mask = expanded_attention_mask.to(prompt_attention_mask.dtype)
+    position_ids = (expanded_attention_mask.cumsum(-1) - 1).masked_fill_((expanded_attention_mask == 0), 1)
+
+    return {
+        "input_ids": expanded_input_ids,
+        "attention_mask": expanded_attention_mask,
+        "position_ids": position_ids,
+    }
+
+
 # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.retrieve_timesteps
 def retrieve_timesteps(
     scheduler,
@@ -251,6 +295,12 @@ class HunyuanVideoImageToVideoPipeline(DiffusionPipeline, HunyuanVideoLoraLoader
         prompt = [prompt_template["template"].format(p) for p in prompt]
 
         crop_start = prompt_template.get("crop_start", None)
+
+        image_emb_len = prompt_template.get("image_emb_len", 576)
+        image_emb_start = prompt_template.get("image_emb_start", 5)
+        image_emb_end = prompt_template.get("image_emb_end", 581)
+        double_return_token_id = prompt_template.get("double_return_token_id", 271)
+
         if crop_start is None:
             prompt_template_input = self.tokenizer(
                 prompt_template["template"],
@@ -280,19 +330,25 @@ class HunyuanVideoImageToVideoPipeline(DiffusionPipeline, HunyuanVideoLoraLoader
 
         image_embeds = self.image_processor(image, return_tensors="pt").pixel_values.to(device)
 
+        image_token_index = self.text_encoder.config.image_token_index
+        pad_token_id = self.text_encoder.config.pad_token_id
+        expanded_inputs = _expand_input_ids_with_image_tokens(
+            text_input_ids,
+            prompt_attention_mask,
+            max_sequence_length,
+            image_token_index,
+            image_emb_len,
+            image_emb_start,
+            image_emb_end,
+            pad_token_id,
+        )
         prompt_embeds = self.text_encoder(
-            input_ids=text_input_ids,
-            attention_mask=prompt_attention_mask,
+            **expanded_inputs,
             pixel_values=image_embeds,
             output_hidden_states=True,
         ).hidden_states[-(num_hidden_layers_to_skip + 1)]
         prompt_embeds = prompt_embeds.to(dtype=dtype)
 
-        image_emb_len = prompt_template.get("image_emb_len", 576)
-        image_emb_start = prompt_template.get("image_emb_start", 5)
-        image_emb_end = prompt_template.get("image_emb_end", 581)
-        double_return_token_id = prompt_template.get("double_return_token_id", 271)
-
         if crop_start is not None and crop_start > 0:
             text_crop_start = crop_start - 1 + image_emb_len
             batch_indices, last_double_return_token_indices = torch.where(text_input_ids == double_return_token_id)
@@ -655,13 +711,13 @@ class HunyuanVideoImageToVideoPipeline(DiffusionPipeline, HunyuanVideoLoraLoader
             true_cfg_scale (`float`, *optional*, defaults to 1.0):
                 When > 1.0 and a provided `negative_prompt`, enables true classifier-free guidance.
             guidance_scale (`float`, defaults to `1.0`):
-                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. Note that the only available HunyuanVideo model is
-                CFG-distilled, which means that traditional guidance between unconditional and conditional latent is
-                not applied.
+                Guidance scale as defined in [Classifier-Free Diffusion
+                Guidance](https://huggingface.co/papers/2207.12598). `guidance_scale` is defined as `w` of equation 2.
+                of [Imagen Paper](https://huggingface.co/papers/2205.11487). 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. Note that the only available
+                HunyuanVideo model is CFG-distilled, which means that traditional guidance between unconditional and
+                conditional latent is not applied.
             num_videos_per_prompt (`int`, *optional*, defaults to 1):
                 The number of images to generate per prompt.
             generator (`torch.Generator` or `List[torch.Generator]`, *optional*):

diffusers/pipelines/hunyuan_video/pipeline_output.py

@@ -1,5 +1,8 @@
 from dataclasses import dataclass
+from typing import List, Union
 
+import numpy as np
+import PIL.Image
 import torch
 
 from diffusers.utils import BaseOutput
@@ -18,3 +21,19 @@ class HunyuanVideoPipelineOutput(BaseOutput):
     """
 
     frames: torch.Tensor
+
+
+@dataclass
+class HunyuanVideoFramepackPipelineOutput(BaseOutput):
+    r"""
+    Output class for HunyuanVideo pipelines.
+
+    Args:
+        frames (`torch.Tensor`, `np.ndarray`, or List[List[PIL.Image.Image]]):
+            List of video outputs - It can be a nested list of length `batch_size,` with each sub-list containing
+            denoised PIL image sequences of length `num_frames.` It can also be a NumPy array or Torch tensor of shape
+            `(batch_size, num_frames, channels, height, width)`. Or, a list of torch tensors where each tensor
+            corresponds to a latent that decodes to multiple frames.
+    """
+
+    frames: Union[torch.Tensor, np.ndarray, List[List[PIL.Image.Image]], List[torch.Tensor]]

diffusers/pipelines/hunyuandit/pipeline_hunyuandit.py

@@ -1,4 +1,4 @@
-# Copyright 2024 HunyuanDiT Authors and The HuggingFace Team. All rights reserved.
+# Copyright 2025 HunyuanDiT Authors and 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.
@@ -128,7 +128,7 @@ def rescale_noise_cfg(noise_cfg, noise_pred_text, guidance_rescale=0.0):
     r"""
     Rescales `noise_cfg` tensor based on `guidance_rescale` to improve image quality and fix overexposure. Based on
     Section 3.4 from [Common Diffusion Noise Schedules and Sample Steps are
-    Flawed](https://arxiv.org/pdf/2305.08891.pdf).
+    Flawed](https://huggingface.co/papers/2305.08891).
 
     Args:
         noise_cfg (`torch.Tensor`):
@@ -433,7 +433,7 @@ class HunyuanDiTPipeline(DiffusionPipeline):
     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
+        # eta corresponds to η in DDIM paper: https://huggingface.co/papers/2010.02502
         # and should be between [0, 1]
 
         accepts_eta = "eta" in set(inspect.signature(self.scheduler.step).parameters.keys())
@@ -555,7 +555,7 @@ class HunyuanDiTPipeline(DiffusionPipeline):
         return self._guidance_rescale
 
     # 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`
+    # of the Imagen paper: https://huggingface.co/papers/2205.11487 . `guidance_scale = 1`
     # corresponds to doing no classifier free guidance.
     @property
     def do_classifier_free_guidance(self):
@@ -625,8 +625,8 @@ class HunyuanDiTPipeline(DiffusionPipeline):
             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) from the [DDIM](https://arxiv.org/abs/2010.02502) paper. Only applies
-                to the [`~schedulers.DDIMScheduler`], and is ignored in other schedulers.
+                Corresponds to parameter eta (η) from the [DDIM](https://huggingface.co/papers/2010.02502) paper. Only
+                applies to the [`~schedulers.DDIMScheduler`], and is ignored in other schedulers.
             generator (`torch.Generator` or `List[torch.Generator]`, *optional*):
                 A [`torch.Generator`](https://pytorch.org/docs/stable/generated/torch.Generator.html) to make
                 generation deterministic.
@@ -662,7 +662,7 @@ class HunyuanDiTPipeline(DiffusionPipeline):
                 inputs will be passed.
             guidance_rescale (`float`, *optional*, defaults to 0.0):
                 Rescale the 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
+                Schedules and Sample Steps are Flawed](https://huggingface.co/papers/2305.08891). See Section 3.4
             original_size (`Tuple[int, int]`, *optional*, defaults to `(1024, 1024)`):
                 The original size of the image. Used to calculate the time ids.
             target_size (`Tuple[int, int]`, *optional*):
@@ -865,7 +865,7 @@ class HunyuanDiTPipeline(DiffusionPipeline):
                     noise_pred = noise_pred_uncond + guidance_scale * (noise_pred_text - noise_pred_uncond)
 
                 if self.do_classifier_free_guidance and guidance_rescale > 0.0:
-                    # Based on 3.4. in https://arxiv.org/pdf/2305.08891.pdf
+                    # Based on 3.4. in https://huggingface.co/papers/2305.08891
                     noise_pred = rescale_noise_cfg(noise_pred, noise_pred_text, guidance_rescale=guidance_rescale)
 
                 # compute the previous noisy sample x_t -> x_t-1

diffusers/pipelines/i2vgen_xl/pipeline_i2vgen_xl.py

@@ -1,4 +1,4 @@
-# Copyright 2024 Alibaba DAMO-VILAB and The HuggingFace Team. All rights reserved.
+# Copyright 2025 Alibaba DAMO-VILAB and 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.
@@ -33,7 +33,7 @@ from ...utils import (
 )
 from ...utils.torch_utils import randn_tensor
 from ...video_processor import VideoProcessor
-from ..pipeline_utils import DiffusionPipeline, StableDiffusionMixin
+from ..pipeline_utils import DeprecatedPipelineMixin, DiffusionPipeline, StableDiffusionMixin
 
 
 if is_torch_xla_available():
@@ -97,9 +97,11 @@ class I2VGenXLPipelineOutput(BaseOutput):
 
 
 class I2VGenXLPipeline(
+    DeprecatedPipelineMixin,
     DiffusionPipeline,
     StableDiffusionMixin,
 ):
+    _last_supported_version = "0.33.1"
     r"""
     Pipeline for image-to-video generation as proposed in [I2VGenXL](https://i2vgen-xl.github.io/).
 
@@ -151,7 +153,7 @@ class I2VGenXLPipeline(
         return self._guidance_scale
 
     # 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`
+    # of the Imagen paper: https://huggingface.co/papers/2205.11487 . `guidance_scale = 1`
     # corresponds to doing no classifier free guidance.
     @property
     def do_classifier_free_guidance(self):
@@ -384,7 +386,7 @@ class I2VGenXLPipeline(
     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
+        # eta corresponds to η in DDIM paper: https://huggingface.co/papers/2010.02502
         # and should be between [0, 1]
 
         accepts_eta = "eta" in set(inspect.signature(self.scheduler.step).parameters.keys())
@@ -462,7 +464,7 @@ class I2VGenXLPipeline(
         image_latents = image_latents.unsqueeze(2)
 
         # Append a position mask for each subsequent frame
-        # after the intial image latent frame
+        # after the initial image latent frame
         frame_position_mask = []
         for frame_idx in range(num_frames - 1):
             scale = (frame_idx + 1) / (num_frames - 1)
@@ -557,8 +559,8 @@ class I2VGenXLPipeline(
                 The prompt or prompts to guide what to not include in image generation. If not defined, you need to
                 pass `negative_prompt_embeds` instead. Ignored when not using guidance (`guidance_scale < 1`).
             eta (`float`, *optional*):
-                Corresponds to parameter eta (η) from the [DDIM](https://arxiv.org/abs/2010.02502) paper. Only applies
-                to the [`~schedulers.DDIMScheduler`], and is ignored in other schedulers.
+                Corresponds to parameter eta (η) from the [DDIM](https://huggingface.co/papers/2010.02502) paper. Only
+                applies to the [`~schedulers.DDIMScheduler`], and is ignored in other schedulers.
             num_videos_per_prompt (`int`, *optional*):
                 The number of images to generate per prompt.
             decode_chunk_size (`int`, *optional*):
@@ -614,7 +616,7 @@ class I2VGenXLPipeline(
 
         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`
+        # of the Imagen paper: https://huggingface.co/papers/2205.11487 . `guidance_scale = 1`
         # corresponds to doing no classifier free guidance.
         self._guidance_scale = guidance_scale
 

diffusers/pipelines/kandinsky/pipeline_kandinsky.py

@@ -1,4 +1,4 @@
-# Copyright 2024 The HuggingFace Team. All rights reserved.
+# Copyright 2025 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.
@@ -278,11 +278,11 @@ class KandinskyPipeline(DiffusionPipeline):
                 The number of denoising steps. More denoising steps usually lead to a higher quality image at the
                 expense of slower inference.
             guidance_scale (`float`, *optional*, defaults to 4.0):
-                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.
+                Guidance scale as defined in [Classifier-Free Diffusion
+                Guidance](https://huggingface.co/papers/2207.12598). `guidance_scale` is defined as `w` of equation 2.
+                of [Imagen Paper](https://huggingface.co/papers/2205.11487). 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.
             num_images_per_prompt (`int`, *optional*, defaults to 1):
                 The number of images to generate per prompt.
             generator (`torch.Generator` or `List[torch.Generator]`, *optional*):

diffusers/pipelines/kandinsky/pipeline_kandinsky_combined.py

@@ -1,4 +1,4 @@
-# Copyright 2024 The HuggingFace Team. All rights reserved.
+# Copyright 2025 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.
@@ -193,7 +193,7 @@ class KandinskyCombinedPipeline(DiffusionPipeline):
     def enable_xformers_memory_efficient_attention(self, attention_op: Optional[Callable] = None):
         self.decoder_pipe.enable_xformers_memory_efficient_attention(attention_op)
 
-    def enable_sequential_cpu_offload(self, gpu_id: Optional[int] = None, device: Union[torch.device, str] = "cuda"):
+    def enable_sequential_cpu_offload(self, gpu_id: Optional[int] = None, device: Union[torch.device, str] = None):
         r"""
         Offloads all models (`unet`, `text_encoder`, `vae`, and `safety checker` state dicts) to CPU using 🤗
         Accelerate, significantly reducing memory usage. Models are moved to a `torch.device('meta')` and loaded on a
@@ -251,20 +251,20 @@ class KandinskyCombinedPipeline(DiffusionPipeline):
             width (`int`, *optional*, defaults to 512):
                 The width in pixels of the generated image.
             prior_guidance_scale (`float`, *optional*, defaults to 4.0):
-                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.
+                Guidance scale as defined in [Classifier-Free Diffusion
+                Guidance](https://huggingface.co/papers/2207.12598). `guidance_scale` is defined as `w` of equation 2.
+                of [Imagen Paper](https://huggingface.co/papers/2205.11487). 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.
             prior_num_inference_steps (`int`, *optional*, defaults to 100):
                 The number of denoising steps. More denoising steps usually lead to a higher quality image at the
                 expense of slower inference.
             guidance_scale (`float`, *optional*, defaults to 4.0):
-                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.
+                Guidance scale as defined in [Classifier-Free Diffusion
+                Guidance](https://huggingface.co/papers/2207.12598). `guidance_scale` is defined as `w` of equation 2.
+                of [Imagen Paper](https://huggingface.co/papers/2205.11487). 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.
             generator (`torch.Generator` or `List[torch.Generator]`, *optional*):
                 One or a list of [torch generator(s)](https://pytorch.org/docs/stable/generated/torch.Generator.html)
                 to make generation deterministic.
@@ -411,7 +411,7 @@ class KandinskyImg2ImgCombinedPipeline(DiffusionPipeline):
     def enable_xformers_memory_efficient_attention(self, attention_op: Optional[Callable] = None):
         self.decoder_pipe.enable_xformers_memory_efficient_attention(attention_op)
 
-    def enable_sequential_cpu_offload(self, gpu_id: Optional[int] = None, device: Union[torch.device, str] = "cuda"):
+    def enable_sequential_cpu_offload(self, gpu_id: Optional[int] = None, device: Union[torch.device, str] = None):
         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
@@ -482,20 +482,20 @@ class KandinskyImg2ImgCombinedPipeline(DiffusionPipeline):
                 be maximum and the denoising process will run for the full number of iterations specified in
                 `num_inference_steps`. A value of 1, therefore, essentially ignores `image`.
             prior_guidance_scale (`float`, *optional*, defaults to 4.0):
-                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.
+                Guidance scale as defined in [Classifier-Free Diffusion
+                Guidance](https://huggingface.co/papers/2207.12598). `guidance_scale` is defined as `w` of equation 2.
+                of [Imagen Paper](https://huggingface.co/papers/2205.11487). 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.
             prior_num_inference_steps (`int`, *optional*, defaults to 100):
                 The number of denoising steps. More denoising steps usually lead to a higher quality image at the
                 expense of slower inference.
             guidance_scale (`float`, *optional*, defaults to 4.0):
-                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.
+                Guidance scale as defined in [Classifier-Free Diffusion
+                Guidance](https://huggingface.co/papers/2207.12598). `guidance_scale` is defined as `w` of equation 2.
+                of [Imagen Paper](https://huggingface.co/papers/2205.11487). 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.
             generator (`torch.Generator` or `List[torch.Generator]`, *optional*):
                 One or a list of [torch generator(s)](https://pytorch.org/docs/stable/generated/torch.Generator.html)
                 to make generation deterministic.
@@ -652,7 +652,7 @@ class KandinskyInpaintCombinedPipeline(DiffusionPipeline):
     def enable_xformers_memory_efficient_attention(self, attention_op: Optional[Callable] = None):
         self.decoder_pipe.enable_xformers_memory_efficient_attention(attention_op)
 
-    def enable_sequential_cpu_offload(self, gpu_id: Optional[int] = None, device: Union[torch.device, str] = "cuda"):
+    def enable_sequential_cpu_offload(self, gpu_id: Optional[int] = None, device: Union[torch.device, str] = None):
         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
@@ -722,20 +722,20 @@ class KandinskyInpaintCombinedPipeline(DiffusionPipeline):
             width (`int`, *optional*, defaults to 512):
                 The width in pixels of the generated image.
             prior_guidance_scale (`float`, *optional*, defaults to 4.0):
-                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.
+                Guidance scale as defined in [Classifier-Free Diffusion
+                Guidance](https://huggingface.co/papers/2207.12598). `guidance_scale` is defined as `w` of equation 2.
+                of [Imagen Paper](https://huggingface.co/papers/2205.11487). 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.
             prior_num_inference_steps (`int`, *optional*, defaults to 100):
                 The number of denoising steps. More denoising steps usually lead to a higher quality image at the
                 expense of slower inference.
             guidance_scale (`float`, *optional*, defaults to 4.0):
-                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.
+                Guidance scale as defined in [Classifier-Free Diffusion
+                Guidance](https://huggingface.co/papers/2207.12598). `guidance_scale` is defined as `w` of equation 2.
+                of [Imagen Paper](https://huggingface.co/papers/2205.11487). 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.
             generator (`torch.Generator` or `List[torch.Generator]`, *optional*):
                 One or a list of [torch generator(s)](https://pytorch.org/docs/stable/generated/torch.Generator.html)
                 to make generation deterministic.

diffusers/pipelines/kandinsky/pipeline_kandinsky_img2img.py

@@ -1,4 +1,4 @@
-# Copyright 2024 The HuggingFace Team. All rights reserved.
+# Copyright 2025 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.
@@ -14,14 +14,13 @@
 
 from typing import Callable, List, Optional, Union
 
-import numpy as np
 import PIL.Image
 import torch
-from PIL import Image
 from transformers import (
     XLMRobertaTokenizer,
 )
 
+from ...image_processor import VaeImageProcessor
 from ...models import UNet2DConditionModel, VQModel
 from ...schedulers import DDIMScheduler
 from ...utils import (
@@ -95,15 +94,6 @@ def get_new_h_w(h, w, scale_factor=8):
     return new_h * scale_factor, new_w * scale_factor
 
 
-def prepare_image(pil_image, w=512, h=512):
-    pil_image = pil_image.resize((w, h), resample=Image.BICUBIC, reducing_gap=1)
-    arr = np.array(pil_image.convert("RGB"))
-    arr = arr.astype(np.float32) / 127.5 - 1
-    arr = np.transpose(arr, [2, 0, 1])
-    image = torch.from_numpy(arr).unsqueeze(0)
-    return image
-
-
 class KandinskyImg2ImgPipeline(DiffusionPipeline):
     """
     Pipeline for image-to-image generation using Kandinsky
@@ -143,7 +133,16 @@ class KandinskyImg2ImgPipeline(DiffusionPipeline):
             scheduler=scheduler,
             movq=movq,
         )
-        self.movq_scale_factor = 2 ** (len(self.movq.config.block_out_channels) - 1)
+        self.movq_scale_factor = (
+            2 ** (len(self.movq.config.block_out_channels) - 1) if getattr(self, "movq", None) else 8
+        )
+        movq_latent_channels = self.movq.config.latent_channels if getattr(self, "movq", None) else 4
+        self.image_processor = VaeImageProcessor(
+            vae_scale_factor=self.movq_scale_factor,
+            vae_latent_channels=movq_latent_channels,
+            resample="bicubic",
+            reducing_gap=1,
+        )
 
     def get_timesteps(self, num_inference_steps, strength, device):
         # get the original timestep using init_timestep
@@ -350,11 +349,11 @@ class KandinskyImg2ImgPipeline(DiffusionPipeline):
                 be maximum and the denoising process will run for the full number of iterations specified in
                 `num_inference_steps`. A value of 1, therefore, essentially ignores `image`.
             guidance_scale (`float`, *optional*, defaults to 4.0):
-                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.
+                Guidance scale as defined in [Classifier-Free Diffusion
+                Guidance](https://huggingface.co/papers/2207.12598). `guidance_scale` is defined as `w` of equation 2.
+                of [Imagen Paper](https://huggingface.co/papers/2205.11487). 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.
             num_images_per_prompt (`int`, *optional*, defaults to 1):
                 The number of images to generate per prompt.
             generator (`torch.Generator` or `List[torch.Generator]`, *optional*):
@@ -417,7 +416,7 @@ class KandinskyImg2ImgPipeline(DiffusionPipeline):
                 f"Input is in incorrect format: {[type(i) for i in image]}. Currently, we only support  PIL image and pytorch tensor"
             )
 
-        image = torch.cat([prepare_image(i, width, height) for i in image], dim=0)
+        image = torch.cat([self.image_processor.preprocess(i, width, height) for i in image], dim=0)
         image = image.to(dtype=prompt_embeds.dtype, device=device)
 
         latents = self.movq.encode(image)["latents"]
@@ -498,13 +497,7 @@ class KandinskyImg2ImgPipeline(DiffusionPipeline):
         if output_type not in ["pt", "np", "pil"]:
             raise ValueError(f"Only the output types `pt`, `pil` and `np` are supported not output_type={output_type}")
 
-        if output_type in ["np", "pil"]:
-            image = image * 0.5 + 0.5
-            image = image.clamp(0, 1)
-            image = image.cpu().permute(0, 2, 3, 1).float().numpy()
-
-        if output_type == "pil":
-            image = self.numpy_to_pil(image)
+        image = self.image_processor.postprocess(image, output_type)
 
         if not return_dict:
             return (image,)

diffusers/pipelines/kandinsky/pipeline_kandinsky_inpaint.py

@@ -1,4 +1,4 @@
-# Copyright 2024 The HuggingFace Team. All rights reserved.
+# Copyright 2025 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.
@@ -456,11 +456,11 @@ class KandinskyInpaintPipeline(DiffusionPipeline):
                 The number of denoising steps. More denoising steps usually lead to a higher quality image at the
                 expense of slower inference.
             guidance_scale (`float`, *optional*, defaults to 4.0):
-                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.
+                Guidance scale as defined in [Classifier-Free Diffusion
+                Guidance](https://huggingface.co/papers/2207.12598). `guidance_scale` is defined as `w` of equation 2.
+                of [Imagen Paper](https://huggingface.co/papers/2205.11487). 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.
             num_images_per_prompt (`int`, *optional*, defaults to 1):
                 The number of images to generate per prompt.
             generator (`torch.Generator` or `List[torch.Generator]`, *optional*):
@@ -496,7 +496,7 @@ class KandinskyInpaintPipeline(DiffusionPipeline):
                 "As of diffusers==0.19.0 this behavior has been inverted. Now white pixels are repainted and black pixels are preserved. "
                 "This way, Kandinsky's masking behavior is aligned with Stable Diffusion. "
                 "THIS means that you HAVE to invert the input mask to have the same behavior as before as explained in https://github.com/huggingface/diffusers/pull/4207. "
-                "This warning will be surpressed after the first inference call and will be removed in diffusers>0.23.0"
+                "This warning will be suppressed after the first inference call and will be removed in diffusers>0.23.0"
             )
             self._warn_has_been_called = True
 

diffusers/pipelines/kandinsky/pipeline_kandinsky_prior.py

@@ -1,4 +1,4 @@
-# Copyright 2024 The HuggingFace Team. All rights reserved.
+# Copyright 2025 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.
@@ -220,11 +220,11 @@ class KandinskyPriorPipeline(DiffusionPipeline):
                 The prompt not to guide the image generation. Ignored when not using guidance (i.e., ignored if
                 `guidance_scale` is less than `1`).
             guidance_scale (`float`, *optional*, defaults to 4.0):
-                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.
+                Guidance scale as defined in [Classifier-Free Diffusion
+                Guidance](https://huggingface.co/papers/2207.12598). `guidance_scale` is defined as `w` of equation 2.
+                of [Imagen Paper](https://huggingface.co/papers/2205.11487). 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.
 
         Examples:
 
@@ -439,11 +439,11 @@ class KandinskyPriorPipeline(DiffusionPipeline):
                 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`.
             guidance_scale (`float`, *optional*, defaults to 4.0):
-                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.
+                Guidance scale as defined in [Classifier-Free Diffusion
+                Guidance](https://huggingface.co/papers/2207.12598). `guidance_scale` is defined as `w` of equation 2.
+                of [Imagen Paper](https://huggingface.co/papers/2205.11487). 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.
             output_type (`str`, *optional*, defaults to `"pt"`):
                 The output format of the generate image. Choose between: `"np"` (`np.array`) or `"pt"`
                 (`torch.Tensor`).

diffusers/pipelines/kandinsky2_2/pipeline_kandinsky2_2.py

@@ -1,4 +1,4 @@
-# Copyright 2024 The HuggingFace Team. All rights reserved.
+# Copyright 2025 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.
@@ -162,11 +162,11 @@ class KandinskyV22Pipeline(DiffusionPipeline):
                 The number of denoising steps. More denoising steps usually lead to a higher quality image at the
                 expense of slower inference.
             guidance_scale (`float`, *optional*, defaults to 4.0):
-                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.
+                Guidance scale as defined in [Classifier-Free Diffusion
+                Guidance](https://huggingface.co/papers/2207.12598). `guidance_scale` is defined as `w` of equation 2.
+                of [Imagen Paper](https://huggingface.co/papers/2205.11487). 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.
             num_images_per_prompt (`int`, *optional*, defaults to 1):
                 The number of images to generate per prompt.
             generator (`torch.Generator` or `List[torch.Generator]`, *optional*):