diffusers 0.34.0__py3-none-any.whl → 0.35.1__py3-none-any.whl

diffusers/modular_pipelines/wan/modular_pipeline.py

@@ -0,0 +1,90 @@
+# 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.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+
+from ...loaders import WanLoraLoaderMixin
+from ...pipelines.pipeline_utils import StableDiffusionMixin
+from ...utils import logging
+from ..modular_pipeline import ModularPipeline
+
+
+logger = logging.get_logger(__name__)  # pylint: disable=invalid-name
+
+
+class WanModularPipeline(
+    ModularPipeline,
+    StableDiffusionMixin,
+    WanLoraLoaderMixin,
+):
+    """
+    A ModularPipeline for Wan.
+
+    <Tip warning={true}>
+
+        This is an experimental feature and is likely to change in the future.
+
+    </Tip>
+    """
+
+    @property
+    def default_height(self):
+        return self.default_sample_height * self.vae_scale_factor_spatial
+
+    @property
+    def default_width(self):
+        return self.default_sample_width * self.vae_scale_factor_spatial
+
+    @property
+    def default_num_frames(self):
+        return (self.default_sample_num_frames - 1) * self.vae_scale_factor_temporal + 1
+
+    @property
+    def default_sample_height(self):
+        return 60
+
+    @property
+    def default_sample_width(self):
+        return 104
+
+    @property
+    def default_sample_num_frames(self):
+        return 21
+
+    @property
+    def vae_scale_factor_spatial(self):
+        vae_scale_factor = 8
+        if hasattr(self, "vae") and self.vae is not None:
+            vae_scale_factor = 2 ** len(self.vae.temperal_downsample)
+        return vae_scale_factor
+
+    @property
+    def vae_scale_factor_temporal(self):
+        vae_scale_factor = 4
+        if hasattr(self, "vae") and self.vae is not None:
+            vae_scale_factor = 2 ** sum(self.vae.temperal_downsample)
+        return vae_scale_factor
+
+    @property
+    def num_channels_transformer(self):
+        num_channels_transformer = 16
+        if hasattr(self, "transformer") and self.transformer is not None:
+            num_channels_transformer = self.transformer.config.in_channels
+        return num_channels_transformer
+
+    @property
+    def num_channels_latents(self):
+        num_channels_latents = 16
+        if hasattr(self, "vae") and self.vae is not None:
+            num_channels_latents = self.vae.config.z_dim
+        return num_channels_latents

diffusers/pipelines/__init__.py

@@ -140,6 +140,8 @@ else:
         "FluxFillPipeline",
         "FluxPriorReduxPipeline",
         "ReduxImageEncoder",
+        "FluxKontextPipeline",
+        "FluxKontextInpaintPipeline",
     ]
     _import_structure["audioldm"] = ["AudioLDMPipeline"]
     _import_structure["audioldm2"] = [
@@ -378,6 +380,19 @@ else:
         "WuerstchenPriorPipeline",
     ]
     _import_structure["wan"] = ["WanPipeline", "WanImageToVideoPipeline", "WanVideoToVideoPipeline", "WanVACEPipeline"]
+    _import_structure["skyreels_v2"] = [
+        "SkyReelsV2DiffusionForcingPipeline",
+        "SkyReelsV2DiffusionForcingImageToVideoPipeline",
+        "SkyReelsV2DiffusionForcingVideoToVideoPipeline",
+        "SkyReelsV2ImageToVideoPipeline",
+        "SkyReelsV2Pipeline",
+    ]
+    _import_structure["qwenimage"] = [
+        "QwenImagePipeline",
+        "QwenImageImg2ImgPipeline",
+        "QwenImageInpaintPipeline",
+        "QwenImageEditPipeline",
+    ]
 try:
     if not is_onnx_available():
         raise OptionalDependencyNotAvailable()
@@ -609,6 +624,8 @@ if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT:
             FluxFillPipeline,
             FluxImg2ImgPipeline,
             FluxInpaintPipeline,
+            FluxKontextInpaintPipeline,
+            FluxKontextPipeline,
             FluxPipeline,
             FluxPriorReduxPipeline,
             ReduxImageEncoder,
@@ -692,6 +709,12 @@ if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT:
         from .paint_by_example import PaintByExamplePipeline
         from .pia import PIAPipeline
         from .pixart_alpha import PixArtAlphaPipeline, PixArtSigmaPipeline
+        from .qwenimage import (
+            QwenImageEditPipeline,
+            QwenImageImg2ImgPipeline,
+            QwenImageInpaintPipeline,
+            QwenImagePipeline,
+        )
         from .sana import SanaControlNetPipeline, SanaPipeline, SanaSprintImg2ImgPipeline, SanaSprintPipeline
         from .semantic_stable_diffusion import SemanticStableDiffusionPipeline
         from .shap_e import ShapEImg2ImgPipeline, ShapEPipeline
@@ -847,6 +870,14 @@ if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT:
                 SpectrogramDiffusionPipeline,
             )
 
+        from .skyreels_v2 import (
+            SkyReelsV2DiffusionForcingImageToVideoPipeline,
+            SkyReelsV2DiffusionForcingPipeline,
+            SkyReelsV2DiffusionForcingVideoToVideoPipeline,
+            SkyReelsV2ImageToVideoPipeline,
+            SkyReelsV2Pipeline,
+        )
+
 else:
     import sys
 

diffusers/pipelines/audioldm2/pipeline_audioldm2.py

@@ -312,15 +312,14 @@ class AudioLDM2Pipeline(DiffusionPipeline):
                 The sequence of generated hidden-states.
         """
         cache_position_kwargs = {}
-        if is_transformers_version("<", "4.52.0.dev0"):
+        if is_transformers_version("<", "4.52.1"):
             cache_position_kwargs["input_ids"] = inputs_embeds
-            cache_position_kwargs["model_kwargs"] = model_kwargs
         else:
             cache_position_kwargs["seq_length"] = inputs_embeds.shape[0]
             cache_position_kwargs["device"] = (
                 self.language_model.device if getattr(self, "language_model", None) is not None else self.device
             )
-            cache_position_kwargs["model_kwargs"] = model_kwargs
+        cache_position_kwargs["model_kwargs"] = model_kwargs
         max_new_tokens = max_new_tokens if max_new_tokens is not None else self.language_model.config.max_new_tokens
         model_kwargs = self.language_model._get_initial_cache_position(**cache_position_kwargs)
 

diffusers/pipelines/auto_pipeline.py

@@ -49,6 +49,7 @@ from .flux import (
     FluxControlPipeline,
     FluxImg2ImgPipeline,
     FluxInpaintPipeline,
+    FluxKontextPipeline,
     FluxPipeline,
 )
 from .hunyuandit import HunyuanDiTPipeline
@@ -142,6 +143,7 @@ AUTO_TEXT2IMAGE_PIPELINES_MAPPING = OrderedDict(
         ("flux", FluxPipeline),
         ("flux-control", FluxControlPipeline),
         ("flux-controlnet", FluxControlNetPipeline),
+        ("flux-kontext", FluxKontextPipeline),
         ("lumina", LuminaPipeline),
         ("lumina2", Lumina2Pipeline),
         ("chroma", ChromaPipeline),
@@ -171,6 +173,7 @@ AUTO_IMAGE2IMAGE_PIPELINES_MAPPING = OrderedDict(
         ("flux", FluxImg2ImgPipeline),
         ("flux-controlnet", FluxControlNetImg2ImgPipeline),
         ("flux-control", FluxControlImg2ImgPipeline),
+        ("flux-kontext", FluxKontextPipeline),
     ]
 )
 
@@ -248,14 +251,15 @@ def _get_connected_pipeline(pipeline_cls):
         return _get_task_class(AUTO_INPAINT_PIPELINES_MAPPING, pipeline_cls.__name__, throw_error_if_not_exist=False)
 
 
-def _get_task_class(mapping, pipeline_class_name, throw_error_if_not_exist: bool = True):
-    def get_model(pipeline_class_name):
-        for task_mapping in SUPPORTED_TASKS_MAPPINGS:
-            for model_name, pipeline in task_mapping.items():
-                if pipeline.__name__ == pipeline_class_name:
-                    return model_name
+def _get_model(pipeline_class_name):
+    for task_mapping in SUPPORTED_TASKS_MAPPINGS:
+        for model_name, pipeline in task_mapping.items():
+            if pipeline.__name__ == pipeline_class_name:
+                return model_name
+
 
-    model_name = get_model(pipeline_class_name)
+def _get_task_class(mapping, pipeline_class_name, throw_error_if_not_exist: bool = True):
+    model_name = _get_model(pipeline_class_name)
 
     if model_name is not None:
         task_class = mapping.get(model_name, None)
@@ -391,8 +395,8 @@ class AutoPipelineForText2Image(ConfigMixin):
 
         <Tip>
 
-        To use private or [gated](https://huggingface.co/docs/hub/models-gated#gated-models) models, log-in with
-        `huggingface-cli login`.
+        To use private or [gated](https://huggingface.co/docs/hub/models-gated#gated-models) models, log-in with `hf
+        auth login`.
 
         </Tip>
 
@@ -686,8 +690,8 @@ class AutoPipelineForImage2Image(ConfigMixin):
 
         <Tip>
 
-        To use private or [gated](https://huggingface.co/docs/hub/models-gated#gated-models) models, log-in with
-        `huggingface-cli login`.
+        To use private or [gated](https://huggingface.co/docs/hub/models-gated#gated-models) models, log-in with `hf
+        auth login`.
 
         </Tip>
 
@@ -996,8 +1000,8 @@ class AutoPipelineForInpainting(ConfigMixin):
 
         <Tip>
 
-        To use private or [gated](https://huggingface.co/docs/hub/models-gated#gated-models) models, log-in with
-        `huggingface-cli login`.
+        To use private or [gated](https://huggingface.co/docs/hub/models-gated#gated-models) models, log-in with `hf
+        auth login`.
 
         </Tip>
 

diffusers/pipelines/chroma/pipeline_chroma.py

@@ -663,11 +663,11 @@ class ChromaPipeline(
                 their `set_timesteps` method. If not defined, the default behavior when `num_inference_steps` is passed
                 will be used.
             guidance_scale (`float`, *optional*, defaults to 3.5):
-                Guidance scale as defined in [Classifier-Free Diffusion Guidance](https://arxiv.org/abs/2207.12598).
-                `guidance_scale` is defined as `w` of equation 2. of [Imagen
-                Paper](https://arxiv.org/pdf/2205.11487.pdf). Guidance scale is enabled by setting `guidance_scale >
-                1`. Higher guidance scale encourages to generate images that are closely linked to the text `prompt`,
-                usually at the expense of lower image quality.
+                Embedded guiddance scale is enabled by setting `guidance_scale` > 1. Higher `guidance_scale` encourages
+                a model to generate images more aligned with `prompt` at the expense of lower image quality.
+
+                Guidance-distilled models approximates true classifer-free guidance for `guidance_scale` > 1. Refer to
+                the [paper](https://huggingface.co/papers/2210.03142) to learn more.
             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/chroma/pipeline_chroma_img2img.py

@@ -725,11 +725,11 @@ class ChromaImg2ImgPipeline(
                 their `set_timesteps` method. If not defined, the default behavior when `num_inference_steps` is passed
                 will be used.
             guidance_scale (`float`, *optional*, defaults to 5.0):
-                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.
+                Embedded guiddance scale is enabled by setting `guidance_scale` > 1. Higher `guidance_scale` encourages
+                a model to generate images more aligned with `prompt` at the expense of lower image quality.
+
+                Guidance-distilled models approximates true classifer-free guidance for `guidance_scale` > 1. Refer to
+                the [paper](https://huggingface.co/papers/2210.03142) to learn more.
             strength (`float, *optional*, defaults to 0.9):
                 Conceptually, indicates how much to transform the reference image. Must be between 0 and 1. image will
                 be used as a starting point, adding more noise to it the larger the strength. The number of denoising

diffusers/pipelines/cogvideo/pipeline_cogvideox.py

@@ -718,14 +718,15 @@ class CogVideoXPipeline(DiffusionPipeline, CogVideoXLoraLoaderMixin):
                 timestep = t.expand(latent_model_input.shape[0])
 
                 # predict noise model_output
-                noise_pred = self.transformer(
-                    hidden_states=latent_model_input,
-                    encoder_hidden_states=prompt_embeds,
-                    timestep=timestep,
-                    image_rotary_emb=image_rotary_emb,
-                    attention_kwargs=attention_kwargs,
-                    return_dict=False,
-                )[0]
+                with self.transformer.cache_context("cond_uncond"):
+                    noise_pred = self.transformer(
+                        hidden_states=latent_model_input,
+                        encoder_hidden_states=prompt_embeds,
+                        timestep=timestep,
+                        image_rotary_emb=image_rotary_emb,
+                        attention_kwargs=attention_kwargs,
+                        return_dict=False,
+                    )[0]
                 noise_pred = noise_pred.float()
 
                 # perform guidance

diffusers/pipelines/cogvideo/pipeline_cogvideox_fun_control.py

@@ -784,14 +784,15 @@ class CogVideoXFunControlPipeline(DiffusionPipeline, CogVideoXLoraLoaderMixin):
                 timestep = t.expand(latent_model_input.shape[0])
 
                 # predict noise model_output
-                noise_pred = self.transformer(
-                    hidden_states=latent_model_input,
-                    encoder_hidden_states=prompt_embeds,
-                    timestep=timestep,
-                    image_rotary_emb=image_rotary_emb,
-                    attention_kwargs=attention_kwargs,
-                    return_dict=False,
-                )[0]
+                with self.transformer.cache_context("cond_uncond"):
+                    noise_pred = self.transformer(
+                        hidden_states=latent_model_input,
+                        encoder_hidden_states=prompt_embeds,
+                        timestep=timestep,
+                        image_rotary_emb=image_rotary_emb,
+                        attention_kwargs=attention_kwargs,
+                        return_dict=False,
+                    )[0]
                 noise_pred = noise_pred.float()
 
                 # perform guidance

diffusers/pipelines/cogvideo/pipeline_cogvideox_image2video.py

@@ -831,15 +831,16 @@ class CogVideoXImageToVideoPipeline(DiffusionPipeline, CogVideoXLoraLoaderMixin)
                 timestep = t.expand(latent_model_input.shape[0])
 
                 # predict noise model_output
-                noise_pred = self.transformer(
-                    hidden_states=latent_model_input,
-                    encoder_hidden_states=prompt_embeds,
-                    timestep=timestep,
-                    ofs=ofs_emb,
-                    image_rotary_emb=image_rotary_emb,
-                    attention_kwargs=attention_kwargs,
-                    return_dict=False,
-                )[0]
+                with self.transformer.cache_context("cond_uncond"):
+                    noise_pred = self.transformer(
+                        hidden_states=latent_model_input,
+                        encoder_hidden_states=prompt_embeds,
+                        timestep=timestep,
+                        ofs=ofs_emb,
+                        image_rotary_emb=image_rotary_emb,
+                        attention_kwargs=attention_kwargs,
+                        return_dict=False,
+                    )[0]
                 noise_pred = noise_pred.float()
 
                 # perform guidance

diffusers/pipelines/cogvideo/pipeline_cogvideox_video2video.py

@@ -799,14 +799,15 @@ class CogVideoXVideoToVideoPipeline(DiffusionPipeline, CogVideoXLoraLoaderMixin)
                 timestep = t.expand(latent_model_input.shape[0])
 
                 # predict noise model_output
-                noise_pred = self.transformer(
-                    hidden_states=latent_model_input,
-                    encoder_hidden_states=prompt_embeds,
-                    timestep=timestep,
-                    image_rotary_emb=image_rotary_emb,
-                    attention_kwargs=attention_kwargs,
-                    return_dict=False,
-                )[0]
+                with self.transformer.cache_context("cond_uncond"):
+                    noise_pred = self.transformer(
+                        hidden_states=latent_model_input,
+                        encoder_hidden_states=prompt_embeds,
+                        timestep=timestep,
+                        image_rotary_emb=image_rotary_emb,
+                        attention_kwargs=attention_kwargs,
+                        return_dict=False,
+                    )[0]
                 noise_pred = noise_pred.float()
 
                 # perform guidance

diffusers/pipelines/cogview4/pipeline_cogview4.py

@@ -619,22 +619,10 @@ class CogView4Pipeline(DiffusionPipeline, CogView4LoraLoaderMixin):
                 # broadcast to batch dimension in a way that's compatible with ONNX/Core ML
                 timestep = t.expand(latents.shape[0])
 
-                noise_pred_cond = self.transformer(
-                    hidden_states=latent_model_input,
-                    encoder_hidden_states=prompt_embeds,
-                    timestep=timestep,
-                    original_size=original_size,
-                    target_size=target_size,
-                    crop_coords=crops_coords_top_left,
-                    attention_kwargs=attention_kwargs,
-                    return_dict=False,
-                )[0]
-
-                # perform guidance
-                if self.do_classifier_free_guidance:
-                    noise_pred_uncond = self.transformer(
+                with self.transformer.cache_context("cond"):
+                    noise_pred_cond = self.transformer(
                         hidden_states=latent_model_input,
-                        encoder_hidden_states=negative_prompt_embeds,
+                        encoder_hidden_states=prompt_embeds,
                         timestep=timestep,
                         original_size=original_size,
                         target_size=target_size,
@@ -643,6 +631,19 @@ class CogView4Pipeline(DiffusionPipeline, CogView4LoraLoaderMixin):
                         return_dict=False,
                     )[0]
 
+                # perform guidance
+                if self.do_classifier_free_guidance:
+                    with self.transformer.cache_context("uncond"):
+                        noise_pred_uncond = self.transformer(
+                            hidden_states=latent_model_input,
+                            encoder_hidden_states=negative_prompt_embeds,
+                            timestep=timestep,
+                            original_size=original_size,
+                            target_size=target_size,
+                            crop_coords=crops_coords_top_left,
+                            attention_kwargs=attention_kwargs,
+                            return_dict=False,
+                        )[0]
                     noise_pred = noise_pred_uncond + self.guidance_scale * (noise_pred_cond - noise_pred_uncond)
                 else:
                     noise_pred = noise_pred_cond

diffusers/pipelines/controlnet/pipeline_controlnet_blip_diffusion.py

@@ -29,7 +29,7 @@ from ...utils.torch_utils import randn_tensor
 from ..blip_diffusion.blip_image_processing import BlipImageProcessor
 from ..blip_diffusion.modeling_blip2 import Blip2QFormerModel
 from ..blip_diffusion.modeling_ctx_clip import ContextCLIPTextModel
-from ..pipeline_utils import DiffusionPipeline, ImagePipelineOutput
+from ..pipeline_utils import DeprecatedPipelineMixin, DiffusionPipeline, ImagePipelineOutput
 
 
 if is_torch_xla_available():
@@ -88,7 +88,7 @@ EXAMPLE_DOC_STRING = """
 """
 
 
-class BlipDiffusionControlNetPipeline(DiffusionPipeline):
+class BlipDiffusionControlNetPipeline(DeprecatedPipelineMixin, DiffusionPipeline):
     """
     Pipeline for Canny Edge based Controlled subject-driven generation using Blip Diffusion.
 
@@ -116,6 +116,7 @@ class BlipDiffusionControlNetPipeline(DiffusionPipeline):
             Position of the context token in the text encoder.
     """
 
+    _last_supported_version = "0.33.1"
     model_cpu_offload_seq = "qformer->text_encoder->unet->vae"
 
     def __init__(