diffusers 0.33.1__py3-none-any.whl → 0.35.0__py3-none-any.whl

diffusers/pipelines/ltx/pipeline_ltx_image2video.py

@@ -1,4 +1,4 @@
-# Copyright 2024 Lightricks and The HuggingFace Team. All rights reserved.
+# Copyright 2025 Lightricks 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.
@@ -159,6 +159,33 @@ def retrieve_latents(
         raise AttributeError("Could not access latents of provided encoder_output")
 
 
+# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.rescale_noise_cfg
+def rescale_noise_cfg(noise_cfg, noise_pred_text, guidance_rescale=0.0):
+    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://huggingface.co/papers/2305.08891).
+
+    Args:
+        noise_cfg (`torch.Tensor`):
+            The predicted noise tensor for the guided diffusion process.
+        noise_pred_text (`torch.Tensor`):
+            The predicted noise tensor for the text-guided diffusion process.
+        guidance_rescale (`float`, *optional*, defaults to 0.0):
+            A rescale factor applied to the noise predictions.
+
+    Returns:
+        noise_cfg (`torch.Tensor`): The rescaled noise prediction tensor.
+    """
+    std_text = noise_pred_text.std(dim=list(range(1, noise_pred_text.ndim)), keepdim=True)
+    std_cfg = noise_cfg.std(dim=list(range(1, noise_cfg.ndim)), keepdim=True)
+    # rescale the results from guidance (fixes overexposure)
+    noise_pred_rescaled = noise_cfg * (std_text / std_cfg)
+    # mix with the original results from guidance by factor guidance_rescale to avoid "plain looking" images
+    noise_cfg = guidance_rescale * noise_pred_rescaled + (1 - guidance_rescale) * noise_cfg
+    return noise_cfg
+
+
 class LTXImageToVideoPipeline(DiffusionPipeline, FromSingleFileMixin, LTXVideoLoraLoaderMixin):
     r"""
     Pipeline for image-to-video generation.
@@ -542,6 +569,10 @@ class LTXImageToVideoPipeline(DiffusionPipeline, FromSingleFileMixin, LTXVideoLo
     def guidance_scale(self):
         return self._guidance_scale
 
+    @property
+    def guidance_rescale(self):
+        return self._guidance_rescale
+
     @property
     def do_classifier_free_guidance(self):
         return self._guidance_scale > 1.0
@@ -576,6 +607,7 @@ class LTXImageToVideoPipeline(DiffusionPipeline, FromSingleFileMixin, LTXVideoLo
         num_inference_steps: int = 50,
         timesteps: List[int] = None,
         guidance_scale: float = 3,
+        guidance_rescale: float = 0.0,
         num_videos_per_prompt: Optional[int] = 1,
         generator: Optional[Union[torch.Generator, List[torch.Generator]]] = None,
         latents: Optional[torch.Tensor] = None,
@@ -615,11 +647,16 @@ class LTXImageToVideoPipeline(DiffusionPipeline, FromSingleFileMixin, LTXVideoLo
                 in their `set_timesteps` method. If not defined, the default behavior when `num_inference_steps` is
                 passed will be used. Must be in descending order.
             guidance_scale (`float`, defaults to `3 `):
-                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.
+            guidance_rescale (`float`, *optional*, defaults to 0.0):
+                Guidance rescale factor proposed by [Common Diffusion Noise Schedules and Sample Steps are
+                Flawed](https://arxiv.org/pdf/2305.08891.pdf) `guidance_scale` is defined as `φ` in equation 16. of
+                [Common Diffusion Noise Schedules and Sample Steps are Flawed](https://arxiv.org/pdf/2305.08891.pdf).
+                Guidance rescale factor should fix overexposure when using zero terminal SNR.
             num_videos_per_prompt (`int`, *optional*, defaults to 1):
                 The number of videos to generate per prompt.
             generator (`torch.Generator` or `List[torch.Generator]`, *optional*):
@@ -688,6 +725,7 @@ class LTXImageToVideoPipeline(DiffusionPipeline, FromSingleFileMixin, LTXVideoLo
         )
 
         self._guidance_scale = guidance_scale
+        self._guidance_rescale = guidance_rescale
         self._attention_kwargs = attention_kwargs
         self._interrupt = False
         self._current_timestep = None
@@ -792,18 +830,19 @@ class LTXImageToVideoPipeline(DiffusionPipeline, FromSingleFileMixin, LTXVideoLo
                 timestep = t.expand(latent_model_input.shape[0])
                 timestep = timestep.unsqueeze(-1) * (1 - conditioning_mask)
 
-                noise_pred = self.transformer(
-                    hidden_states=latent_model_input,
-                    encoder_hidden_states=prompt_embeds,
-                    timestep=timestep,
-                    encoder_attention_mask=prompt_attention_mask,
-                    num_frames=latent_num_frames,
-                    height=latent_height,
-                    width=latent_width,
-                    rope_interpolation_scale=rope_interpolation_scale,
-                    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,
+                        encoder_attention_mask=prompt_attention_mask,
+                        num_frames=latent_num_frames,
+                        height=latent_height,
+                        width=latent_width,
+                        rope_interpolation_scale=rope_interpolation_scale,
+                        attention_kwargs=attention_kwargs,
+                        return_dict=False,
+                    )[0]
                 noise_pred = noise_pred.float()
 
                 if self.do_classifier_free_guidance:
@@ -811,6 +850,12 @@ class LTXImageToVideoPipeline(DiffusionPipeline, FromSingleFileMixin, LTXVideoLo
                     noise_pred = noise_pred_uncond + self.guidance_scale * (noise_pred_text - noise_pred_uncond)
                     timestep, _ = timestep.chunk(2)
 
+                    if self.guidance_rescale > 0:
+                        # Based on 3.4. in https://arxiv.org/pdf/2305.08891.pdf
+                        noise_pred = rescale_noise_cfg(
+                            noise_pred, noise_pred_text, guidance_rescale=self.guidance_rescale
+                        )
+
                 # compute the previous noisy sample x_t -> x_t-1
                 noise_pred = self._unpack_latents(
                     noise_pred,

diffusers/pipelines/ltx/pipeline_ltx_latent_upsample.py

@@ -0,0 +1,277 @@
+# Copyright 2025 Lightricks 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.
+# 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 typing import List, Optional, Union
+
+import torch
+
+from ...image_processor import PipelineImageInput
+from ...models import AutoencoderKLLTXVideo
+from ...utils import get_logger
+from ...utils.torch_utils import randn_tensor
+from ...video_processor import VideoProcessor
+from ..pipeline_utils import DiffusionPipeline
+from .modeling_latent_upsampler import LTXLatentUpsamplerModel
+from .pipeline_output import LTXPipelineOutput
+
+
+logger = get_logger(__name__)  # pylint: disable=invalid-name
+
+
+# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion_img2img.retrieve_latents
+def retrieve_latents(
+    encoder_output: torch.Tensor, generator: Optional[torch.Generator] = None, sample_mode: str = "sample"
+):
+    if hasattr(encoder_output, "latent_dist") and sample_mode == "sample":
+        return encoder_output.latent_dist.sample(generator)
+    elif hasattr(encoder_output, "latent_dist") and sample_mode == "argmax":
+        return encoder_output.latent_dist.mode()
+    elif hasattr(encoder_output, "latents"):
+        return encoder_output.latents
+    else:
+        raise AttributeError("Could not access latents of provided encoder_output")
+
+
+class LTXLatentUpsamplePipeline(DiffusionPipeline):
+    model_cpu_offload_seq = ""
+
+    def __init__(
+        self,
+        vae: AutoencoderKLLTXVideo,
+        latent_upsampler: LTXLatentUpsamplerModel,
+    ) -> None:
+        super().__init__()
+
+        self.register_modules(vae=vae, latent_upsampler=latent_upsampler)
+
+        self.vae_spatial_compression_ratio = (
+            self.vae.spatial_compression_ratio if getattr(self, "vae", None) is not None else 32
+        )
+        self.vae_temporal_compression_ratio = (
+            self.vae.temporal_compression_ratio if getattr(self, "vae", None) is not None else 8
+        )
+        self.video_processor = VideoProcessor(vae_scale_factor=self.vae_spatial_compression_ratio)
+
+    def prepare_latents(
+        self,
+        video: Optional[torch.Tensor] = None,
+        batch_size: int = 1,
+        dtype: Optional[torch.dtype] = None,
+        device: Optional[torch.device] = None,
+        generator: Optional[torch.Generator] = None,
+        latents: Optional[torch.Tensor] = None,
+    ) -> torch.Tensor:
+        if latents is not None:
+            return latents.to(device=device, dtype=dtype)
+
+        video = video.to(device=device, dtype=self.vae.dtype)
+        if isinstance(generator, list):
+            if len(generator) != batch_size:
+                raise ValueError(
+                    f"You have passed a list of generators of length {len(generator)}, but requested an effective batch"
+                    f" size of {batch_size}. Make sure the batch size matches the length of the generators."
+                )
+
+            init_latents = [
+                retrieve_latents(self.vae.encode(video[i].unsqueeze(0)), generator[i]) for i in range(batch_size)
+            ]
+        else:
+            init_latents = [retrieve_latents(self.vae.encode(vid.unsqueeze(0)), generator) for vid in video]
+
+        init_latents = torch.cat(init_latents, dim=0).to(dtype)
+        init_latents = self._normalize_latents(init_latents, self.vae.latents_mean, self.vae.latents_std)
+        return init_latents
+
+    def adain_filter_latent(self, latents: torch.Tensor, reference_latents: torch.Tensor, factor: float = 1.0):
+        """
+        Applies Adaptive Instance Normalization (AdaIN) to a latent tensor based on statistics from a reference latent
+        tensor.
+
+        Args:
+            latent (`torch.Tensor`):
+                Input latents to normalize
+            reference_latents (`torch.Tensor`):
+                The reference latents providing style statistics.
+            factor (`float`):
+                Blending factor between original and transformed latent. Range: -10.0 to 10.0, Default: 1.0
+
+        Returns:
+            torch.Tensor: The transformed latent tensor
+        """
+        result = latents.clone()
+
+        for i in range(latents.size(0)):
+            for c in range(latents.size(1)):
+                r_sd, r_mean = torch.std_mean(reference_latents[i, c], dim=None)  # index by original dim order
+                i_sd, i_mean = torch.std_mean(result[i, c], dim=None)
+
+                result[i, c] = ((result[i, c] - i_mean) / i_sd) * r_sd + r_mean
+
+        result = torch.lerp(latents, result, factor)
+        return result
+
+    @staticmethod
+    # Copied from diffusers.pipelines.ltx.pipeline_ltx.LTXPipeline._normalize_latents
+    def _normalize_latents(
+        latents: torch.Tensor, latents_mean: torch.Tensor, latents_std: torch.Tensor, scaling_factor: float = 1.0
+    ) -> torch.Tensor:
+        # Normalize latents across the channel dimension [B, C, F, H, W]
+        latents_mean = latents_mean.view(1, -1, 1, 1, 1).to(latents.device, latents.dtype)
+        latents_std = latents_std.view(1, -1, 1, 1, 1).to(latents.device, latents.dtype)
+        latents = (latents - latents_mean) * scaling_factor / latents_std
+        return latents
+
+    @staticmethod
+    # Copied from diffusers.pipelines.ltx.pipeline_ltx.LTXPipeline._denormalize_latents
+    def _denormalize_latents(
+        latents: torch.Tensor, latents_mean: torch.Tensor, latents_std: torch.Tensor, scaling_factor: float = 1.0
+    ) -> torch.Tensor:
+        # Denormalize latents across the channel dimension [B, C, F, H, W]
+        latents_mean = latents_mean.view(1, -1, 1, 1, 1).to(latents.device, latents.dtype)
+        latents_std = latents_std.view(1, -1, 1, 1, 1).to(latents.device, latents.dtype)
+        latents = latents * latents_std / scaling_factor + latents_mean
+        return latents
+
+    def enable_vae_slicing(self):
+        r"""
+        Enable sliced VAE decoding. When this option is enabled, the VAE will split the input tensor in slices to
+        compute decoding in several steps. This is useful to save some memory and allow larger batch sizes.
+        """
+        self.vae.enable_slicing()
+
+    def disable_vae_slicing(self):
+        r"""
+        Disable sliced VAE decoding. If `enable_vae_slicing` was previously enabled, this method will go back to
+        computing decoding in one step.
+        """
+        self.vae.disable_slicing()
+
+    def enable_vae_tiling(self):
+        r"""
+        Enable tiled VAE decoding. When this option is enabled, the VAE will split the input tensor into tiles to
+        compute decoding and encoding in several steps. This is useful for saving a large amount of memory and to allow
+        processing larger images.
+        """
+        self.vae.enable_tiling()
+
+    def disable_vae_tiling(self):
+        r"""
+        Disable tiled VAE decoding. If `enable_vae_tiling` was previously enabled, this method will go back to
+        computing decoding in one step.
+        """
+        self.vae.disable_tiling()
+
+    def check_inputs(self, video, height, width, latents):
+        if height % self.vae_spatial_compression_ratio != 0 or width % self.vae_spatial_compression_ratio != 0:
+            raise ValueError(f"`height` and `width` have to be divisible by 32 but are {height} and {width}.")
+
+        if video is not None and latents is not None:
+            raise ValueError("Only one of `video` or `latents` can be provided.")
+        if video is None and latents is None:
+            raise ValueError("One of `video` or `latents` has to be provided.")
+
+    @torch.no_grad()
+    def __call__(
+        self,
+        video: Optional[List[PipelineImageInput]] = None,
+        height: int = 512,
+        width: int = 704,
+        latents: Optional[torch.Tensor] = None,
+        decode_timestep: Union[float, List[float]] = 0.0,
+        decode_noise_scale: Optional[Union[float, List[float]]] = None,
+        adain_factor: float = 0.0,
+        generator: Optional[Union[torch.Generator, List[torch.Generator]]] = None,
+        output_type: Optional[str] = "pil",
+        return_dict: bool = True,
+    ):
+        self.check_inputs(
+            video=video,
+            height=height,
+            width=width,
+            latents=latents,
+        )
+
+        if video is not None:
+            # Batched video input is not yet tested/supported. TODO: take a look later
+            batch_size = 1
+        else:
+            batch_size = latents.shape[0]
+        device = self._execution_device
+
+        if video is not None:
+            num_frames = len(video)
+            if num_frames % self.vae_temporal_compression_ratio != 1:
+                num_frames = (
+                    num_frames // self.vae_temporal_compression_ratio * self.vae_temporal_compression_ratio + 1
+                )
+                video = video[:num_frames]
+                logger.warning(
+                    f"Video length expected to be of the form `k * {self.vae_temporal_compression_ratio} + 1` but is {len(video)}. Truncating to {num_frames} frames."
+                )
+            video = self.video_processor.preprocess_video(video, height=height, width=width)
+            video = video.to(device=device, dtype=torch.float32)
+
+        latents = self.prepare_latents(
+            video=video,
+            batch_size=batch_size,
+            dtype=torch.float32,
+            device=device,
+            generator=generator,
+            latents=latents,
+        )
+
+        latents = self._denormalize_latents(
+            latents, self.vae.latents_mean, self.vae.latents_std, self.vae.config.scaling_factor
+        )
+        latents = latents.to(self.latent_upsampler.dtype)
+        latents_upsampled = self.latent_upsampler(latents)
+
+        if adain_factor > 0.0:
+            latents = self.adain_filter_latent(latents_upsampled, latents, adain_factor)
+        else:
+            latents = latents_upsampled
+
+        if output_type == "latent":
+            latents = self._normalize_latents(
+                latents, self.vae.latents_mean, self.vae.latents_std, self.vae.config.scaling_factor
+            )
+            video = latents
+        else:
+            if not self.vae.config.timestep_conditioning:
+                timestep = None
+            else:
+                noise = randn_tensor(latents.shape, generator=generator, device=device, dtype=latents.dtype)
+                if not isinstance(decode_timestep, list):
+                    decode_timestep = [decode_timestep] * batch_size
+                if decode_noise_scale is None:
+                    decode_noise_scale = decode_timestep
+                elif not isinstance(decode_noise_scale, list):
+                    decode_noise_scale = [decode_noise_scale] * batch_size
+
+                timestep = torch.tensor(decode_timestep, device=device, dtype=latents.dtype)
+                decode_noise_scale = torch.tensor(decode_noise_scale, device=device, dtype=latents.dtype)[
+                    :, None, None, None, None
+                ]
+                latents = (1 - decode_noise_scale) * latents + decode_noise_scale * noise
+
+            video = self.vae.decode(latents, timestep, return_dict=False)[0]
+            video = self.video_processor.postprocess_video(video, output_type=output_type)
+
+        # Offload all models
+        self.maybe_free_model_hooks()
+
+        if not return_dict:
+            return (video,)
+
+        return LTXPipelineOutput(frames=video)

diffusers/pipelines/lumina/pipeline_lumina.py

@@ -1,4 +1,4 @@
-# Copyright 2024 Alpha-VLLM and The HuggingFace Team. All rights reserved.
+# Copyright 2025 Alpha-VLLM 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.
@@ -372,7 +372,7 @@ class LuminaPipeline(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())
@@ -534,7 +534,7 @@ class LuminaPipeline(DiffusionPipeline):
         # &amp
         caption = re.sub(r"&amp", "", caption)
 
-        # ip adresses:
+        # ip addresses:
         caption = re.sub(r"\d{1,3}\.\d{1,3}\.\d{1,3}\.\d{1,3}", " ", caption)
 
         # article ids:
@@ -619,7 +619,7 @@ class LuminaPipeline(DiffusionPipeline):
         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):
@@ -677,11 +677,11 @@ class LuminaPipeline(DiffusionPipeline):
                 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 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.
             height (`int`, *optional*, defaults to self.unet.config.sample_size):
@@ -689,8 +689,8 @@ class LuminaPipeline(DiffusionPipeline):
             width (`int`, *optional*, defaults to self.unet.config.sample_size):
                 The width in pixels of the generated image.
             eta (`float`, *optional*, defaults to 0.0):
-                Corresponds to parameter eta (η) in the DDIM paper: https://arxiv.org/abs/2010.02502. Only applies to
-                [`schedulers.DDIMScheduler`], will be ignored for others.
+                Corresponds to parameter eta (η) in the DDIM paper: https://huggingface.co/papers/2010.02502. Only
+                applies to [`schedulers.DDIMScheduler`], will be ignored for others.
             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.
@@ -771,7 +771,7 @@ class LuminaPipeline(DiffusionPipeline):
         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.
         do_classifier_free_guidance = guidance_scale > 1.0
 
@@ -848,7 +848,7 @@ class LuminaPipeline(DiffusionPipeline):
                 # prepare image_rotary_emb for positional encoding
                 # dynamic scaling_factor for different resolution.
                 # NOTE: For `Time-aware` denosing mechanism from Lumina-Next
-                # https://arxiv.org/abs/2406.18583, Sec 2.3
+                # https://huggingface.co/papers/2406.18583, Sec 2.3
                 # NOTE: We should compute different image_rotary_emb with different timestep.
                 if current_timestep[0] < scaling_watershed:
                     linear_factor = scaling_factor

diffusers/pipelines/lumina2/pipeline_lumina2.py

@@ -1,4 +1,4 @@
-# Copyright 2024 Alpha-VLLM and The HuggingFace Team. All rights reserved.
+# Copyright 2025 Alpha-VLLM 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.
@@ -342,7 +342,7 @@ class Lumina2Pipeline(DiffusionPipeline, Lumina2LoraLoaderMixin):
     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())
@@ -487,7 +487,7 @@ class Lumina2Pipeline(DiffusionPipeline, Lumina2LoraLoaderMixin):
         return self._attention_kwargs
 
     # 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):
@@ -544,11 +544,11 @@ class Lumina2Pipeline(DiffusionPipeline, Lumina2LoraLoaderMixin):
                 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 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.
             height (`int`, *optional*, defaults to self.unet.config.sample_size):
@@ -556,8 +556,8 @@ class Lumina2Pipeline(DiffusionPipeline, Lumina2LoraLoaderMixin):
             width (`int`, *optional*, defaults to self.unet.config.sample_size):
                 The width in pixels of the generated image.
             eta (`float`, *optional*, defaults to 0.0):
-                Corresponds to parameter eta (η) in the DDIM paper: https://arxiv.org/abs/2010.02502. Only applies to
-                [`schedulers.DDIMScheduler`], will be ignored for others.
+                Corresponds to parameter eta (η) in the DDIM paper: https://huggingface.co/papers/2010.02502. Only
+                applies to [`schedulers.DDIMScheduler`], will be ignored for others.
             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/marigold/marigold_image_processing.py

@@ -426,7 +426,7 @@ class MarigoldImageProcessor(ConfigMixin):
                 if isinstance(img, np.ndarray):
                     img = torch.from_numpy(img)
                 if not torch.is_floating_point(img):
-                    raise ValueError(f"{prefix}: unexected dtype={img.dtype}.")
+                    raise ValueError(f"{prefix}: unexpected dtype={img.dtype}.")
             else:
                 raise ValueError(f"{prefix}: unexpected type={type(img)}.")
             if val_min != 0.0 or val_max != 1.0:
@@ -464,7 +464,7 @@ class MarigoldImageProcessor(ConfigMixin):
             if torch.is_tensor(img):
                 img = img.cpu().numpy()
             if not np.issubdtype(img.dtype, np.floating):
-                raise ValueError(f"{prefix}: unexected dtype={img.dtype}.")
+                raise ValueError(f"{prefix}: unexpected dtype={img.dtype}.")
             if val_min != 0.0 or val_max != 1.0:
                 img = (img - val_min) / (val_max - val_min)
             img = (img * (2**16 - 1)).astype(np.uint16)

diffusers/pipelines/mochi/pipeline_mochi.py

@@ -1,4 +1,4 @@
-# Copyright 2024 Genmo and The HuggingFace Team. All rights reserved.
+# Copyright 2025 Genmo 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.
@@ -521,11 +521,11 @@ class MochiPipeline(DiffusionPipeline, Mochi1LoraLoaderMixin):
                 in their `set_timesteps` method. If not defined, the default behavior when `num_inference_steps` is
                 passed will be used. Must be in descending order.
             guidance_scale (`float`, defaults to `4.5`):
-                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.
+                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_videos_per_prompt (`int`, *optional*, defaults to 1):
                 The number of videos to generate per prompt.
             generator (`torch.Generator` or `List[torch.Generator]`, *optional*):
@@ -671,14 +671,15 @@ class MochiPipeline(DiffusionPipeline, Mochi1LoraLoaderMixin):
                 # broadcast to batch dimension in a way that's compatible with ONNX/Core ML
                 timestep = t.expand(latent_model_input.shape[0]).to(latents.dtype)
 
-                noise_pred = self.transformer(
-                    hidden_states=latent_model_input,
-                    encoder_hidden_states=prompt_embeds,
-                    timestep=timestep,
-                    encoder_attention_mask=prompt_attention_mask,
-                    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,
+                        encoder_attention_mask=prompt_attention_mask,
+                        attention_kwargs=attention_kwargs,
+                        return_dict=False,
+                    )[0]
                 # Mochi CFG + Sampling runs in FP32
                 noise_pred = noise_pred.to(torch.float32)