Add SkyReels V2: Infinite-Length Film Generative Model (#11518)

* style * Fix class name casing for SkyReelsV2 components in multiple files to ensure consistency and correct functionality. * cleaning * cleansing * Refactor `get_timestep_embedding` to move modifications into `SkyReelsV2TimeTextImageEmbedding`. * Remove unnecessary line break in `get_timestep_embedding` function for cleaner code. * Remove `skyreels_v2` entry from `_import_structure` and update its initialization to directly assign the list of SkyReelsV2 components. * cleansing * Refactor attention processing in `SkyReelsV2AttnProcessor2_0` to always convert query, key, and value to `torch.bfloat16`, simplifying the code and improving clarity. * Enhance example usage in `pipeline_skyreels_v2_diffusion_forcing.py` by adding VAE initialization and detailed prompt for video generation, improving clarity and usability of the documentation. * Refactor import structure in `__init__.py` for SkyReelsV2 components and improve formatting in `pipeline_skyreels_v2_diffusion_forcing.py` to enhance code readability and maintainability. * Update `guidance_scale` parameter in `SkyReelsV2DiffusionForcingPipeline` from 5.0 to 6.0 to enhance video generation quality. * Update `guidance_scale` parameter in example documentation and class definition of `SkyReelsV2DiffusionForcingPipeline` to ensure consistency and improve video generation quality. * Update `causal_block_size` parameter in `SkyReelsV2DiffusionForcingPipeline` to default to `None`. * up * Fix dtype conversion for `timestep_proj` in `SkyReelsV2Transformer3DModel` to *ensure* correct tensor operations. * Optimize causal mask generation by replacing repeated tensor with `repeat_interleave` for improved efficiency in `SkyReelsV2Transformer3DModel`. * style * Enhance example documentation in `SkyReelsV2DiffusionForcingPipeline` with guidance scale and shift parameters for T2V and I2V. Remove unused `retrieve_latents` function to streamline the code. * Refactor sample scheduler creation in `SkyReelsV2DiffusionForcingPipeline` to use `deepcopy` for improved state management during inference steps. * Enhance error handling and documentation in `SkyReelsV2DiffusionForcingPipeline` for `overlap_history` and `addnoise_condition` parameters to improve long video generation guidance. * Update documentation and progress bar handling in `SkyReelsV2DiffusionForcingPipeline` to clarify asynchronous inference settings and improve progress tracking during denoising steps. * Refine progress bar calculation in `SkyReelsV2DiffusionForcingPipeline` by rounding the step size to one decimal place for improved readability during denoising steps. * Update import statements in `SkyReelsV2DiffusionForcingPipeline` documentation for improved clarity and organization. * Refactor progress bar handling in `SkyReelsV2DiffusionForcingPipeline` to use total steps instead of calculated step size. * update templates for i2v, v2v * Add `retrieve_latents` function to streamline latent retrieval in `SkyReelsV2DiffusionForcingPipeline`. Update video latent processing to utilize this new function for improved clarity and maintainability. * Add `retrieve_latents` function to both i2v and v2v pipelines for consistent latent retrieval. Update video latent processing to utilize this function, enhancing clarity and maintainability across the SkyReelsV2DiffusionForcingPipeline implementations. * Remove redundant ValueError for `overlap_history` in `SkyReelsV2DiffusionForcingPipeline` to streamline error handling and improve user guidance for long video generation. * Update default video dimensions and flow matching scheduler parameter in `SkyReelsV2DiffusionForcingPipeline` to enhance video generation capabilities. * Refactor `SkyReelsV2DiffusionForcingPipeline` to support Image-to-Video (i2v) generation. Update class name, add image encoding functionality, and adjust parameters for improved video generation. Enhance error handling for image inputs and update documentation accordingly. * Improve organization for image-last_image condition. * Refactor `SkyReelsV2DiffusionForcingImageToVideoPipeline` to improve latent preparation and video condition handling integration. * style * style * Add example usage of PIL for image input in `SkyReelsV2DiffusionForcingImageToVideoPipeline` documentation. * Refactor `SkyReelsV2DiffusionForcingPipeline` to `SkyReelsV2DiffusionForcingVideoToVideoPipeline`, enhancing support for Video-to-Video (v2v) generation. Introduce video input handling, update latent preparation logic, and improve error handling for input parameters. * Refactor `SkyReelsV2DiffusionForcingImageToVideoPipeline` by removing the `image_encoder` and `image_processor` dependencies. Update the CPU offload sequence accordingly. * Refactor `SkyReelsV2DiffusionForcingImageToVideoPipeline` to enhance latent preparation logic and condition handling. Update image input type to `Optional`, streamline video condition processing, and improve handling of `last_image` during latent generation. * Enhance `SkyReelsV2DiffusionForcingPipeline` by refining latent preparation for long video generation. Introduce new parameters for video handling, overlap history, and causal block size. Update logic to accommodate both short and long video scenarios, ensuring compatibility and improved processing. * refactor * fix num_frames * fix prefix_video_latents * up * refactor * Fix typo in scheduler method call within `SkyReelsV2DiffusionForcingVideoToVideoPipeline` to ensure proper noise scaling during latent generation. * up * Enhance `SkyReelsV2DiffusionForcingImageToVideoPipeline` by adding support for `last_image` parameter and refining latent frame calculations. Update preprocessing logic. * add statistics * Refine latent frame handling in `SkyReelsV2DiffusionForcingImageToVideoPipeline` by correcting variable names and reintroducing latent mean and standard deviation calculations. Update logic for frame preparation and sampling to ensure accurate video generation. * up * refactor * up * Refactor `SkyReelsV2DiffusionForcingVideoToVideoPipeline` to improve latent handling by enforcing tensor input for video, updating frame preparation logic, and adjusting default frame count. Enhance preprocessing and postprocessing steps for better integration. * style * fix vae output indexing * upup * up * Fix tensor concatenation and repetition logic in `SkyReelsV2DiffusionForcingImageToVideoPipeline` to ensure correct dimensionality for video conditions and latent conditions. * Refactor latent retrieval logic in `SkyReelsV2DiffusionForcingVideoToVideoPipeline` to handle tensor dimensions more robustly, ensuring compatibility with both 3D and 4D video inputs. * Enhance logging in `SkyReelsV2DiffusionForcing` pipelines by adding iteration print statements for better debugging. Clean up unused code related to prefix video latents length calculation in `SkyReelsV2DiffusionForcingImageToVideoPipeline`. * Update latent handling in `SkyReelsV2DiffusionForcingImageToVideoPipeline` to conditionally set latents based on video iteration state, improving flexibility for video input processing. * Refactor `SkyReelsV2TimeTextImageEmbedding` to utilize `get_1d_sincos_pos_embed_from_grid` for timestep projection. * Enhance `get_1d_sincos_pos_embed_from_grid` function to include an optional parameter `flip_sin_to_cos` for flipping sine and cosine embeddings, improving flexibility in positional embedding generation. * Update timestep projection in `SkyReelsV2TimeTextImageEmbedding` to include `flip_sin_to_cos` parameter, enhancing the flexibility of time embedding generation. * Refactor tensor type handling in `SkyReelsV2AttnProcessor2_0` and `SkyReelsV2TransformerBlock` to ensure consistent use of `torch.float32` and `torch.bfloat16`, improving integration. * Update tensor type in `SkyReelsV2RotaryPosEmbed` to use `torch.float32` for frequency calculations, ensuring consistency in data types across the model. * Refactor `SkyReelsV2TimeTextImageEmbedding` to utilize automatic mixed precision for timestep projection. * down * down * style * Add debug tensor tracking to `SkyReelsV2Transformer3DModel` for enhanced debugging and output analysis; update `Transformer2DModelOutput` to include debug tensors. * up * Refactor indentation in `SkyReelsV2AttnProcessor2_0` to improve code readability and maintain consistency in style. * Convert query, key, and value tensors to bfloat16 in `SkyReelsV2AttnProcessor2_0` for improved performance. * Add debug print statements in `SkyReelsV2TransformerBlock` to track tensor shapes and values for improved debugging and analysis. * debug * debug * Remove commented-out debug tensor tracking from `SkyReelsV2TransformerBlock` * Add functionality to save processed video latents as a Safetensors file in `SkyReelsV2DiffusionForcingPipeline`. * up * Add functionality to save output latents as a Safetensors file in `SkyReelsV2DiffusionForcingPipeline`. * up * Remove additional commented-out debug tensor tracking from `SkyReelsV2TransformerBlock` and `SkyReelsV2Transformer3DModel` for cleaner code. * style * cleansing * Update example documentation and parameters in `SkyReelsV2Pipeline`. Adjusted example code for loading models, modified default values for height, width, num_frames, and guidance_scale, and improved output video quality settings. * Update shift parameter in example documentation and default values across SkyReels V2 pipelines. Adjusted shift values for I2V from 3.0 to 5.0 and updated related example code for consistency. * Update example documentation in SkyReels V2 pipelines to include available model options and update model references for loading. Adjusted model names to reflect the latest versions across I2V, V2V, and T2V pipelines. * Add test templates * style * Add docs template * Add SkyReels V2 Diffusion Forcing Video-to-Video Pipeline to imports * style * fix-copies * convert i2v 1.3b * Update transformer configuration to include `image_dim` for SkyReels V2 models and refactor imports to use `SkyReelsV2Transformer3DModel`. * Refactor transformer import in SkyReels V2 pipeline to use `SkyReelsV2Transformer3DModel` for consistency. * Update transformer configuration in SkyReels V2 to increase `in_channels` from 16 to 36 for i2v conf. * Update transformer configuration in SkyReels V2 to set `added_kv_proj_dim` values for different model types. * up * up * up * Add SkyReelsV2Pipeline support for T2V model type in conversion script * upp * Refactor model type checks in conversion script to use substring matching for improved flexibility * upp * Fix shard path formatting in conversion script to accommodate varying model types by dynamically adjusting zero padding. * Update sharded safetensors loading logic in conversion script to use substring matching for model directory checks * Update scheduler parameters in SkyReels V2 test files for consistency across image and video pipelines * Refactor conversion script to initialize text encoder, tokenizer, and scheduler for SkyReels pipelines, enhancing model integration * style * Update documentation for SkyReels-V2, introducing the Infinite-length Film Generative model, enhancing text-to-video generation examples, and updating model references throughout the API documentation. * Add SkyReelsV2Transformer3DModel and FlowMatchUniPCMultistepScheduler documentation, updating TOC and introducing new model and scheduler files. * style * Update documentation for SkyReelsV2DiffusionForcingPipeline to correct flow matching scheduler parameter for I2V from 3.0 to 5.0, ensuring clarity in usage examples. * Add documentation for causal_block_size parameter in SkyReelsV2DF pipelines, clarifying its role in asynchronous inference. * Simplify min_ar_step calculation in SkyReelsV2DiffusionForcingPipeline to improve clarity. * style and fix-copies * style * Add documentation for SkyReelsV2Transformer3DModel Introduced a new markdown file detailing the SkyReelsV2Transformer3DModel, including usage instructions and model output specifications. * Update test configurations for SkyReelsV2 pipelines - Adjusted `in_channels` from 36 to 16 in `test_skyreels_v2_df_image_to_video.py`. - Added new parameters: `overlap_history`, `num_frames`, and `base_num_frames` in `test_skyreels_v2_df_video_to_video.py`. - Updated expected output shape in video tests from (17, 3, 16, 16) to (41, 3, 16, 16). * Refines SkyReelsV2DF test parameters * Update src/diffusers/models/modeling_outputs.py Co-authored-by: Aryan <contact.aryanvs@gmail.com> * Refactor `grid_sizes` processing by using already-calculated post-patch parameters to simplify * Update docs/source/en/api/pipelines/skyreels_v2.md Co-authored-by: Aryan <contact.aryanvs@gmail.com> * Refactor parameter naming for diffusion forcing in SkyReelsV2 pipelines - Changed `flag_df` to `enable_diffusion_forcing` for clarity in the SkyReelsV2Transformer3DModel and associated pipelines. - Updated all relevant method calls to reflect the new parameter name. * Revert _toctree.yml to adjust section expansion states * style * Update docs/source/en/api/models/skyreels_v2_transformer_3d.md Co-authored-by: YiYi Xu <yixu310@gmail.com> * Add copying label to SkyReelsV2ImageEmbedding from WanImageEmbedding. * Refactor transformer block processing in SkyReelsV2Transformer3DModel - Ensured proper handling of hidden states during both gradient checkpointing and standard processing. * Update SkyReels V2 documentation to remove VRAM requirement and streamline imports - Removed the mention of ~13GB VRAM requirement for the SkyReels-V2 model. - Simplified import statements by removing unused `load_image` import. * Add SkyReelsV2LoraLoaderMixin for loading and managing LoRA layers in SkyReelsV2Transformer3DModel - Introduced SkyReelsV2LoraLoaderMixin class to handle loading, saving, and fusing of LoRA weights specific to the SkyReelsV2 model. - Implemented methods for state dict management, including compatibility checks for various LoRA formats. - Enhanced functionality for loading weights with options for low CPU memory usage and hotswapping. - Added detailed docstrings for clarity on parameters and usage. * Update SkyReelsV2 documentation and loader mixin references - Corrected the documentation to reference the new `SkyReelsV2LoraLoaderMixin` for loading LoRA weights. - Updated comments in the `SkyReelsV2LoraLoaderMixin` class to reflect changes in model references from `WanTransformer3DModel` to `SkyReelsV2Transformer3DModel`. * Enhance SkyReelsV2 integration by adding SkyReelsV2LoraLoaderMixin references - Added `SkyReelsV2LoraLoaderMixin` to the documentation and loader imports for improved LoRA weight management. - Updated multiple pipeline classes to inherit from `SkyReelsV2LoraLoaderMixin` instead of `WanLoraLoaderMixin`. * Update SkyReelsV2 model references in documentation - Replaced placeholder model paths with actual paths for SkyReels-V2 models in multiple pipeline files. - Ensured consistency across the documentation for loading models in the SkyReelsV2 pipelines. * style * fix-copies * Refactor `fps_projection` in `SkyReelsV2Transformer3DModel` - Replaced the sequential linear layers for `fps_projection` with a `FeedForward` layer using `SiLU` activation for better integration. * Update docs * Refactor video processing in SkyReelsV2DiffusionForcingPipeline - Renamed parameters for clarity: `video` to `video_latents` and `overlap_history` to `overlap_history_latent_frames`. - Updated logic for handling long video generation, including adjustments to latent frame calculations and accumulation. - Consolidated handling of latents for both long and short video generation scenarios. - Final decoding step now consistently converts latents to pixels, ensuring proper output format. * Update activation function in `fps_projection` of `SkyReelsV2Transformer3DModel` - Changed activation function from `silu` to `linear-silu` in the `fps_projection` layer for improved performance and integration. * Add fps_projection layer renaming in convert_skyreelsv2_to_diffusers.py - Updated key mappings for the `fps_projection` layer to align with new naming conventions, ensuring consistency in model integration. * Fix fps_projection assignment in SkyReelsV2Transformer3DModel - Corrected the assignment of the `fps_projection` layer to ensure it is properly cast to the appropriate data type, enhancing model functionality. * Update _keep_in_fp32_modules in SkyReelsV2Transformer3DModel - Added `fps_projection` to the list of modules that should remain in FP32 precision, ensuring proper handling of data types during model operations. * Remove integration test classes from SkyReelsV2 test files - Deleted the `SkyReelsV2DiffusionForcingPipelineIntegrationTests` and `SkyReelsV2PipelineIntegrationTests` classes along with their associated setup, teardown, and test methods, as they were not implemented and not needed for current testing. * style * Refactor: Remove hardcoded `torch.bfloat16` cast in attention * Refactor: Simplify data type handling in transformer model Removes unnecessary data type conversions for the FPS embedding and timestep projection. This change simplifies the forward pass by relying on the inherent data types of the tensors. * Refactor: Remove `fps_projection` from `_keep_in_fp32_modules` in `SkyReelsV2Transformer3DModel` * Update src/diffusers/models/transformers/transformer_skyreels_v2.py Co-authored-by: Aryan <contact.aryanvs@gmail.com> * Refactor: Remove unused flags and simplify attention mask handling in SkyReelsV2AttnProcessor2_0 and SkyReelsV2Transformer3DModel Refactor: Simplify causal attention logic in SkyReelsV2 Removes the `flag_causal_attention` and `_flag_ar_attention` flags to simplify the implementation. The decision to apply a causal attention mask is now based directly on the `num_frame_per_block` configuration, eliminating redundant flags and conditional checks. This streamlines the attention mechanism and simplifies the `set_ar_attention` methods. * Refactor: Clarify variable names for latent frames Renames `base_num_frames` to `base_latent_num_frames` to make it explicit that the variable refers to the number of frames in the latent space. This change improves code readability and reduces potential confusion between latent frames and decoded video frames. The `num_frames` parameter in `generate_timestep_matrix` is also renamed to `num_latent_frames` for consistency. * Enhance documentation: Add detailed docstring for timestep matrix generation in SkyReelsV2DiffusionForcingPipeline * Docs: Clarify long video chunking in pipeline docstring Improves the explanation of long video processing within the pipeline's docstring. The update replaces the abstract description with a concrete example, illustrating how the sliding window mechanism works with overlapping chunks. This makes the roles of `base_num_frames` and `overlap_history` clearer for users. * Docs: Move visual demonstration and processing details for SkyReelsV2DiffusionForcingPipeline to docs page from the code * Docs: Update asynchronous processing timeline and examples for long video handling in SkyReels-V2 documentation * Enhance timestep matrix generation documentation and logic for synchronous/asynchronous video processing * Update timestep matrix documentation and enhance analysis for clarity in SkyReelsV2DiffusionForcingPipeline * Docs: Update visual demonstration section and add detailed step matrix construction example for asynchronous processing in SkyReelsV2DiffusionForcingPipeline * style * fix-copies * Refactor parameter names for clarity in SkyReelsV2DiffusionForcingImageToVideoPipeline and SkyReelsV2DiffusionForcingVideoToVideoPipeline * Refactor: Avoid VAE roundtrip in long video generation Improves performance and quality for long video generation by operating entirely in latent space during the iterative generation process. Instead of decoding latents to video and then re-encoding the overlapping section for the next chunk, this change passes the generated latents directly between iterations. This avoids a computationally expensive and potentially lossy VAE decode/encode cycle within the loop. The full video is now decoded only once from the accumulated latents at the end of the process. * Refactor: Rename prefix_video_latents_length to prefix_video_latents_frames for clarity * Refactor: Rename num_latent_frames to current_num_latent_frames for clarity in SkyReelsV2DiffusionForcingImageToVideoPipeline * Refactor: Enhance long video generation logic and improve latent handling in SkyReelsV2DiffusionForcingImageToVideoPipeline Refactor: Unify video generation and pass latents directly Unifies the separate code paths for short and long video generation into a single, streamlined loop. This change eliminates the inefficient decode-encode cycle during long video generation. Instead of converting latents to pixel-space video between chunks, the pipeline now passes the generated latents directly to the next iteration. This improves performance, avoids potential quality loss from intermediate VAE steps, and enhances code maintainability by removing significant duplication. * style * Refactor: Remove overlap_history parameter and streamline long video generation logic in SkyReelsV2DiffusionForcingImageToVideoPipeline Refactor: Streamline long video generation logic Removes the `overlap_history` parameter and simplifies the conditioning process for long video generation. This change avoids a redundant VAE encoding step by directly using latent frames from the previous chunk for conditioning. It also moves image preprocessing outside the main generation loop to prevent repeated computations and clarifies the handling of prefix latents. * style * Refactor latent handling in i2v diffusion forcing pipeline Improves the latent conditioning and accumulation logic within the image-to-video diffusion forcing loop. - Corrects the splitting of the initial conditioning tensor to robustly handle both even and odd lengths. - Simplifies how latents are accumulated across iterations for long video generation. - Ensures the final latents are trimmed correctly before decoding only when a `last_image` is provided. * Refactor: Remove overlap_history parameter from SkyReelsV2DiffusionForcingImageToVideoPipeline * Refactor: Adjust video_latents parameter handling in prepare_latents method * style * Refactor: Update long video iteration print statements for clarity * Fix: Update transformer config with dynamic causal block size Updates the SkyReelsV2 pipelines to correctly set the `causal_block_size` in the transformer's configuration when it's provided during a pipeline call. This ensures the model configuration reflects the user's specified setting for the inference run. The `set_ar_attention` method is also renamed to `_set_ar_attention` to mark it as an internal helper. * style * Refactor: Adjust video input size and expected output shape in inference test * Refactor: Rename video variables for clarity in SkyReelsV2DiffusionForcingVideoToVideoPipeline * Docs: Clarify time embedding logic in SkyReelsV2 Adds comments to explain the handling of different time embedding tensor dimensions. A 2D tensor is used for standard models with a single time embedding per batch, while a 3D tensor is used for Diffusion Forcing models where each frame has its own time embedding. This clarifies the expected input for different model variations. * Docs: Update SkyReels V2 pipeline examples Updates the docstring examples for the SkyReels V2 pipelines to reflect current best practices and API changes. - Removes the `shift` parameter from pipeline call examples, as it is now configured directly on the scheduler. - Replaces the `set_ar_attention` method call with the `causal_block_size` argument in the pipeline call for diffusion forcing examples. - Adjusts recommended parameters for I2V and V2V examples, including inference steps, guidance scale, and `ar_step`. * Refactor: Remove `shift` parameter from SkyReelsV2 pipelines Removes the `shift` parameter from the call signature of all SkyReelsV2 pipelines. This parameter is a scheduler-specific configuration and should be set directly on the scheduler during its initialization, rather than being passed at runtime through the pipeline. This change simplifies the pipeline API. Usage examples are updated to reflect that the `shift` value should now be passed when creating the `FlowMatchUniPCMultistepScheduler`. * Refactors SkyReelsV2 image-to-video tests and adds last image case Simplifies the test suite by removing a duplicated test class and streamlining the dummy component and input generation. Adds a new test to verify the pipeline's behavior when a `last_image` is provided as input for conditioning. * test: Add image components to SkyReelsV2 pipeline test Adds the `image_encoder` and `image_processor` to the test components for the image-to-video pipeline. Also replaces a hardcoded value for the positional embedding sequence length with a more descriptive calculation, improving clarity. * test: Add callback configuration test for SkyReelsV2DiffusionForcingVideoToVideoPipeline test: Add callback test for SkyReelsV2DFV2V pipeline Adds a test to validate the callback functionality for the `SkyReelsV2DiffusionForcingVideoToVideoPipeline`. This test confirms that `callback_on_step_end` is invoked correctly and can modify the pipeline's state during inference. It uses a callback to dynamically increase the `guidance_scale` and asserts that the final value is as expected. The implementation correctly accounts for the nested denoising loops present in diffusion forcing pipelines. * style * fix: Update image_encoder type to CLIPVisionModelWithProjection in SkyReelsV2ImageToVideoPipeline * UP * Add conversion support for SkyReels-V2-FLF2V models Adds configurations for three new FLF2V model variants (1.3B-540P, 14B-540P, and 14B-720P) to the conversion script. This change also introduces specific handling to zero out the image positional embeddings for these models and updates the main script to correctly initialize the image-to-video pipeline. * Docs: Update and simplify SkyReels V2 usage examples Simplifies the text-to-video example by removing the manual group offloading configuration, making it more straightforward. Adds comments to pipeline parameters to clarify their purpose and provides guidance for different resolutions and long video generation. Introduces a new section with a code example for the video-to-video pipeline. * style * docs: Add SkyReels-V2 FLF2V 1.3B model to supported models list * docs: Update SkyReels-V2 documentation * Move the initialization of the `gradient_checkpointing` attribute to its suggested location. * Refactor: Use logger for long video progress messages Replaces `print()` calls with `logger.debug()` for reporting progress during long video generation in SkyReelsV2DF pipelines. This change reduces console output verbosity for standard runs while allowing developers to view progress by enabling debug-level logging. * Refactor SkyReelsV2 timestep embedding into a module Extract the sinusoidal timestep embedding logic into a new `SkyReelsV2Timesteps` `nn.Module`. This change encapsulates the embedding generation, which simplifies the `SkyReelsV2TimeTextImageEmbedding` class and improves code modularity. * Fix: Preserve original shape in timestep embeddings Reshapes the timestep embedding tensor to match the original input shape. This ensures that batched timestep inputs retain their batch dimension after embedding, preventing potential shape mismatches. * style * Refactor: Move SkyReelsV2Timesteps to model file Colocates the `SkyReelsV2Timesteps` class with the SkyReelsV2 transformer model. This change moves model-specific timestep embedding logic from the general embeddings module to the transformer's own file, improving modularity and making the model more self-contained. * Refactor parameter dtype retrieval to use utility function Replaces manual parameter iteration with the `get_parameter_dtype` helper to determine the time embedder's data type. This change improves code readability and centralizes the logic. * Add comments to track the tensor shape transformations * Add copied froms * style * fix-copies * up * Remove FlowMatchUniPCMultistepScheduler Deletes the `FlowMatchUniPCMultistepScheduler` as it is no longer being used. * Refactor: Replace FlowMatchUniPC scheduler with UniPC Removes the `FlowMatchUniPCMultistepScheduler` and integrates its functionality into the existing `UniPCMultistepScheduler`. This consolidation is achieved by using the `use_flow_sigmas=True` parameter in `UniPCMultistepScheduler`, simplifying the scheduler API and reducing code duplication. All usages, documentation, and tests are updated accordingly. * style * Remove text_encoder parameter from SkyReelsV2DiffusionForcingPipeline initialization * Docs: Rename `pipe` to `pipeline` in SkyReels examples Updates the variable name from `pipe` to `pipeline` across all SkyReels V2 documentation examples. This change improves clarity and consistency. * Fix: Rename shift parameter to flow_shift in SkyReels-V2 examples * Fix: Rename shift parameter to flow_shift in example documentation across SkyReels-V2 files * Fix: Rename shift parameter to flow_shift in UniPCMultistepScheduler initialization across SkyReels test files * Removes unused generator argument from scheduler step The `generator` parameter is not used by the scheduler's `step` method within the SkyReelsV2 diffusion forcing pipelines. This change removes the unnecessary argument from the method call for code clarity and consistency. * Fix: Update time_embedder_dtype assignment to use the first parameter's dtype in SkyReelsV2TimeTextImageEmbedding * style * Refactor: Use get_parameter_dtype utility function Replaces manual parameter iteration with the `get_parameter_dtype` helper. * Fix: Prevent (potential) error in parameter dtype check Adds a check to ensure the `_keep_in_fp32_modules` attribute exists on a parameter before it is accessed. This prevents a potential `AttributeError`, making the utility function more robust when used with models that do not define this attribute. --------- Co-authored-by: YiYi Xu <yixu310@gmail.com> Co-authored-by: Aryan <contact.aryanvs@gmail.com>

Add SkyReels V2: Infinite-Length Film Generative Model (#11518)
* style * Fix class name casing for SkyReelsV2 components in multiple files to ensure consistency and correct functionality. * cleaning * cleansing * Refactor `get_timestep_embedding` to move modifications into `SkyReelsV2TimeTextImageEmbedding`. * Remove unnecessary line break in `get_timestep_embedding` function for cleaner code. * Remove `skyreels_v2` entry from `_import_structure` and update its initialization to directly assign the list of SkyReelsV2 components. * cleansing * Refactor attention processing in `SkyReelsV2AttnProcessor2_0` to always convert query, key, and value to `torch.bfloat16`, simplifying the code and improving clarity. * Enhance example usage in `pipeline_skyreels_v2_diffusion_forcing.py` by adding VAE initialization and detailed prompt for video generation, improving clarity and usability of the documentation. * Refactor import structure in `__init__.py` for SkyReelsV2 components and improve formatting in `pipeline_skyreels_v2_diffusion_forcing.py` to enhance code readability and maintainability. * Update `guidance_scale` parameter in `SkyReelsV2DiffusionForcingPipeline` from 5.0 to 6.0 to enhance video generation quality. * Update `guidance_scale` parameter in example documentation and class definition of `SkyReelsV2DiffusionForcingPipeline` to ensure consistency and improve video generation quality. * Update `causal_block_size` parameter in `SkyReelsV2DiffusionForcingPipeline` to default to `None`. * up * Fix dtype conversion for `timestep_proj` in `SkyReelsV2Transformer3DModel` to *ensure* correct tensor operations. * Optimize causal mask generation by replacing repeated tensor with `repeat_interleave` for improved efficiency in `SkyReelsV2Transformer3DModel`. * style * Enhance example documentation in `SkyReelsV2DiffusionForcingPipeline` with guidance scale and shift parameters for T2V and I2V. Remove unused `retrieve_latents` function to streamline the code. * Refactor sample scheduler creation in `SkyReelsV2DiffusionForcingPipeline` to use `deepcopy` for improved state management during inference steps. * Enhance error handling and documentation in `SkyReelsV2DiffusionForcingPipeline` for `overlap_history` and `addnoise_condition` parameters to improve long video generation guidance. * Update documentation and progress bar handling in `SkyReelsV2DiffusionForcingPipeline` to clarify asynchronous inference settings and improve progress tracking during denoising steps. * Refine progress bar calculation in `SkyReelsV2DiffusionForcingPipeline` by rounding the step size to one decimal place for improved readability during denoising steps. * Update import statements in `SkyReelsV2DiffusionForcingPipeline` documentation for improved clarity and organization. * Refactor progress bar handling in `SkyReelsV2DiffusionForcingPipeline` to use total steps instead of calculated step size. * update templates for i2v, v2v * Add `retrieve_latents` function to streamline latent retrieval in `SkyReelsV2DiffusionForcingPipeline`. Update video latent processing to utilize this new function for improved clarity and maintainability. * Add `retrieve_latents` function to both i2v and v2v pipelines for consistent latent retrieval. Update video latent processing to utilize this function, enhancing clarity and maintainability across the SkyReelsV2DiffusionForcingPipeline implementations. * Remove redundant ValueError for `overlap_history` in `SkyReelsV2DiffusionForcingPipeline` to streamline error handling and improve user guidance for long video generation. * Update default video dimensions and flow matching scheduler parameter in `SkyReelsV2DiffusionForcingPipeline` to enhance video generation capabilities. * Refactor `SkyReelsV2DiffusionForcingPipeline` to support Image-to-Video (i2v) generation. Update class name, add image encoding functionality, and adjust parameters for improved video generation. Enhance error handling for image inputs and update documentation accordingly. * Improve organization for image-last_image condition. * Refactor `SkyReelsV2DiffusionForcingImageToVideoPipeline` to improve latent preparation and video condition handling integration. * style * style * Add example usage of PIL for image input in `SkyReelsV2DiffusionForcingImageToVideoPipeline` documentation. * Refactor `SkyReelsV2DiffusionForcingPipeline` to `SkyReelsV2DiffusionForcingVideoToVideoPipeline`, enhancing support for Video-to-Video (v2v) generation. Introduce video input handling, update latent preparation logic, and improve error handling for input parameters. * Refactor `SkyReelsV2DiffusionForcingImageToVideoPipeline` by removing the `image_encoder` and `image_processor` dependencies. Update the CPU offload sequence accordingly. * Refactor `SkyReelsV2DiffusionForcingImageToVideoPipeline` to enhance latent preparation logic and condition handling. Update image input type to `Optional`, streamline video condition processing, and improve handling of `last_image` during latent generation. * Enhance `SkyReelsV2DiffusionForcingPipeline` by refining latent preparation for long video generation. Introduce new parameters for video handling, overlap history, and causal block size. Update logic to accommodate both short and long video scenarios, ensuring compatibility and improved processing. * refactor * fix num_frames * fix prefix_video_latents * up * refactor * Fix typo in scheduler method call within `SkyReelsV2DiffusionForcingVideoToVideoPipeline` to ensure proper noise scaling during latent generation. * up * Enhance `SkyReelsV2DiffusionForcingImageToVideoPipeline` by adding support for `last_image` parameter and refining latent frame calculations. Update preprocessing logic. * add statistics * Refine latent frame handling in `SkyReelsV2DiffusionForcingImageToVideoPipeline` by correcting variable names and reintroducing latent mean and standard deviation calculations. Update logic for frame preparation and sampling to ensure accurate video generation. * up * refactor * up * Refactor `SkyReelsV2DiffusionForcingVideoToVideoPipeline` to improve latent handling by enforcing tensor input for video, updating frame preparation logic, and adjusting default frame count. Enhance preprocessing and postprocessing steps for better integration. * style * fix vae output indexing * upup * up * Fix tensor concatenation and repetition logic in `SkyReelsV2DiffusionForcingImageToVideoPipeline` to ensure correct dimensionality for video conditions and latent conditions. * Refactor latent retrieval logic in `SkyReelsV2DiffusionForcingVideoToVideoPipeline` to handle tensor dimensions more robustly, ensuring compatibility with both 3D and 4D video inputs. * Enhance logging in `SkyReelsV2DiffusionForcing` pipelines by adding iteration print statements for better debugging. Clean up unused code related to prefix video latents length calculation in `SkyReelsV2DiffusionForcingImageToVideoPipeline`. * Update latent handling in `SkyReelsV2DiffusionForcingImageToVideoPipeline` to conditionally set latents based on video iteration state, improving flexibility for video input processing. * Refactor `SkyReelsV2TimeTextImageEmbedding` to utilize `get_1d_sincos_pos_embed_from_grid` for timestep projection. * Enhance `get_1d_sincos_pos_embed_from_grid` function to include an optional parameter `flip_sin_to_cos` for flipping sine and cosine embeddings, improving flexibility in positional embedding generation. * Update timestep projection in `SkyReelsV2TimeTextImageEmbedding` to include `flip_sin_to_cos` parameter, enhancing the flexibility of time embedding generation. * Refactor tensor type handling in `SkyReelsV2AttnProcessor2_0` and `SkyReelsV2TransformerBlock` to ensure consistent use of `torch.float32` and `torch.bfloat16`, improving integration. * Update tensor type in `SkyReelsV2RotaryPosEmbed` to use `torch.float32` for frequency calculations, ensuring consistency in data types across the model. * Refactor `SkyReelsV2TimeTextImageEmbedding` to utilize automatic mixed precision for timestep projection. * down * down * style * Add debug tensor tracking to `SkyReelsV2Transformer3DModel` for enhanced debugging and output analysis; update `Transformer2DModelOutput` to include debug tensors. * up * Refactor indentation in `SkyReelsV2AttnProcessor2_0` to improve code readability and maintain consistency in style. * Convert query, key, and value tensors to bfloat16 in `SkyReelsV2AttnProcessor2_0` for improved performance. * Add debug print statements in `SkyReelsV2TransformerBlock` to track tensor shapes and values for improved debugging and analysis. * debug * debug * Remove commented-out debug tensor tracking from `SkyReelsV2TransformerBlock` * Add functionality to save processed video latents as a Safetensors file in `SkyReelsV2DiffusionForcingPipeline`. * up * Add functionality to save output latents as a Safetensors file in `SkyReelsV2DiffusionForcingPipeline`. * up * Remove additional commented-out debug tensor tracking from `SkyReelsV2TransformerBlock` and `SkyReelsV2Transformer3DModel` for cleaner code. * style * cleansing * Update example documentation and parameters in `SkyReelsV2Pipeline`. Adjusted example code for loading models, modified default values for height, width, num_frames, and guidance_scale, and improved output video quality settings. * Update shift parameter in example documentation and default values across SkyReels V2 pipelines. Adjusted shift values for I2V from 3.0 to 5.0 and updated related example code for consistency. * Update example documentation in SkyReels V2 pipelines to include available model options and update model references for loading. Adjusted model names to reflect the latest versions across I2V, V2V, and T2V pipelines. * Add test templates * style * Add docs template * Add SkyReels V2 Diffusion Forcing Video-to-Video Pipeline to imports * style * fix-copies * convert i2v 1.3b * Update transformer configuration to include `image_dim` for SkyReels V2 models and refactor imports to use `SkyReelsV2Transformer3DModel`. * Refactor transformer import in SkyReels V2 pipeline to use `SkyReelsV2Transformer3DModel` for consistency. * Update transformer configuration in SkyReels V2 to increase `in_channels` from 16 to 36 for i2v conf. * Update transformer configuration in SkyReels V2 to set `added_kv_proj_dim` values for different model types. * up * up * up * Add SkyReelsV2Pipeline support for T2V model type in conversion script * upp * Refactor model type checks in conversion script to use substring matching for improved flexibility * upp * Fix shard path formatting in conversion script to accommodate varying model types by dynamically adjusting zero padding. * Update sharded safetensors loading logic in conversion script to use substring matching for model directory checks * Update scheduler parameters in SkyReels V2 test files for consistency across image and video pipelines * Refactor conversion script to initialize text encoder, tokenizer, and scheduler for SkyReels pipelines, enhancing model integration * style * Update documentation for SkyReels-V2, introducing the Infinite-length Film Generative model, enhancing text-to-video generation examples, and updating model references throughout the API documentation. * Add SkyReelsV2Transformer3DModel and FlowMatchUniPCMultistepScheduler documentation, updating TOC and introducing new model and scheduler files. * style * Update documentation for SkyReelsV2DiffusionForcingPipeline to correct flow matching scheduler parameter for I2V from 3.0 to 5.0, ensuring clarity in usage examples. * Add documentation for causal_block_size parameter in SkyReelsV2DF pipelines, clarifying its role in asynchronous inference. * Simplify min_ar_step calculation in SkyReelsV2DiffusionForcingPipeline to improve clarity. * style and fix-copies * style * Add documentation for SkyReelsV2Transformer3DModel Introduced a new markdown file detailing the SkyReelsV2Transformer3DModel, including usage instructions and model output specifications. * Update test configurations for SkyReelsV2 pipelines - Adjusted `in_channels` from 36 to 16 in `test_skyreels_v2_df_image_to_video.py`. - Added new parameters: `overlap_history`, `num_frames`, and `base_num_frames` in `test_skyreels_v2_df_video_to_video.py`. - Updated expected output shape in video tests from (17, 3, 16, 16) to (41, 3, 16, 16). * Refines SkyReelsV2DF test parameters * Update src/diffusers/models/modeling_outputs.py Co-authored-by: Aryan <contact.aryanvs@gmail.com> * Refactor `grid_sizes` processing by using already-calculated post-patch parameters to simplify * Update docs/source/en/api/pipelines/skyreels_v2.md Co-authored-by: Aryan <contact.aryanvs@gmail.com> * Refactor parameter naming for diffusion forcing in SkyReelsV2 pipelines - Changed `flag_df` to `enable_diffusion_forcing` for clarity in the SkyReelsV2Transformer3DModel and associated pipelines. - Updated all relevant method calls to reflect the new parameter name. * Revert _toctree.yml to adjust section expansion states * style * Update docs/source/en/api/models/skyreels_v2_transformer_3d.md Co-authored-by: YiYi Xu <yixu310@gmail.com> * Add copying label to SkyReelsV2ImageEmbedding from WanImageEmbedding. * Refactor transformer block processing in SkyReelsV2Transformer3DModel - Ensured proper handling of hidden states during both gradient checkpointing and standard processing. * Update SkyReels V2 documentation to remove VRAM requirement and streamline imports - Removed the mention of ~13GB VRAM requirement for the SkyReels-V2 model. - Simplified import statements by removing unused `load_image` import. * Add SkyReelsV2LoraLoaderMixin for loading and managing LoRA layers in SkyReelsV2Transformer3DModel - Introduced SkyReelsV2LoraLoaderMixin class to handle loading, saving, and fusing of LoRA weights specific to the SkyReelsV2 model. - Implemented methods for state dict management, including compatibility checks for various LoRA formats. - Enhanced functionality for loading weights with options for low CPU memory usage and hotswapping. - Added detailed docstrings for clarity on parameters and usage. * Update SkyReelsV2 documentation and loader mixin references - Corrected the documentation to reference the new `SkyReelsV2LoraLoaderMixin` for loading LoRA weights. - Updated comments in the `SkyReelsV2LoraLoaderMixin` class to reflect changes in model references from `WanTransformer3DModel` to `SkyReelsV2Transformer3DModel`. * Enhance SkyReelsV2 integration by adding SkyReelsV2LoraLoaderMixin references - Added `SkyReelsV2LoraLoaderMixin` to the documentation and loader imports for improved LoRA weight management. - Updated multiple pipeline classes to inherit from `SkyReelsV2LoraLoaderMixin` instead of `WanLoraLoaderMixin`. * Update SkyReelsV2 model references in documentation - Replaced placeholder model paths with actual paths for SkyReels-V2 models in multiple pipeline files. - Ensured consistency across the documentation for loading models in the SkyReelsV2 pipelines. * style * fix-copies * Refactor `fps_projection` in `SkyReelsV2Transformer3DModel` - Replaced the sequential linear layers for `fps_projection` with a `FeedForward` layer using `SiLU` activation for better integration. * Update docs * Refactor video processing in SkyReelsV2DiffusionForcingPipeline - Renamed parameters for clarity: `video` to `video_latents` and `overlap_history` to `overlap_history_latent_frames`. - Updated logic for handling long video generation, including adjustments to latent frame calculations and accumulation. - Consolidated handling of latents for both long and short video generation scenarios. - Final decoding step now consistently converts latents to pixels, ensuring proper output format. * Update activation function in `fps_projection` of `SkyReelsV2Transformer3DModel` - Changed activation function from `silu` to `linear-silu` in the `fps_projection` layer for improved performance and integration. * Add fps_projection layer renaming in convert_skyreelsv2_to_diffusers.py - Updated key mappings for the `fps_projection` layer to align with new naming conventions, ensuring consistency in model integration. * Fix fps_projection assignment in SkyReelsV2Transformer3DModel - Corrected the assignment of the `fps_projection` layer to ensure it is properly cast to the appropriate data type, enhancing model functionality. * Update _keep_in_fp32_modules in SkyReelsV2Transformer3DModel - Added `fps_projection` to the list of modules that should remain in FP32 precision, ensuring proper handling of data types during model operations. * Remove integration test classes from SkyReelsV2 test files - Deleted the `SkyReelsV2DiffusionForcingPipelineIntegrationTests` and `SkyReelsV2PipelineIntegrationTests` classes along with their associated setup, teardown, and test methods, as they were not implemented and not needed for current testing. * style * Refactor: Remove hardcoded `torch.bfloat16` cast in attention * Refactor: Simplify data type handling in transformer model Removes unnecessary data type conversions for the FPS embedding and timestep projection. This change simplifies the forward pass by relying on the inherent data types of the tensors. * Refactor: Remove `fps_projection` from `_keep_in_fp32_modules` in `SkyReelsV2Transformer3DModel` * Update src/diffusers/models/transformers/transformer_skyreels_v2.py Co-authored-by: Aryan <contact.aryanvs@gmail.com> * Refactor: Remove unused flags and simplify attention mask handling in SkyReelsV2AttnProcessor2_0 and SkyReelsV2Transformer3DModel Refactor: Simplify causal attention logic in SkyReelsV2 Removes the `flag_causal_attention` and `_flag_ar_attention` flags to simplify the implementation. The decision to apply a causal attention mask is now based directly on the `num_frame_per_block` configuration, eliminating redundant flags and conditional checks. This streamlines the attention mechanism and simplifies the `set_ar_attention` methods. * Refactor: Clarify variable names for latent frames Renames `base_num_frames` to `base_latent_num_frames` to make it explicit that the variable refers to the number of frames in the latent space. This change improves code readability and reduces potential confusion between latent frames and decoded video frames. The `num_frames` parameter in `generate_timestep_matrix` is also renamed to `num_latent_frames` for consistency. * Enhance documentation: Add detailed docstring for timestep matrix generation in SkyReelsV2DiffusionForcingPipeline * Docs: Clarify long video chunking in pipeline docstring Improves the explanation of long video processing within the pipeline's docstring. The update replaces the abstract description with a concrete example, illustrating how the sliding window mechanism works with overlapping chunks. This makes the roles of `base_num_frames` and `overlap_history` clearer for users. * Docs: Move visual demonstration and processing details for SkyReelsV2DiffusionForcingPipeline to docs page from the code * Docs: Update asynchronous processing timeline and examples for long video handling in SkyReels-V2 documentation * Enhance timestep matrix generation documentation and logic for synchronous/asynchronous video processing * Update timestep matrix documentation and enhance analysis for clarity in SkyReelsV2DiffusionForcingPipeline * Docs: Update visual demonstration section and add detailed step matrix construction example for asynchronous processing in SkyReelsV2DiffusionForcingPipeline * style * fix-copies * Refactor parameter names for clarity in SkyReelsV2DiffusionForcingImageToVideoPipeline and SkyReelsV2DiffusionForcingVideoToVideoPipeline * Refactor: Avoid VAE roundtrip in long video generation Improves performance and quality for long video generation by operating entirely in latent space during the iterative generation process. Instead of decoding latents to video and then re-encoding the overlapping section for the next chunk, this change passes the generated latents directly between iterations. This avoids a computationally expensive and potentially lossy VAE decode/encode cycle within the loop. The full video is now decoded only once from the accumulated latents at the end of the process. * Refactor: Rename prefix_video_latents_length to prefix_video_latents_frames for clarity * Refactor: Rename num_latent_frames to current_num_latent_frames for clarity in SkyReelsV2DiffusionForcingImageToVideoPipeline * Refactor: Enhance long video generation logic and improve latent handling in SkyReelsV2DiffusionForcingImageToVideoPipeline Refactor: Unify video generation and pass latents directly Unifies the separate code paths for short and long video generation into a single, streamlined loop. This change eliminates the inefficient decode-encode cycle during long video generation. Instead of converting latents to pixel-space video between chunks, the pipeline now passes the generated latents directly to the next iteration. This improves performance, avoids potential quality loss from intermediate VAE steps, and enhances code maintainability by removing significant duplication. * style * Refactor: Remove overlap_history parameter and streamline long video generation logic in SkyReelsV2DiffusionForcingImageToVideoPipeline Refactor: Streamline long video generation logic Removes the `overlap_history` parameter and simplifies the conditioning process for long video generation. This change avoids a redundant VAE encoding step by directly using latent frames from the previous chunk for conditioning. It also moves image preprocessing outside the main generation loop to prevent repeated computations and clarifies the handling of prefix latents. * style * Refactor latent handling in i2v diffusion forcing pipeline Improves the latent conditioning and accumulation logic within the image-to-video diffusion forcing loop. - Corrects the splitting of the initial conditioning tensor to robustly handle both even and odd lengths. - Simplifies how latents are accumulated across iterations for long video generation. - Ensures the final latents are trimmed correctly before decoding only when a `last_image` is provided. * Refactor: Remove overlap_history parameter from SkyReelsV2DiffusionForcingImageToVideoPipeline * Refactor: Adjust video_latents parameter handling in prepare_latents method * style * Refactor: Update long video iteration print statements for clarity * Fix: Update transformer config with dynamic causal block size Updates the SkyReelsV2 pipelines to correctly set the `causal_block_size` in the transformer's configuration when it's provided during a pipeline call. This ensures the model configuration reflects the user's specified setting for the inference run. The `set_ar_attention` method is also renamed to `_set_ar_attention` to mark it as an internal helper. * style * Refactor: Adjust video input size and expected output shape in inference test * Refactor: Rename video variables for clarity in SkyReelsV2DiffusionForcingVideoToVideoPipeline * Docs: Clarify time embedding logic in SkyReelsV2 Adds comments to explain the handling of different time embedding tensor dimensions. A 2D tensor is used for standard models with a single time embedding per batch, while a 3D tensor is used for Diffusion Forcing models where each frame has its own time embedding. This clarifies the expected input for different model variations. * Docs: Update SkyReels V2 pipeline examples Updates the docstring examples for the SkyReels V2 pipelines to reflect current best practices and API changes. - Removes the `shift` parameter from pipeline call examples, as it is now configured directly on the scheduler. - Replaces the `set_ar_attention` method call with the `causal_block_size` argument in the pipeline call for diffusion forcing examples. - Adjusts recommended parameters for I2V and V2V examples, including inference steps, guidance scale, and `ar_step`. * Refactor: Remove `shift` parameter from SkyReelsV2 pipelines Removes the `shift` parameter from the call signature of all SkyReelsV2 pipelines. This parameter is a scheduler-specific configuration and should be set directly on the scheduler during its initialization, rather than being passed at runtime through the pipeline. This change simplifies the pipeline API. Usage examples are updated to reflect that the `shift` value should now be passed when creating the `FlowMatchUniPCMultistepScheduler`. * Refactors SkyReelsV2 image-to-video tests and adds last image case Simplifies the test suite by removing a duplicated test class and streamlining the dummy component and input generation. Adds a new test to verify the pipeline's behavior when a `last_image` is provided as input for conditioning. * test: Add image components to SkyReelsV2 pipeline test Adds the `image_encoder` and `image_processor` to the test components for the image-to-video pipeline. Also replaces a hardcoded value for the positional embedding sequence length with a more descriptive calculation, improving clarity. * test: Add callback configuration test for SkyReelsV2DiffusionForcingVideoToVideoPipeline test: Add callback test for SkyReelsV2DFV2V pipeline Adds a test to validate the callback functionality for the `SkyReelsV2DiffusionForcingVideoToVideoPipeline`. This test confirms that `callback_on_step_end` is invoked correctly and can modify the pipeline's state during inference. It uses a callback to dynamically increase the `guidance_scale` and asserts that the final value is as expected. The implementation correctly accounts for the nested denoising loops present in diffusion forcing pipelines. * style * fix: Update image_encoder type to CLIPVisionModelWithProjection in SkyReelsV2ImageToVideoPipeline * UP * Add conversion support for SkyReels-V2-FLF2V models Adds configurations for three new FLF2V model variants (1.3B-540P, 14B-540P, and 14B-720P) to the conversion script. This change also introduces specific handling to zero out the image positional embeddings for these models and updates the main script to correctly initialize the image-to-video pipeline. * Docs: Update and simplify SkyReels V2 usage examples Simplifies the text-to-video example by removing the manual group offloading configuration, making it more straightforward. Adds comments to pipeline parameters to clarify their purpose and provides guidance for different resolutions and long video generation. Introduces a new section with a code example for the video-to-video pipeline. * style * docs: Add SkyReels-V2 FLF2V 1.3B model to supported models list * docs: Update SkyReels-V2 documentation * Move the initialization of the `gradient_checkpointing` attribute to its suggested location. * Refactor: Use logger for long video progress messages Replaces `print()` calls with `logger.debug()` for reporting progress during long video generation in SkyReelsV2DF pipelines. This change reduces console output verbosity for standard runs while allowing developers to view progress by enabling debug-level logging. * Refactor SkyReelsV2 timestep embedding into a module Extract the sinusoidal timestep embedding logic into a new `SkyReelsV2Timesteps` `nn.Module`. This change encapsulates the embedding generation, which simplifies the `SkyReelsV2TimeTextImageEmbedding` class and improves code modularity. * Fix: Preserve original shape in timestep embeddings Reshapes the timestep embedding tensor to match the original input shape. This ensures that batched timestep inputs retain their batch dimension after embedding, preventing potential shape mismatches. * style * Refactor: Move SkyReelsV2Timesteps to model file Colocates the `SkyReelsV2Timesteps` class with the SkyReelsV2 transformer model. This change moves model-specific timestep embedding logic from the general embeddings module to the transformer's own file, improving modularity and making the model more self-contained. * Refactor parameter dtype retrieval to use utility function Replaces manual parameter iteration with the `get_parameter_dtype` helper to determine the time embedder's data type. This change improves code readability and centralizes the logic. * Add comments to track the tensor shape transformations * Add copied froms * style * fix-copies * up * Remove FlowMatchUniPCMultistepScheduler Deletes the `FlowMatchUniPCMultistepScheduler` as it is no longer being used. * Refactor: Replace FlowMatchUniPC scheduler with UniPC Removes the `FlowMatchUniPCMultistepScheduler` and integrates its functionality into the existing `UniPCMultistepScheduler`. This consolidation is achieved by using the `use_flow_sigmas=True` parameter in `UniPCMultistepScheduler`, simplifying the scheduler API and reducing code duplication. All usages, documentation, and tests are updated accordingly. * style * Remove text_encoder parameter from SkyReelsV2DiffusionForcingPipeline initialization * Docs: Rename `pipe` to `pipeline` in SkyReels examples Updates the variable name from `pipe` to `pipeline` across all SkyReels V2 documentation examples. This change improves clarity and consistency. * Fix: Rename shift parameter to flow_shift in SkyReels-V2 examples * Fix: Rename shift parameter to flow_shift in example documentation across SkyReels-V2 files * Fix: Rename shift parameter to flow_shift in UniPCMultistepScheduler initialization across SkyReels test files * Removes unused generator argument from scheduler step The `generator` parameter is not used by the scheduler's `step` method within the SkyReelsV2 diffusion forcing pipelines. This change removes the unnecessary argument from the method call for code clarity and consistency. * Fix: Update time_embedder_dtype assignment to use the first parameter's dtype in SkyReelsV2TimeTextImageEmbedding * style * Refactor: Use get_parameter_dtype utility function Replaces manual parameter iteration with the `get_parameter_dtype` helper. * Fix: Prevent (potential) error in parameter dtype check Adds a check to ensure the `_keep_in_fp32_modules` attribute exists on a parameter before it is accessed. This prevents a potential `AttributeError`, making the utility function more robust when used with models that do not define this attribute. --------- Co-authored-by: YiYi Xu <yixu310@gmail.com> Co-authored-by: Aryan <contact.aryanvs@gmail.com>
7298bdd8 · Tolga Cangöz · GitHub · 9c13f865 · 7298bdd8 · 7298bdd8
Unverified Commit 7298bdd8 authored Jul 16, 2025 by Tolga Cangöz Committed by GitHub Jul 16, 2025
20 changed files
--- a/docs/source/en/_toctree.yml
+++ b/docs/source/en/_toctree.yml
@@ -353,6 +353,8 @@
        title: SanaTransformer2DModel
      - local: api/models/sd3_transformer2d
        title: SD3Transformer2DModel
+      - local: api/models/skyreels_v2_transformer_3d
+        title: SkyReelsV2Transformer3DModel
      - local: api/models/stable_audio_transformer
        title: StableAudioDiTModel
      - local: api/models/transformer2d
@@ -547,6 +549,8 @@
      title: Semantic Guidance
    - local: api/pipelines/shap_e
      title: Shap-E
+    - local: api/pipelines/skyreels_v2
+      title: SkyReels-V2
    - local: api/pipelines/stable_audio
      title: Stable Audio
    - local: api/pipelines/stable_cascade

--- a/docs/source/en/api/loaders/lora.md
+++ b/docs/source/en/api/loaders/lora.md
@@ -26,6 +26,7 @@ LoRA is a fast and lightweight training method that inserts and trains a signifi
 - [`HunyuanVideoLoraLoaderMixin`] provides similar functions for [HunyuanVideo](https://huggingface.co/docs/diffusers/main/en/api/pipelines/hunyuan_video).
 - [`Lumina2LoraLoaderMixin`] provides similar functions for [Lumina2](https://huggingface.co/docs/diffusers/main/en/api/pipelines/lumina2).
 - [`WanLoraLoaderMixin`] provides similar functions for [Wan](https://huggingface.co/docs/diffusers/main/en/api/pipelines/wan).
+- [`SkyReelsV2LoraLoaderMixin`] provides similar functions for [SkyReels-V2](https://huggingface.co/docs/diffusers/main/en/api/pipelines/skyreels_v2).
 - [`CogView4LoraLoaderMixin`] provides similar functions for [CogView4](https://huggingface.co/docs/diffusers/main/en/api/pipelines/cogview4).
 - [`AmusedLoraLoaderMixin`] is for the [`AmusedPipeline`].
 - [`HiDreamImageLoraLoaderMixin`] provides similar functions for [HiDream Image](https://huggingface.co/docs/diffusers/main/en/api/pipelines/hidream)
@@ -92,6 +93,10 @@ To learn more about how to load LoRA weights, see the [LoRA](../../using-diffuse
 [[autodoc]] loaders.lora_pipeline.WanLoraLoaderMixin
+## SkyReelsV2LoraLoaderMixin
+[[autodoc]] loaders.lora_pipeline.SkyReelsV2LoraLoaderMixin
 ## AmusedLoraLoaderMixin
 [[autodoc]] loaders.lora_pipeline.AmusedLoraLoaderMixin
@@ -100,6 +105,6 @@ To learn more about how to load LoRA weights, see the [LoRA](../../using-diffuse
 [[autodoc]] loaders.lora_pipeline.HiDreamImageLoraLoaderMixin
-## WanLoraLoaderMixin
+## LoraBaseMixin
-[[autodoc]] loaders.lora_pipeline.WanLoraLoaderMixin
+[[autodoc]] loaders.lora_base.LoraBaseMixin
\ No newline at end of file
--- a/docs/source/en/api/models/skyreels_v2_transformer_3d.md
+++ b/docs/source/en/api/models/skyreels_v2_transformer_3d.md
+<!-- Copyright 2024 The HuggingFace Team. All rights reserved.
+Licensed under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with
+the License. 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. -->
+# SkyReelsV2Transformer3DModel
+A Diffusion Transformer model for 3D video-like data was introduced in [SkyReels-V2](https://github.com/SkyworkAI/SkyReels-V2) by the Skywork AI.
+The model can be loaded with the following code snippet.
+```python
+from diffusers import SkyReelsV2Transformer3DModel
+transformer = SkyReelsV2Transformer3DModel.from_pretrained("Skywork/SkyReels-V2-DF-1.3B-540P-Diffusers", subfolder="transformer", torch_dtype=torch.bfloat16)
+```
+## SkyReelsV2Transformer3DModel
+[[autodoc]] SkyReelsV2Transformer3DModel
+## Transformer2DModelOutput
+[[autodoc]] models.modeling_outputs.Transformer2DModelOutput
--- a/docs/source/en/api/pipelines/skyreels_v2.md
+++ b/docs/source/en/api/pipelines/skyreels_v2.md
+<!-- Copyright 2024 The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# 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. -->
+<div style="float: right;">
+  <div class="flex flex-wrap space-x-1">
+    <a href="https://huggingface.co/docs/diffusers/main/en/tutorials/using_peft_for_inference" target="_blank" rel="noopener">
+      <img alt="LoRA" src="https://img.shields.io/badge/LoRA-d8b4fe?style=flat"/>
+    </a>
+  </div>
+</div>
+# SkyReels-V2: Infinite-length Film Generative model
+[SkyReels-V2](https://huggingface.co/papers/2504.13074) by the SkyReels Team.
+*Recent advances in video generation have been driven by diffusion models and autoregressive frameworks, yet critical challenges persist in harmonizing prompt adherence, visual quality, motion dynamics, and duration: compromises in motion dynamics to enhance temporal visual quality, constrained video duration (5-10 seconds) to prioritize resolution, and inadequate shot-aware generation stemming from general-purpose MLLMs' inability to interpret cinematic grammar, such as shot composition, actor expressions, and camera motions. These intertwined limitations hinder realistic long-form synthesis and professional film-style generation. To address these limitations, we propose SkyReels-V2, an Infinite-length Film Generative Model, that synergizes Multi-modal Large Language Model (MLLM), Multi-stage Pretraining, Reinforcement Learning, and Diffusion Forcing Framework. Firstly, we design a comprehensive structural representation of video that combines the general descriptions by the Multi-modal LLM and the detailed shot language by sub-expert models. Aided with human annotation, we then train a unified Video Captioner, named SkyCaptioner-V1, to efficiently label the video data. Secondly, we establish progressive-resolution pretraining for the fundamental video generation, followed by a four-stage post-training enhancement: Initial concept-balanced Supervised Fine-Tuning (SFT) improves baseline quality; Motion-specific Reinforcement Learning (RL) training with human-annotated and synthetic distortion data addresses dynamic artifacts; Our diffusion forcing framework with non-decreasing noise schedules enables long-video synthesis in an efficient search space; Final high-quality SFT refines visual fidelity. All the code and models are available at [this https URL](https://github.com/SkyworkAI/SkyReels-V2).*
+You can find all the original SkyReels-V2 checkpoints under the [Skywork](https://huggingface.co/collections/Skywork/skyreels-v2-6801b1b93df627d441d0d0d9) organization.
+The following SkyReels-V2 models are supported in Diffusers:
+- [SkyReels-V2 DF 1.3B - 540P](https://huggingface.co/Skywork/SkyReels-V2-DF-1.3B-540P-Diffusers)
+- [SkyReels-V2 DF 14B - 540P](https://huggingface.co/Skywork/SkyReels-V2-DF-14B-540P-Diffusers)
+- [SkyReels-V2 DF 14B - 720P](https://huggingface.co/Skywork/SkyReels-V2-DF-14B-720P-Diffusers)
+- [SkyReels-V2 T2V 14B - 540P](https://huggingface.co/Skywork/SkyReels-V2-T2V-14B-540P-Diffusers)
+- [SkyReels-V2 T2V 14B - 720P](https://huggingface.co/Skywork/SkyReels-V2-T2V-14B-720P-Diffusers)
+- [SkyReels-V2 I2V 1.3B - 540P](https://huggingface.co/Skywork/SkyReels-V2-I2V-1.3B-540P-Diffusers)
+- [SkyReels-V2 I2V 14B - 540P](https://huggingface.co/Skywork/SkyReels-V2-I2V-14B-540P-Diffusers)
+- [SkyReels-V2 I2V 14B - 720P](https://huggingface.co/Skywork/SkyReels-V2-I2V-14B-720P-Diffusers)
+- [SkyReels-V2 FLF2V 1.3B - 540P](https://huggingface.co/Skywork/SkyReels-V2-FLF2V-1.3B-540P-Diffusers)
+> [!TIP]
+> Click on the SkyReels-V2 models in the right sidebar for more examples of video generation.
+### A _Visual_ Demonstration
+        An example with these parameters:
+        base_num_frames=97, num_frames=97, num_inference_steps=30, ar_step=5, causal_block_size=5
+        vae_scale_factor_temporal -> 4
+        num_latent_frames: (97-1)//vae_scale_factor_temporal+1 = 25 frames -> 5 blocks of 5 frames each
+        base_num_latent_frames = (97-1)//vae_scale_factor_temporal+1 = 25 → blocks = 25//5 = 5 blocks
+        This 5 blocks means the maximum context length of the model is 25 frames in the latent space.
+        Asynchronous Processing Timeline:
+        ┌─────────────────────────────────────────────────────────────────┐
+        │ Steps:    1    6   11   16   21   26   31   36   41   46   50   │
+        │ Block 1: [■■■■■■■■■■■■■■■■■■■■■■■■■■■■■■]                       │
+        │ Block 2:      [■■■■■■■■■■■■■■■■■■■■■■■■■■■■■■]                  │
+        │ Block 3:           [■■■■■■■■■■■■■■■■■■■■■■■■■■■■■■]             │
+        │ Block 4:                [■■■■■■■■■■■■■■■■■■■■■■■■■■■■■■]        │
+        │ Block 5:                     [■■■■■■■■■■■■■■■■■■■■■■■■■■■■■■]   │
+        └─────────────────────────────────────────────────────────────────┘
+        For Long Videos (num_frames > base_num_frames):
+        base_num_frames acts as the "sliding window size" for processing long videos.
+        Example: 257-frame video with base_num_frames=97, overlap_history=17
+        ┌──── Iteration 1 (frames 1-97) ────┐
+        │ Processing window: 97 frames      │ → 5 blocks, async processing
+        │ Generates: frames 1-97            │
+        └───────────────────────────────────┘
+                    ┌────── Iteration 2 (frames 81-177) ──────┐
+                    │ Processing window: 97 frames            │
+                    │ Overlap: 17 frames (81-97) from prev    │ → 5 blocks, async processing
+                    │ Generates: frames 98-177                │
+                    └─────────────────────────────────────────┘
+                                ┌────── Iteration 3 (frames 161-257) ──────┐
+                                │ Processing window: 97 frames             │
+                                │ Overlap: 17 frames (161-177) from prev   │ → 5 blocks, async processing
+                                │ Generates: frames 178-257                │
+                                └──────────────────────────────────────────┘
+        Each iteration independently runs the asynchronous processing with its own 5 blocks.
+        base_num_frames controls:
+        1. Memory usage (larger window = more VRAM)
+        2. Model context length (must match training constraints)
+        3. Number of blocks per iteration (base_num_latent_frames // causal_block_size)
+        Each block takes 30 steps to complete denoising.
+        Block N starts at step: 1 + (N-1) x ar_step
+        Total steps: 30 + (5-1) x 5 = 50 steps
+        Synchronous mode (ar_step=0) would process all blocks/frames simultaneously:
+        ┌──────────────────────────────────────────────┐
+        │ Steps:       1            ...            30  │
+        │ All blocks: [■■■■■■■■■■■■■■■■■■■■■■■■■■■■■■] │
+        └──────────────────────────────────────────────┘
+        Total steps: 30 steps
+        An example on how the step matrix is constructed for asynchronous processing:
+        Given the parameters: (num_inference_steps=30, flow_shift=8, num_frames=97, ar_step=5, causal_block_size=5)
+        - num_latent_frames = (97 frames - 1) // (4 temporal downsampling) + 1 = 25
+        - step_template = [999, 995, 991, 986, 980, 975, 969, 963, 956, 948,
+                           941, 932, 922, 912, 901, 888, 874, 859, 841, 822,
+                           799, 773, 743, 708, 666, 615, 551, 470, 363, 216]
+        The algorithm creates a 50x25 step_matrix where:
+        - Row 1:  [999, 999, 999, 999, 999, 999, 999, 999, 999, 999, 999, 999, 999, 999, 999, 999, 999, 999, 999, 999, 999, 999, 999, 999, 999]
+        - Row 2:  [995, 995, 995, 995, 995, 999, 999, 999, 999, 999, 999, 999, 999, 999, 999, 999, 999, 999, 999, 999, 999, 999, 999, 999, 999]
+        - Row 3:  [991, 991, 991, 991, 991, 999, 999, 999, 999, 999, 999, 999, 999, 999, 999, 999, 999, 999, 999, 999, 999, 999, 999, 999, 999]
+        - ...
+        - Row 7:  [969, 969, 969, 969, 969, 995, 995, 995, 995, 995, 999, 999, 999, 999, 999, 999, 999, 999, 999, 999, 999, 999, 999, 999, 999]
+        - ...
+        - Row 21: [799, 799, 799, 799, 799, 888, 888, 888, 888, 888, 941, 941, 941, 941, 941, 975, 975, 975, 975, 975, 999, 999, 999, 999, 999]
+        - ...
+        - Row 35: [  0,   0,   0,   0,   0, 216, 216, 216, 216, 216, 666, 666, 666, 666, 666, 822, 822, 822, 822, 822, 901, 901, 901, 901, 901]
+        - ...
+        - Row 42: [  0,   0,   0,   0,   0,   0,   0,   0,   0,   0,   0,   0,   0,   0,   0, 551, 551, 551, 551, 551, 773, 773, 773, 773, 773]
+        - ...
+        - Row 50: [  0,   0,   0,   0,   0,   0,   0,   0,   0,   0,   0,   0,   0,   0,   0,   0,   0,   0,   0,   0, 216, 216, 216, 216, 216]
+        Detailed Row 6 Analysis:
+        - step_matrix[5]:       [ 975, 975, 975, 975, 975, 999, 999, 999, 999, 999, 999,  ...,  999]
+        - step_index[5]:        [   6,   6,   6,   6,   6,   1,   1,   1,   1,   1,   0,  ...,    0]
+        - step_update_mask[5]:  [True,True,True,True,True,True,True,True,True,True,False, ...,False]
+        - valid_interval[5]:    (0, 25)
+        Key Pattern: Block i lags behind Block i-1 by exactly ar_step=5 timesteps, creating the
+        staggered "diffusion forcing" effect where later blocks condition on cleaner earlier blocks.
+### Text-to-Video Generation
+The example below demonstrates how to generate a video from text.
+<hfoptions id="T2V usage">
+<hfoption id="T2V memory">
+Refer to the [Reduce memory usage](../../optimization/memory) guide for more details about the various memory saving techniques.
+From the original repo:
+>You can use --ar_step 5 to enable asynchronous inference. When asynchronous inference, --causal_block_size 5 is recommended while it is not supposed to be set for synchronous generation... Asynchronous inference will take more steps to diffuse the whole sequence which means it will be SLOWER than synchronous mode. In our experiments, asynchronous inference may improve the instruction following and visual consistent performance.
+```py
+# pip install ftfy
+import torch
+from diffusers import AutoModel, SkyReelsV2DiffusionForcingPipeline, UniPCMultistepScheduler
+from diffusers.utils import export_to_video
+vae = AutoModel.from_pretrained("Skywork/SkyReels-V2-DF-14B-540P-Diffusers", subfolder="vae", torch_dtype=torch.float32)
+transformer = AutoModel.from_pretrained("Skywork/SkyReels-V2-DF-14B-540P-Diffusers", subfolder="transformer", torch_dtype=torch.bfloat16)
+pipeline = SkyReelsV2DiffusionForcingPipeline.from_pretrained(
+    "Skywork/SkyReels-V2-DF-14B-540P-Diffusers",
+    vae=vae,
+    transformer=transformer,
+    torch_dtype=torch.bfloat16
+)
+flow_shift = 8.0  # 8.0 for T2V, 5.0 for I2V
+pipeline.scheduler = UniPCMultistepScheduler.from_config(pipeline.scheduler.config, flow_shift=flow_shift)
+pipeline = pipeline.to("cuda")
+prompt = "A cat and a dog baking a cake together in a kitchen. The cat is carefully measuring flour, while the dog is stirring the batter with a wooden spoon. The kitchen is cozy, with sunlight streaming through the window."
+output = pipeline(
+    prompt=prompt,
+    num_inference_steps=30,
+    height=544,  # 720 for 720P
+    width=960,   # 1280 for 720P
+    num_frames=97,
+    base_num_frames=97,  # 121 for 720P
+    ar_step=5,  # Controls asynchronous inference (0 for synchronous mode)
+    causal_block_size=5,  # Number of frames in each block for asynchronous processing
+    overlap_history=None,  # Number of frames to overlap for smooth transitions in long videos; 17 for long video generations
+    addnoise_condition=20,  # Improves consistency in long video generation
+).frames[0]
+export_to_video(output, "T2V.mp4", fps=24, quality=8)
+```
+</hfoption>
+</hfoptions>
+### First-Last-Frame-to-Video Generation
+The example below demonstrates how to use the image-to-video pipeline to generate a video using a text description, a starting frame, and an ending frame.
+<hfoptions id="FLF2V usage">
+<hfoption id="usage">
+```python
+import numpy as np
+import torch
+import torchvision.transforms.functional as TF
+from diffusers import AutoencoderKLWan, SkyReelsV2DiffusionForcingImageToVideoPipeline, UniPCMultistepScheduler
+from diffusers.utils import export_to_video, load_image
+model_id = "Skywork/SkyReels-V2-DF-14B-720P-Diffusers"
+vae = AutoencoderKLWan.from_pretrained(model_id, subfolder="vae", torch_dtype=torch.float32)
+pipeline = SkyReelsV2DiffusionForcingImageToVideoPipeline.from_pretrained(
+    model_id, vae=vae, torch_dtype=torch.bfloat16
+)
+flow_shift = 5.0  # 8.0 for T2V, 5.0 for I2V
+pipeline.scheduler = UniPCMultistepScheduler.from_config(pipeline.scheduler.config, flow_shift=flow_shift)
+pipeline.to("cuda")
+first_frame = load_image("https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/diffusers/flf2v_input_first_frame.png")
+last_frame = load_image("https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/diffusers/flf2v_input_last_frame.png")
+def aspect_ratio_resize(image, pipeline, max_area=720 * 1280):
+    aspect_ratio = image.height / image.width
+    mod_value = pipeline.vae_scale_factor_spatial * pipeline.transformer.config.patch_size[1]
+    height = round(np.sqrt(max_area * aspect_ratio)) // mod_value * mod_value
+    width = round(np.sqrt(max_area / aspect_ratio)) // mod_value * mod_value
+    image = image.resize((width, height))
+    return image, height, width
+def center_crop_resize(image, height, width):
+    # Calculate resize ratio to match first frame dimensions
+    resize_ratio = max(width / image.width, height / image.height)
+    # Resize the image
+    width = round(image.width * resize_ratio)
+    height = round(image.height * resize_ratio)
+    size = [width, height]
+    image = TF.center_crop(image, size)
+    return image, height, width
+first_frame, height, width = aspect_ratio_resize(first_frame, pipeline)
+if last_frame.size != first_frame.size:
+    last_frame, _, _ = center_crop_resize(last_frame, height, width)
+prompt = "CG animation style, a small blue bird takes off from the ground, flapping its wings. The bird's feathers are delicate, with a unique pattern on its chest. The background shows a blue sky with white clouds under bright sunshine. The camera follows the bird upward, capturing its flight and the vastness of the sky from a close-up, low-angle perspective."
+output = pipeline(
+    image=first_frame, last_image=last_frame, prompt=prompt, height=height, width=width, guidance_scale=5.0
+).frames[0]
+export_to_video(output, "output.mp4", fps=24, quality=8)
+```
+</hfoption>
+</hfoptions>
+### Video-to-Video Generation
+<hfoptions id="V2V usage">
+<hfoption id="usage">
+`SkyReelsV2DiffusionForcingVideoToVideoPipeline` extends a given video.
+```python
+import numpy as np
+import torch
+import torchvision.transforms.functional as TF
+from diffusers import AutoencoderKLWan, SkyReelsV2DiffusionForcingVideoToVideoPipeline, UniPCMultistepScheduler
+from diffusers.utils import export_to_video, load_video
+model_id = "Skywork/SkyReels-V2-DF-14B-540P-Diffusers"
+vae = AutoencoderKLWan.from_pretrained(model_id, subfolder="vae", torch_dtype=torch.float32)
+pipeline = SkyReelsV2DiffusionForcingVideoToVideoPipeline.from_pretrained(
+    model_id, vae=vae, torch_dtype=torch.bfloat16
+)
+flow_shift = 5.0  # 8.0 for T2V, 5.0 for I2V
+pipeline.scheduler = UniPCMultistepScheduler.from_config(pipeline.scheduler.config, flow_shift=flow_shift)
+pipeline.to("cuda")
+video = load_video("input_video.mp4")
+prompt = "CG animation style, a small blue bird takes off from the ground, flapping its wings. The bird's feathers are delicate, with a unique pattern on its chest. The background shows a blue sky with white clouds under bright sunshine. The camera follows the bird upward, capturing its flight and the vastness of the sky from a close-up, low-angle perspective."
+output = pipeline(
+    video=video, prompt=prompt, height=544, width=960, guidance_scale=5.0,
+    num_inference_steps=30, num_frames=257, base_num_frames=97#, ar_step=5, causal_block_size=5,
+).frames[0]
+export_to_video(output, "output.mp4", fps=24, quality=8)
+# Total frames will be the number of frames of given video + 257
+```
+</hfoption>
+</hfoptions>
+## Notes
+- SkyReels-V2 supports LoRAs with [`~loaders.SkyReelsV2LoraLoaderMixin.load_lora_weights`].
+  <details>
+  <summary>Show example code</summary>
+  ```py
+  # pip install ftfy
+  import torch
+  from diffusers import AutoModel, SkyReelsV2DiffusionForcingPipeline
+  from diffusers.utils import export_to_video
+  vae = AutoModel.from_pretrained(
+      "Skywork/SkyReels-V2-DF-1.3B-540P-Diffusers", subfolder="vae", torch_dtype=torch.float32
+  )
+  pipeline = SkyReelsV2DiffusionForcingPipeline.from_pretrained(
+      "Skywork/SkyReels-V2-DF-1.3B-540P-Diffusers", vae=vae, torch_dtype=torch.bfloat16
+  )
+  pipeline.to("cuda")
+  pipeline.load_lora_weights("benjamin-paine/steamboat-willie-1.3b", adapter_name="steamboat-willie")
+  pipeline.set_adapters("steamboat-willie")
+  pipeline.enable_model_cpu_offload()
+  # use "steamboat willie style" to trigger the LoRA
+  prompt = """
+  steamboat willie style, golden era animation, The camera rushes from far to near in a low-angle shot,
+  revealing a white ferret on a log. It plays, leaps into the water, and emerges, as the camera zooms in
+  for a close-up. Water splashes berry bushes nearby, while moss, snow, and leaves blanket the ground.
+  Birch trees and a light blue sky frame the scene, with ferns in the foreground. Side lighting casts dynamic
+  shadows and warm highlights. Medium composition, front view, low angle, with depth of field.
+  """
+  output = pipeline(
+      prompt=prompt,
+      num_frames=97,
+      guidance_scale=6.0,
+  ).frames[0]
+  export_to_video(output, "output.mp4", fps=24)
+  ```
+  </details>
+## SkyReelsV2DiffusionForcingPipeline
+[[autodoc]] SkyReelsV2DiffusionForcingPipeline
+  - all
+  - __call__
+## SkyReelsV2DiffusionForcingImageToVideoPipeline
+[[autodoc]] SkyReelsV2DiffusionForcingImageToVideoPipeline
+  - all
+  - __call__
+## SkyReelsV2DiffusionForcingVideoToVideoPipeline
+[[autodoc]] SkyReelsV2DiffusionForcingVideoToVideoPipeline
+  - all
+  - __call__
+## SkyReelsV2Pipeline
+[[autodoc]] SkyReelsV2Pipeline
+  - all
+  - __call__
+## SkyReelsV2ImageToVideoPipeline
+[[autodoc]] SkyReelsV2ImageToVideoPipeline
+  - all
+  - __call__
+## SkyReelsV2PipelineOutput
+[[autodoc]] pipelines.skyreels_v2.pipeline_output.SkyReelsV2PipelineOutput
\ No newline at end of file
--- a/scripts/convert_skyreelsv2_to_diffusers.py
+++ b/scripts/convert_skyreelsv2_to_diffusers.py
--- a/src/diffusers/__init__.py
+++ b/src/diffusers/__init__.py
@@ -219,6 +219,7 @@ else:
            "SD3ControlNetModel",
            "SD3MultiControlNetModel",
            "SD3Transformer2DModel",
+            "SkyReelsV2Transformer3DModel",
            "SparseControlNetModel",
            "StableAudioDiTModel",
            "StableCascadeUNet",
@@ -488,6 +489,11 @@ else:
            "SemanticStableDiffusionPipeline",
            "ShapEImg2ImgPipeline",
            "ShapEPipeline",
+            "SkyReelsV2DiffusionForcingImageToVideoPipeline",
+            "SkyReelsV2DiffusionForcingPipeline",
+            "SkyReelsV2DiffusionForcingVideoToVideoPipeline",
+            "SkyReelsV2ImageToVideoPipeline",
+            "SkyReelsV2Pipeline",
            "StableAudioPipeline",
            "StableAudioProjectionModel",
            "StableCascadeCombinedPipeline",
@@ -865,6 +871,7 @@ if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT:
            SD3ControlNetModel,
            SD3MultiControlNetModel,
            SD3Transformer2DModel,
+            SkyReelsV2Transformer3DModel,
            SparseControlNetModel,
            StableAudioDiTModel,
            T2IAdapter,
@@ -1109,6 +1116,11 @@ if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT:
            SemanticStableDiffusionPipeline,
            ShapEImg2ImgPipeline,
            ShapEPipeline,
+            SkyReelsV2DiffusionForcingImageToVideoPipeline,
+            SkyReelsV2DiffusionForcingPipeline,
+            SkyReelsV2DiffusionForcingVideoToVideoPipeline,
+            SkyReelsV2ImageToVideoPipeline,
+            SkyReelsV2Pipeline,
            StableAudioPipeline,
            StableAudioProjectionModel,
            StableCascadeCombinedPipeline,

--- a/src/diffusers/loaders/__init__.py
+++ b/src/diffusers/loaders/__init__.py
@@ -78,6 +78,7 @@ if is_torch_available():
            "Lumina2LoraLoaderMixin",
            "WanLoraLoaderMixin",
            "HiDreamImageLoraLoaderMixin",
+            "SkyReelsV2LoraLoaderMixin",
        ]
        _import_structure["textual_inversion"] = ["TextualInversionLoaderMixin"]
        _import_structure["ip_adapter"] = [
@@ -119,6 +120,7 @@ if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT:
                Mochi1LoraLoaderMixin,
                SanaLoraLoaderMixin,
                SD3LoraLoaderMixin,
+                SkyReelsV2LoraLoaderMixin,
                StableDiffusionLoraLoaderMixin,
                StableDiffusionXLLoraLoaderMixin,
                WanLoraLoaderMixin,

--- a/src/diffusers/loaders/lora_pipeline.py
+++ b/src/diffusers/loaders/lora_pipeline.py
@@ -5454,6 +5454,404 @@ class WanLoraLoaderMixin(LoraBaseMixin):
        super().unfuse_lora(components=components, **kwargs)
+class SkyReelsV2LoraLoaderMixin(LoraBaseMixin):
+    r"""
+    Load LoRA layers into [`SkyReelsV2Transformer3DModel`].
+    """
+    _lora_loadable_modules = ["transformer"]
+    transformer_name = TRANSFORMER_NAME
+    @classmethod
+    @validate_hf_hub_args
+    # Copied from diffusers.loaders.lora_pipeline.WanLoraLoaderMixin.lora_state_dict
+    def lora_state_dict(
+        cls,
+        pretrained_model_name_or_path_or_dict: Union[str, Dict[str, torch.Tensor]],
+        **kwargs,
+    ):
+        r"""
+        Return state dict for lora weights and the network alphas.
+        <Tip warning={true}>
+        We support loading A1111 formatted LoRA checkpoints in a limited capacity.
+        This function is experimental and might change in the future.
+        </Tip>
+        Parameters:
+            pretrained_model_name_or_path_or_dict (`str` or `os.PathLike` or `dict`):
+                Can be either:
+                    - A string, the *model id* (for example `google/ddpm-celebahq-256`) of a pretrained model hosted on
+                      the Hub.
+                    - A path to a *directory* (for example `./my_model_directory`) containing the model weights saved
+                      with [`ModelMixin.save_pretrained`].
+                    - A [torch state
+                      dict](https://pytorch.org/tutorials/beginner/saving_loading_models.html#what-is-a-state-dict).
+            cache_dir (`Union[str, os.PathLike]`, *optional*):
+                Path to a directory where a downloaded pretrained model configuration is cached if the standard cache
+                is not used.
+            force_download (`bool`, *optional*, defaults to `False`):
+                Whether or not to force the (re-)download of the model weights and configuration files, overriding the
+                cached versions if they exist.
+            proxies (`Dict[str, str]`, *optional*):
+                A dictionary of proxy servers to use by protocol or endpoint, for example, `{'http': 'foo.bar:3128',
+                'http://hostname': 'foo.bar:4012'}`. The proxies are used on each request.
+            local_files_only (`bool`, *optional*, defaults to `False`):
+                Whether to only load local model weights and configuration files or not. If set to `True`, the model
+                won't be downloaded from the Hub.
+            token (`str` or *bool*, *optional*):
+                The token to use as HTTP bearer authorization for remote files. If `True`, the token generated from
+                `diffusers-cli login` (stored in `~/.huggingface`) is used.
+            revision (`str`, *optional*, defaults to `"main"`):
+                The specific model version to use. It can be a branch name, a tag name, a commit id, or any identifier
+                allowed by Git.
+            subfolder (`str`, *optional*, defaults to `""`):
+                The subfolder location of a model file within a larger model repository on the Hub or locally.
+            return_lora_metadata (`bool`, *optional*, defaults to False):
+                When enabled, additionally return the LoRA adapter metadata, typically found in the state dict.
+        """
+        # Load the main state dict first which has the LoRA layers for either of
+        # transformer and text encoder or both.
+        cache_dir = kwargs.pop("cache_dir", None)
+        force_download = kwargs.pop("force_download", False)
+        proxies = kwargs.pop("proxies", None)
+        local_files_only = kwargs.pop("local_files_only", None)
+        token = kwargs.pop("token", None)
+        revision = kwargs.pop("revision", None)
+        subfolder = kwargs.pop("subfolder", None)
+        weight_name = kwargs.pop("weight_name", None)
+        use_safetensors = kwargs.pop("use_safetensors", None)
+        return_lora_metadata = kwargs.pop("return_lora_metadata", False)
+        allow_pickle = False
+        if use_safetensors is None:
+            use_safetensors = True
+            allow_pickle = True
+        user_agent = {"file_type": "attn_procs_weights", "framework": "pytorch"}
+        state_dict, metadata = _fetch_state_dict(
+            pretrained_model_name_or_path_or_dict=pretrained_model_name_or_path_or_dict,
+            weight_name=weight_name,
+            use_safetensors=use_safetensors,
+            local_files_only=local_files_only,
+            cache_dir=cache_dir,
+            force_download=force_download,
+            proxies=proxies,
+            token=token,
+            revision=revision,
+            subfolder=subfolder,
+            user_agent=user_agent,
+            allow_pickle=allow_pickle,
+        )
+        if any(k.startswith("diffusion_model.") for k in state_dict):
+            state_dict = _convert_non_diffusers_wan_lora_to_diffusers(state_dict)
+        elif any(k.startswith("lora_unet_") for k in state_dict):
+            state_dict = _convert_musubi_wan_lora_to_diffusers(state_dict)
+        is_dora_scale_present = any("dora_scale" in k for k in state_dict)
+        if is_dora_scale_present:
+            warn_msg = "It seems like you are using a DoRA checkpoint that is not compatible in Diffusers at the moment. So, we are going to filter out the keys associated to 'dora_scale` from the state dict. If you think this is a mistake please open an issue https://github.com/huggingface/diffusers/issues/new."
+            logger.warning(warn_msg)
+            state_dict = {k: v for k, v in state_dict.items() if "dora_scale" not in k}
+        out = (state_dict, metadata) if return_lora_metadata else state_dict
+        return out
+    @classmethod
+    # Copied from diffusers.loaders.lora_pipeline.WanLoraLoaderMixin._maybe_expand_t2v_lora_for_i2v
+    def _maybe_expand_t2v_lora_for_i2v(
+        cls,
+        transformer: torch.nn.Module,
+        state_dict,
+    ):
+        if transformer.config.image_dim is None:
+            return state_dict
+        target_device = transformer.device
+        if any(k.startswith("transformer.blocks.") for k in state_dict):
+            num_blocks = len({k.split("blocks.")[1].split(".")[0] for k in state_dict if "blocks." in k})
+            is_i2v_lora = any("add_k_proj" in k for k in state_dict) and any("add_v_proj" in k for k in state_dict)
+            has_bias = any(".lora_B.bias" in k for k in state_dict)
+            if is_i2v_lora:
+                return state_dict
+            for i in range(num_blocks):
+                for o, c in zip(["k_img", "v_img"], ["add_k_proj", "add_v_proj"]):
+                    # These keys should exist if the block `i` was part of the T2V LoRA.
+                    ref_key_lora_A = f"transformer.blocks.{i}.attn2.to_k.lora_A.weight"
+                    ref_key_lora_B = f"transformer.blocks.{i}.attn2.to_k.lora_B.weight"
+                    if ref_key_lora_A not in state_dict or ref_key_lora_B not in state_dict:
+                        continue
+                    state_dict[f"transformer.blocks.{i}.attn2.{c}.lora_A.weight"] = torch.zeros_like(
+                        state_dict[f"transformer.blocks.{i}.attn2.to_k.lora_A.weight"], device=target_device
+                    )
+                    state_dict[f"transformer.blocks.{i}.attn2.{c}.lora_B.weight"] = torch.zeros_like(
+                        state_dict[f"transformer.blocks.{i}.attn2.to_k.lora_B.weight"], device=target_device
+                    )
+                    # If the original LoRA had biases (indicated by has_bias)
+                    # AND the specific reference bias key exists for this block.
+                    ref_key_lora_B_bias = f"transformer.blocks.{i}.attn2.to_k.lora_B.bias"
+                    if has_bias and ref_key_lora_B_bias in state_dict:
+                        ref_lora_B_bias_tensor = state_dict[ref_key_lora_B_bias]
+                        state_dict[f"transformer.blocks.{i}.attn2.{c}.lora_B.bias"] = torch.zeros_like(
+                            ref_lora_B_bias_tensor,
+                            device=target_device,
+                        )
+        return state_dict
+    # Copied from diffusers.loaders.lora_pipeline.WanLoraLoaderMixin.load_lora_weights
+    def load_lora_weights(
+        self,
+        pretrained_model_name_or_path_or_dict: Union[str, Dict[str, torch.Tensor]],
+        adapter_name: Optional[str] = None,
+        hotswap: bool = False,
+        **kwargs,
+    ):
+        """
+        Load LoRA weights specified in `pretrained_model_name_or_path_or_dict` into `self.transformer` and
+        `self.text_encoder`. All kwargs are forwarded to `self.lora_state_dict`. See
+        [`~loaders.StableDiffusionLoraLoaderMixin.lora_state_dict`] for more details on how the state dict is loaded.
+        See [`~loaders.StableDiffusionLoraLoaderMixin.load_lora_into_transformer`] for more details on how the state
+        dict is loaded into `self.transformer`.
+        Parameters:
+            pretrained_model_name_or_path_or_dict (`str` or `os.PathLike` or `dict`):
+                See [`~loaders.StableDiffusionLoraLoaderMixin.lora_state_dict`].
+            adapter_name (`str`, *optional*):
+                Adapter name to be used for referencing the loaded adapter model. If not specified, it will use
+                `default_{i}` where i is the total number of adapters being loaded.
+            low_cpu_mem_usage (`bool`, *optional*):
+                Speed up model loading by only loading the pretrained LoRA weights and not initializing the random
+                weights.
+            hotswap (`bool`, *optional*):
+                See [`~loaders.StableDiffusionLoraLoaderMixin.load_lora_weights`].
+            kwargs (`dict`, *optional*):
+                See [`~loaders.StableDiffusionLoraLoaderMixin.lora_state_dict`].
+        """
+        if not USE_PEFT_BACKEND:
+            raise ValueError("PEFT backend is required for this method.")
+        low_cpu_mem_usage = kwargs.pop("low_cpu_mem_usage", _LOW_CPU_MEM_USAGE_DEFAULT_LORA)
+        if low_cpu_mem_usage and is_peft_version("<", "0.13.0"):
+            raise ValueError(
+                "`low_cpu_mem_usage=True` is not compatible with this `peft` version. Please update it with `pip install -U peft`."
+            )
+        # if a dict is passed, copy it instead of modifying it inplace
+        if isinstance(pretrained_model_name_or_path_or_dict, dict):
+            pretrained_model_name_or_path_or_dict = pretrained_model_name_or_path_or_dict.copy()
+        # First, ensure that the checkpoint is a compatible one and can be successfully loaded.
+        kwargs["return_lora_metadata"] = True
+        state_dict, metadata = self.lora_state_dict(pretrained_model_name_or_path_or_dict, **kwargs)
+        # convert T2V LoRA to I2V LoRA (when loaded to Wan I2V) by adding zeros for the additional (missing) _img layers
+        state_dict = self._maybe_expand_t2v_lora_for_i2v(
+            transformer=getattr(self, self.transformer_name) if not hasattr(self, "transformer") else self.transformer,
+            state_dict=state_dict,
+        )
+        is_correct_format = all("lora" in key for key in state_dict.keys())
+        if not is_correct_format:
+            raise ValueError("Invalid LoRA checkpoint.")
+        self.load_lora_into_transformer(
+            state_dict,
+            transformer=getattr(self, self.transformer_name) if not hasattr(self, "transformer") else self.transformer,
+            adapter_name=adapter_name,
+            metadata=metadata,
+            _pipeline=self,
+            low_cpu_mem_usage=low_cpu_mem_usage,
+            hotswap=hotswap,
+        )
+    @classmethod
+    # Copied from diffusers.loaders.lora_pipeline.SD3LoraLoaderMixin.load_lora_into_transformer with SD3Transformer2DModel->SkyReelsV2Transformer3DModel
+    def load_lora_into_transformer(
+        cls,
+        state_dict,
+        transformer,
+        adapter_name=None,
+        _pipeline=None,
+        low_cpu_mem_usage=False,
+        hotswap: bool = False,
+        metadata=None,
+    ):
+        """
+        This will load the LoRA layers specified in `state_dict` into `transformer`.
+        Parameters:
+            state_dict (`dict`):
+                A standard state dict containing the lora layer parameters. The keys can either be indexed directly
+                into the unet or prefixed with an additional `unet` which can be used to distinguish between text
+                encoder lora layers.
+            transformer (`SkyReelsV2Transformer3DModel`):
+                The Transformer model to load the LoRA layers into.
+            adapter_name (`str`, *optional*):
+                Adapter name to be used for referencing the loaded adapter model. If not specified, it will use
+                `default_{i}` where i is the total number of adapters being loaded.
+            low_cpu_mem_usage (`bool`, *optional*):
+                Speed up model loading by only loading the pretrained LoRA weights and not initializing the random
+                weights.
+            hotswap (`bool`, *optional*):
+                See [`~loaders.StableDiffusionLoraLoaderMixin.load_lora_weights`].
+            metadata (`dict`):
+                Optional LoRA adapter metadata. When supplied, the `LoraConfig` arguments of `peft` won't be derived
+                from the state dict.
+        """
+        if low_cpu_mem_usage and is_peft_version("<", "0.13.0"):
+            raise ValueError(
+                "`low_cpu_mem_usage=True` is not compatible with this `peft` version. Please update it with `pip install -U peft`."
+            )
+        # Load the layers corresponding to transformer.
+        logger.info(f"Loading {cls.transformer_name}.")
+        transformer.load_lora_adapter(
+            state_dict,
+            network_alphas=None,
+            adapter_name=adapter_name,
+            metadata=metadata,
+            _pipeline=_pipeline,
+            low_cpu_mem_usage=low_cpu_mem_usage,
+            hotswap=hotswap,
+        )
+    @classmethod
+    # Copied from diffusers.loaders.lora_pipeline.CogVideoXLoraLoaderMixin.save_lora_weights
+    def save_lora_weights(
+        cls,
+        save_directory: Union[str, os.PathLike],
+        transformer_lora_layers: Dict[str, Union[torch.nn.Module, torch.Tensor]] = None,
+        is_main_process: bool = True,
+        weight_name: str = None,
+        save_function: Callable = None,
+        safe_serialization: bool = True,
+        transformer_lora_adapter_metadata: Optional[dict] = None,
+    ):
+        r"""
+        Save the LoRA parameters corresponding to the transformer.
+        Arguments:
+            save_directory (`str` or `os.PathLike`):
+                Directory to save LoRA parameters to. Will be created if it doesn't exist.
+            transformer_lora_layers (`Dict[str, torch.nn.Module]` or `Dict[str, torch.Tensor]`):
+                State dict of the LoRA layers corresponding to the `transformer`.
+            is_main_process (`bool`, *optional*, defaults to `True`):
+                Whether the process calling this is the main process or not. Useful during distributed training and you
+                need to call this function on all processes. In this case, set `is_main_process=True` only on the main
+                process to avoid race conditions.
+            save_function (`Callable`):
+                The function to use to save the state dictionary. Useful during distributed training when you need to
+                replace `torch.save` with another method. Can be configured with the environment variable
+                `DIFFUSERS_SAVE_MODE`.
+            safe_serialization (`bool`, *optional*, defaults to `True`):
+                Whether to save the model using `safetensors` or the traditional PyTorch way with `pickle`.
+            transformer_lora_adapter_metadata:
+                LoRA adapter metadata associated with the transformer to be serialized with the state dict.
+        """
+        state_dict = {}
+        lora_adapter_metadata = {}
+        if not transformer_lora_layers:
+            raise ValueError("You must pass `transformer_lora_layers`.")
+        state_dict.update(cls.pack_weights(transformer_lora_layers, cls.transformer_name))
+        if transformer_lora_adapter_metadata is not None:
+            lora_adapter_metadata.update(
+                _pack_dict_with_prefix(transformer_lora_adapter_metadata, cls.transformer_name)
+            )
+        # Save the model
+        cls.write_lora_layers(
+            state_dict=state_dict,
+            save_directory=save_directory,
+            is_main_process=is_main_process,
+            weight_name=weight_name,
+            save_function=save_function,
+            safe_serialization=safe_serialization,
+            lora_adapter_metadata=lora_adapter_metadata,
+        )
+    # Copied from diffusers.loaders.lora_pipeline.CogVideoXLoraLoaderMixin.fuse_lora
+    def fuse_lora(
+        self,
+        components: List[str] = ["transformer"],
+        lora_scale: float = 1.0,
+        safe_fusing: bool = False,
+        adapter_names: Optional[List[str]] = None,
+        **kwargs,
+    ):
+        r"""
+        Fuses the LoRA parameters into the original parameters of the corresponding blocks.
+        <Tip warning={true}>
+        This is an experimental API.
+        </Tip>
+        Args:
+            components: (`List[str]`): List of LoRA-injectable components to fuse the LoRAs into.
+            lora_scale (`float`, defaults to 1.0):
+                Controls how much to influence the outputs with the LoRA parameters.
+            safe_fusing (`bool`, defaults to `False`):
+                Whether to check fused weights for NaN values before fusing and if values are NaN not fusing them.
+            adapter_names (`List[str]`, *optional*):
+                Adapter names to be used for fusing. If nothing is passed, all active adapters will be fused.
+        Example:
+        ```py
+        from diffusers import DiffusionPipeline
+        import torch
+        pipeline = DiffusionPipeline.from_pretrained(
+            "stabilityai/stable-diffusion-xl-base-1.0", torch_dtype=torch.float16
+        ).to("cuda")
+        pipeline.load_lora_weights("nerijs/pixel-art-xl", weight_name="pixel-art-xl.safetensors", adapter_name="pixel")
+        pipeline.fuse_lora(lora_scale=0.7)
+        ```
+        """
+        super().fuse_lora(
+            components=components,
+            lora_scale=lora_scale,
+            safe_fusing=safe_fusing,
+            adapter_names=adapter_names,
+            **kwargs,
+        )
+    # Copied from diffusers.loaders.lora_pipeline.CogVideoXLoraLoaderMixin.unfuse_lora
+    def unfuse_lora(self, components: List[str] = ["transformer"], **kwargs):
+        r"""
+        Reverses the effect of
+        [`pipe.fuse_lora()`](https://huggingface.co/docs/diffusers/main/en/api/loaders#diffusers.loaders.LoraBaseMixin.fuse_lora).
+        <Tip warning={true}>
+        This is an experimental API.
+        </Tip>
+        Args:
+            components (`List[str]`): List of LoRA-injectable components to unfuse LoRA from.
+            unfuse_transformer (`bool`, defaults to `True`): Whether to unfuse the UNet LoRA parameters.
+        """
+        super().unfuse_lora(components=components, **kwargs)
 class CogView4LoraLoaderMixin(LoraBaseMixin):
    r"""
    Load LoRA layers into [`WanTransformer3DModel`]. Specific to [`CogView4Pipeline`].

--- a/src/diffusers/models/__init__.py
+++ b/src/diffusers/models/__init__.py
@@ -88,6 +88,7 @@ if is_torch_available():
    _import_structure["transformers.transformer_mochi"] = ["MochiTransformer3DModel"]
    _import_structure["transformers.transformer_omnigen"] = ["OmniGenTransformer2DModel"]
    _import_structure["transformers.transformer_sd3"] = ["SD3Transformer2DModel"]
+    _import_structure["transformers.transformer_skyreels_v2"] = ["SkyReelsV2Transformer3DModel"]
    _import_structure["transformers.transformer_temporal"] = ["TransformerTemporalModel"]
    _import_structure["transformers.transformer_wan"] = ["WanTransformer3DModel"]
    _import_structure["transformers.transformer_wan_vace"] = ["WanVACETransformer3DModel"]
@@ -176,6 +177,7 @@ if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT:
            PriorTransformer,
            SanaTransformer2DModel,
            SD3Transformer2DModel,
+            SkyReelsV2Transformer3DModel,
            StableAudioDiTModel,
            T5FilmDecoder,
            Transformer2DModel,

--- a/src/diffusers/models/embeddings.py
+++ b/src/diffusers/models/embeddings.py
@@ -319,7 +319,7 @@ def get_2d_sincos_pos_embed_from_grid(embed_dim, grid, output_type="np"):
    return emb
-def get_1d_sincos_pos_embed_from_grid(embed_dim, pos, output_type="np"):
+def get_1d_sincos_pos_embed_from_grid(embed_dim, pos, output_type="np", flip_sin_to_cos=False):
    """
    This function generates 1D positional embeddings from a grid.
@@ -352,6 +352,11 @@ def get_1d_sincos_pos_embed_from_grid(embed_dim, pos, output_type="np"):
    emb_cos = torch.cos(out)  # (M, D/2)
    emb = torch.concat([emb_sin, emb_cos], dim=1)  # (M, D)
+    # flip sine and cosine embeddings
+    if flip_sin_to_cos:
+        emb = torch.cat([emb[:, embed_dim // 2 :], emb[:, : embed_dim // 2]], dim=1)
    return emb

--- a/src/diffusers/models/modeling_utils.py
+++ b/src/diffusers/models/modeling_utils.py
@@ -172,7 +172,11 @@ def get_parameter_dtype(parameter: torch.nn.Module) -> torch.dtype:
    for name, param in parameter.named_parameters():
        last_dtype = param.dtype
-        if parameter._keep_in_fp32_modules and any(m in name for m in parameter._keep_in_fp32_modules):
+        if (
+            hasattr(parameter, "_keep_in_fp32_modules")
+            and parameter._keep_in_fp32_modules
+            and any(m in name for m in parameter._keep_in_fp32_modules)
+        ):
            continue
        if param.is_floating_point():

--- a/src/diffusers/models/transformers/__init__.py
+++ b/src/diffusers/models/transformers/__init__.py
@@ -31,6 +31,7 @@ if is_torch_available():
    from .transformer_mochi import MochiTransformer3DModel
    from .transformer_omnigen import OmniGenTransformer2DModel
    from .transformer_sd3 import SD3Transformer2DModel
+    from .transformer_skyreels_v2 import SkyReelsV2Transformer3DModel
    from .transformer_temporal import TransformerTemporalModel
    from .transformer_wan import WanTransformer3DModel
    from .transformer_wan_vace import WanVACETransformer3DModel
--- a/src/diffusers/models/transformers/transformer_skyreels_v2.py
+++ b/src/diffusers/models/transformers/transformer_skyreels_v2.py
--- a/src/diffusers/pipelines/__init__.py
+++ b/src/diffusers/pipelines/__init__.py
@@ -380,6 +380,13 @@ else:
        "WuerstchenPriorPipeline",
    ]
    _import_structure["wan"] = ["WanPipeline", "WanImageToVideoPipeline", "WanVideoToVideoPipeline", "WanVACEPipeline"]
+    _import_structure["skyreels_v2"] = [
+        "SkyReelsV2DiffusionForcingPipeline",
+        "SkyReelsV2DiffusionForcingImageToVideoPipeline",
+        "SkyReelsV2DiffusionForcingVideoToVideoPipeline",
+        "SkyReelsV2ImageToVideoPipeline",
+        "SkyReelsV2Pipeline",
+    ]
 try:
    if not is_onnx_available():
        raise OptionalDependencyNotAvailable()
@@ -851,6 +858,14 @@ if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT:
                SpectrogramDiffusionPipeline,
            )
+        from .skyreels_v2 import (
+            SkyReelsV2DiffusionForcingImageToVideoPipeline,
+            SkyReelsV2DiffusionForcingPipeline,
+            SkyReelsV2DiffusionForcingVideoToVideoPipeline,
+            SkyReelsV2ImageToVideoPipeline,
+            SkyReelsV2Pipeline,
+        )
 else:
    import sys

--- a/src/diffusers/pipelines/skyreels_v2/__init__.py
+++ b/src/diffusers/pipelines/skyreels_v2/__init__.py
+from typing import TYPE_CHECKING
+from ...utils import (
+    DIFFUSERS_SLOW_IMPORT,
+    OptionalDependencyNotAvailable,
+    _LazyModule,
+    get_objects_from_module,
+    is_torch_available,
+    is_transformers_available,
+)
+_dummy_objects = {}
+_import_structure = {}
+try:
+    if not (is_transformers_available() and is_torch_available()):
+        raise OptionalDependencyNotAvailable()
+except OptionalDependencyNotAvailable:
+    from ...utils import dummy_torch_and_transformers_objects  # noqa F403
+    _dummy_objects.update(get_objects_from_module(dummy_torch_and_transformers_objects))
+else:
+    _import_structure["pipeline_skyreels_v2"] = ["SkyReelsV2Pipeline"]
+    _import_structure["pipeline_skyreels_v2_diffusion_forcing"] = ["SkyReelsV2DiffusionForcingPipeline"]
+    _import_structure["pipeline_skyreels_v2_diffusion_forcing_i2v"] = [
+        "SkyReelsV2DiffusionForcingImageToVideoPipeline"
+    ]
+    _import_structure["pipeline_skyreels_v2_diffusion_forcing_v2v"] = [
+        "SkyReelsV2DiffusionForcingVideoToVideoPipeline"
+    ]
+    _import_structure["pipeline_skyreels_v2_i2v"] = ["SkyReelsV2ImageToVideoPipeline"]
+if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT:
+    try:
+        if not (is_transformers_available() and is_torch_available()):
+            raise OptionalDependencyNotAvailable()
+    except OptionalDependencyNotAvailable:
+        from ...utils.dummy_torch_and_transformers_objects import *
+    else:
+        from .pipeline_skyreels_v2 import SkyReelsV2Pipeline
+        from .pipeline_skyreels_v2_diffusion_forcing import SkyReelsV2DiffusionForcingPipeline
+        from .pipeline_skyreels_v2_diffusion_forcing_i2v import SkyReelsV2DiffusionForcingImageToVideoPipeline
+        from .pipeline_skyreels_v2_diffusion_forcing_v2v import SkyReelsV2DiffusionForcingVideoToVideoPipeline
+        from .pipeline_skyreels_v2_i2v import SkyReelsV2ImageToVideoPipeline
+else:
+    import sys
+    sys.modules[__name__] = _LazyModule(
+        __name__,
+        globals()["__file__"],
+        _import_structure,
+        module_spec=__spec__,
+    )
+    for name, value in _dummy_objects.items():
+        setattr(sys.modules[__name__], name, value)
--- a/src/diffusers/pipelines/skyreels_v2/pipeline_output.py
+++ b/src/diffusers/pipelines/skyreels_v2/pipeline_output.py
+from dataclasses import dataclass
+import torch
+from diffusers.utils import BaseOutput
+@dataclass
+class SkyReelsV2PipelineOutput(BaseOutput):
+    r"""
+    Output class for SkyReelsV2 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)`.
+    """
+    frames: torch.Tensor
--- a/src/diffusers/pipelines/skyreels_v2/pipeline_skyreels_v2.py
+++ b/src/diffusers/pipelines/skyreels_v2/pipeline_skyreels_v2.py
--- a/src/diffusers/pipelines/skyreels_v2/pipeline_skyreels_v2_diffusion_forcing.py
+++ b/src/diffusers/pipelines/skyreels_v2/pipeline_skyreels_v2_diffusion_forcing.py
--- a/src/diffusers/pipelines/skyreels_v2/pipeline_skyreels_v2_diffusion_forcing_i2v.py
+++ b/src/diffusers/pipelines/skyreels_v2/pipeline_skyreels_v2_diffusion_forcing_i2v.py
--- a/src/diffusers/pipelines/skyreels_v2/pipeline_skyreels_v2_diffusion_forcing_v2v.py
+++ b/src/diffusers/pipelines/skyreels_v2/pipeline_skyreels_v2_diffusion_forcing_v2v.py