Self Forcing & StreamEdit

This folder contains the StreamEdit implementation built on Self Forcing codebase.

Most files are inherited from the upstream Self Forcing repository. The implementation of StreamEdit is mainly located in:

• inference_edit_streamedit.py: command-line entry point for StreamEdit video editing.
• inference_edit_streamedit.sh: example editing commands.
• pipeline/edit_causal_inference.py: dual-branch editing pipeline, attention bridge, grounding/boosting, source-oriented guidance, and visual prompting logic.
• pipeline/__init__.py: exposes EditCausalInferencePipeline.
• wan/modules/model.py and wan/modules/causal_model.py: attention-level support for grounding, boosting, query/key blending, source KV injection, and mask-aware cache usage.

The original Self Forcing training and generation files are kept for code compatibility and attribution. For full upstream documentation, please refer to the original Self Forcing repository.

Quick Run

Install dependencies from the top-level requirements.txt, then prepare checkpoints:

huggingface-cli download Wan-AI/Wan2.1-T2V-1.3B --local-dir-use-symlinks False --local-dir wan_models/Wan2.1-T2V-1.3B
huggingface-cli download gdhe17/Self-Forcing checkpoints/self_forcing_dmd.pt --local-dir .
python setup.py develop

Run the provided editing examples:

bash inference_edit_streamedit.sh

Refer to Self Forcing Issue 2, we implemented rollout-based long-video editing for Self Forcing, which is different from LongLive's natural adaptation to any length.

Acknowledgements

This folder is based on Self Forcing, which is licensed under Apache-2.0. Please keep the upstream license and cite Self Forcing when using this implementation.