StreamGVE: Training-Free Video Editing via Few-Step Streaming Video Generation

Guanlong Jiao^1,4, Chenyangguang Zhang², Jia Jun Cheng Xian^1,4, Zewei Zhang^1,3, Renjie Liao^1,4,5

¹The University of British Columbia, ²ETH Zürich, ³McMaster University, ⁴Vector Institute, ⁵Canada CIFAR AI Chair

Video Results are all shown in our Project Page.

---

✨ Highlights

StreamGVE is a training-free video editing framework built on few-step streaming video generation models. Instead of treating editing as data-to-data transformation, StreamGVE formulates editing as source-conditioned noise-to-target generation, thus makeing it possible for few-step fast controllable video editing. StreamGVE supports few-step text-driven video editing and optional first-frame visual prompting for videos of any length. It is implemented on streaming video generation models:

• Self-Forcing_StreamGVE: StreamGVE built on Self Forcing.
• LongLive_StreamGVE: StreamGVE built on LongLive v1.0.

🏡 Environment

conda create -n streamgve python=3.12 -y
conda activate streamgve

# Choose the PyTorch command appropriate for your CUDA version, example for CUDA 12.8:
pip install torch==2.8.0 torchvision==0.23.0 --index-url https://download.pytorch.org/whl/cu128

pip install -r requirements.txt
pip install flash-attn --no-build-isolation

For the Self Forcing-based implementation, install the local package in editable mode:

cd Self-Forcing_StreamGVE
python setup.py develop
cd ..

🎯 Checkpoints

Self Forcing

cd Self-Forcing_StreamGVE
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 .
cd ..

LongLive

cd LongLive_StreamGVE
huggingface-cli download Wan-AI/Wan2.1-T2V-1.3B --local-dir wan_models/Wan2.1-T2V-1.3B
huggingface-cli download Efficient-Large-Model/LongLive --local-dir longlive_models
cd ..

You can share the checkpoints of Wan through soft links, as they are the same across projects.

🎨 Running Video Editing

Each implementation provides a ready-to-run example script.

Self Forcing-based editing

cd Self-Forcing_StreamGVE
bash inference_edit_streamgve.sh

LongLive-based editing

cd LongLive_StreamGVE
bash inference_edit_streamgve.sh

🎉 Acknowledgements

This repository builds on the excellent open-source work of Self Forcing, LongLive, and Wan2.1. We also thank UniEdit-Flow and FiVE-Bench for helpful open-soucre code and benchmarks.

🔮 Citation

If you find this work useful, please consider citing:

TODO