VINCIE-NExT: Unlocking Video Editing from Images via In-Context Modeling

Website | Paper | Code

VINCIE-NExT transfers editing capability from images to videos through in-context visual demonstrations. It decomposes video editing into a chain of sub-tasks (Video → Image → Image → Video). An image editing pair, generated by the model or supplied by the user, serves as an appearance blueprint for every output frame.

Files

File	Description
dit_480p.pth	The final model at 640×480 (reported in the paper).
dit_256p.pth	The model of the first training stage at 256×256 (used for ablations in the paper).

Both are 3B MM-DiT weights (fp32). The VAE and the text encoder (Qwen-14B) are shared with VINCIE-3B.

Usage

git clone https://github.com/LgQu/VINCIE-NExT
cd VINCIE-NExT
pip install -r requirements.txt
python download_ckpt.py  # DiT weights from this repo, VAE and text encoder from VINCIE-3B
bash scripts/infer_480p.sh

See the code repository for all inference settings.

Citation

@inproceedings{qu2026vincienext,
  title     = {VINCIE-NExT: Unlocking Video Editing from Images via In-Context Modeling},
  author    = {Qu, Leigang and Cheng, Feng and Yang, Ziyan and Yang, Bangbang and Huang, Zhaoyang and Chow, Wei and Li, Yicong and Wang, Wenjie and Chua, Tat-Seng and Zeng, Yan},
  booktitle = {Advances in Neural Information Processing Systems},
  year      = {2026}
}

License

Apache-2.0, subject to any intellectual property rights in the model owned by ByteDance. The text encoder is adapted from Qwen-14B, and your use of that model must comply with its license.