Kiwi-Edit is a versatile video editing framework built on an MLLM encoder and a video Diffusion Transformer (DiT). It supports both instruction-based video editing and reference-guided editing (using a reference image and instruction).
Kiwi-Edit introduces a unified editing architecture that synergizes learnable queries and latent visual features for reference semantic guidance. It addresses the challenge of precise visual control in instruction-based editing by allowing users to provide a reference image to guide the transformation. The framework achieves significant performance improvements in instruction following and reference fidelity through a scalable data generation pipeline and a multi-stage training curriculum.
This model is compatible with the
diffusers library. To run inference, follow the installation instructions in the
official repository.
You can run a quick test on a demo video using the following command provided in the repository:
1python diffusers_demo.py \
2 --video_path ./demo_data/video/source/0005e4ad9f49814db1d3f2296b911abf.mp4 \
3 --prompt "Remove the monkey." \
4 --save_path output.mp4 \
5 --model_path linyq/kiwi-edit-5b-instruct-only-diffusers
1@misc{kiwiedit,
2 title={Kiwi-Edit: Versatile Video Editing via Instruction and Reference Guidance},
3 author={Yiqi Lin and Guoqiang Liang and Ziyun Zeng and Zechen Bai and Yanzhe Chen and Mike Zheng Shou},
4 year={2026},
5 eprint={2603.02175},
6 archivePrefix={arXiv},
7 primaryClass={cs.CV},
8 url={https://arxiv.org/abs/2603.02175},
9}
