README.md

metadata

license: mit
pipeline_tag: video-to-video
library_name: transformers

ReViSE: Towards Reason-Informed Video Editing in Unified Models with Self-Reflective Learning

This repository contains the official implementation of the paper ReViSE: Towards Reason-Informed Video Editing in Unified Models with Self-Reflective Learning.

ReViSE introduces the Reason-Informed Video Editing (RVE) task, which requires reasoning about physical plausibility and causal dynamics during editing. It proposes a Self-Reflective Reasoning (SRF) framework that unifies generation and evaluation within a single architecture, utilizing an internal VLM for intrinsic feedback. This model significantly enhances editing accuracy and visual fidelity in reason-informed video editing.

GitHub Repository: https://github.com/Liuxinyv/ReViSE

Demos

Reason-informed video editing

What if the the dog ran into the depth of a forest?
What if the girl’s fragrance gently attracted a delicate butterfly, fluttering toward her?
What if the scene transitioned from a magical night to a dawn, causing the northern lights to fade away?

Quick Start (Inference)

To get started with ReViSE inference, follow these steps:

1. Create conda environment

conda create -n revise python=3.10
conda activate revise

pip install -r pip_requirements.txt

2. Set up environment variables for CUDA

# For CUDA (adjust path as needed)
export CUDA_HOME="/usr/local/cuda"
export PATH="${CUDA_HOME}/bin:${PATH}"
export LD_LIBRARY_PATH="${CUDA_HOME}/lib64:${LD_LIBRARY_PATH}"

# Add to inference or training script
export PYTHONPATH="${PWD}:${PWD}/nets/third_party:${PYTHONPATH}"

3. Downloading checkpoints Download our pretrained model checkpoint here.

Inference

# Run inference with sample data
bash tools/inference/inference.sh

Acknowledgement

We would like to thank Omni-Video, VILA and Wan2.1 for their excellent work.

Citation

If you find this project useful, please consider citing:

@misc{liu2025revisereasoninformedvideoediting,
      title={ReViSE: Towards Reason-Informed Video Editing in Unified Models with Self-Reflective Learning}, 
      author={Xinyu Liu and Hangjie Yuan and Yujie Wei and Jiazheng Xing and Yujin Han and Jiahao Pan and Yanbiao Ma and Chi-Min Chan and Kang Zhao and Shiwei Zhang and Wenhan Luo and Yike Guo},
      year={2025},
      eprint={2512.09924},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2512.09924}, 
}