JD.com open-sources real-time video editing model JoyAI-Video-Edit

JD.com announced on August 5 that it has open-sourced its self-developed real-time streaming video editing model, JoyAI-Video-Edit. This model lets users modify characters and scenes while watching a video, shifting video creation from 'editing after recording' to real-time interactive editing. Built on JD's own JoyAI-Video model, it achieves 30 frames per second inference at 720P resolution, keeping up with typical video playback, and supports stable streaming editing of any length, so users can make changes without waiting for the whole video to render.

In benchmark tests against international models like SANA Streaming and LiveEdit, JoyAI-Video-Edit outperformed all streaming editing methods on the OpenVE-Bench, especially in tasks like global style changes and local replacements. The model also offers a new way to synthesize data for embodied intelligence, turning videos of human hands into robotic arm operations, which helps reduce data collection costs for robot training. The open-source code is available on Hugging Face and GitHub.