‹ back to 2026-08-05

JoyAI-Video-Edit: Real-Time Open-Ended Video Editing with Autoregressive Diffusion

arXiv cs.CV Multimodal & Generative Yicheng Xiao, Wenxun Dai, Xinran Qin, Lin Song, Maoquan Zhang, Hang Xu, Yukang Chen, Yitong Li, Guohui Zhang, Yuan Zhang, Xuying Zhang, Tommy Zhang, Jianlong Yuan, Peihao Li, Shuai Lu, Siming Fu, Chuyang Zhao, Xin Han, Jie Huang, Wenbo Li, Guoqing Ma, Wei Huang, Xiaojuan Qi, Haoyang Huang, Nan Duan 2026-08-04

TL;DR - JoyAI-Video-Edit is a 16B-parameter autoregressive diffusion framework for open-ended, causal video editing. It achieves approximately 30 FPS at 720p on a single Nvidia B200 while maintaining source fidelity and long-term temporal consistency.

Processes video in autoregressive chunks without future frames or a predefined duration.
Uses Source-Anchored Distribution Matching Distillation to preserve source fidelity during two-step generation.
Applies Long-Horizon Autoregressive Distillation to reduce temporal drift and train–inference mismatch.
Outperforms existing streaming editors in reported evaluations and remains competitive with strong offline systems.

view merged work →