JD's JoyAI-Video-Edit Does Streaming 720p Video Editing at ~30 FPS on One B200 With No Access to Future Frames

arXiv / HuggingFace Daily Papers·medium signalXBlueskyLinkedInCopy link

arXiv 2608.03974, trending on Hugging Face Daily Papers August 5 with 70 upvotes, tackles causal video editing: bounded compute, low latency, no lookahead. The 16B model combines chunk-wise autoregressive adaptation, Source-Anchored Distribution Matching Distillation and Long-Horizon Autoregressive Distillation to hold source fidelity and temporal consistency across long clips. Reported numbers: ~30 FPS end-to-end 720p on a single Nvidia B200, 2.68s full-pipeline latency for 81 frames, 200 FPS VAE, 3.60 overall on OpenVE-Bench short video and 3.30 on LongV2VBench for one-minute clips, with 87–90% human preference over competing streaming editors. From Joy Future Academy and JD.