JD's 16B JoyAI-Video-Edit streams 720p edits at 30 FPS on B200

A team from JD's Joy Future Academy has put out a paper claiming something the offline video editing world has been reaching for, a diffusion-based editor that emits frames as the source video arrives instead of waiting for the whole clip. The paper, posted on Hugging Face, describes JoyAI-Video-Edit, a 16B-parameter autoregressive diffusion framework that does end-to-end 720p editing at approximately 30 FPS on a single Nvidia B200 GPU.

The engineering trick is not raw scale. The model runs bidirectional attention inside each temporal chunk and causal attention across chunks, with a sliding window that keeps only recent chunks and the first chunk (as a global sink) in memory. That bounds per-step compute regardless of how long the stream runs. On top of that, the team distills the denoising process into a two-step generator using something they call Source-Anchored Distribution Matching Distillation, which guides the teacher independently along text conditioning and source fidelity so the model stays anchored to incoming frames rather than drifting into its own hallucinated history. Add FP8 quantization, bounded KV caching and an optimized VAE pipeline and the deployment cycle lands at 266 ms per chunk (22 ms VAE encode, 185 ms DiT denoise, 19 ms VAE decode, plus 31 ms clean KV-cache construction and 9 ms pseudo encoding).

Why this matters if you are not building diffusion pipelines: high-quality instruction-guided video editing has effectively been an offline post-production activity, which ruled it out of live broadcast, video calls and interactive entertainment. A causal streaming editor at roughly broadcast frame rate changes that assumption. On the paper's own OpenVE-Bench, JoyAI-Video-Edit scores 3.60 overall, ahead of every streaming editor the authors tested and comparable to strong offline systems such as Kling-3.0 Omni and Bernini-R.

The honest caveat is that these numbers come from JD's own reporting and, in the case of the one-minute long-video test, a benchmark (LongV2VBench, 229 tasks across five categories) that the authors themselves constructed. The model still trails the best offline systems in individual categories such as background change, and the 30 FPS figure is tied to the B200 plus FP8 quantization, with no throughput reported on more common hardware and no discussion of content-safety or watermarking for real-time edits.

If the results hold up outside the authors' own evaluation, the interesting downstream is not another flagship benchmark line, it is that the code has been released on GitHub, giving other teams building live video products a concrete recipe for streaming diffusion editing instead of a demo reel.