JD open-sources 16B real-time video editing model
JD.com released Apache-2.0 weights for JoyAI-Video-Edit, which edits 720p video at about 30 frames per second on a single NVIDIA B200.
JD.com's open-source arm published weights, deployment code and a technical report on Wednesday for JoyAI-Video-Edit, a 16-billion-parameter model that performs instruction-guided video editing in real time rather than as an offline render job.
The claim
The system is built on a multimodal diffusion transformer converted to chunk-wise autoregressive generation, so it consumes and emits video in causal segments and never needs to see future frames. That removes the fixed-duration ceiling common to diffusion video editors: the model accepts an unbounded stream. JD reports end-to-end throughput of 30.19 frames per second at 720×1280 on a single NVIDIA B200 accelerator, which is roughly the threshold at which editing can be applied to a live feed instead of a file.
Getting there required two distillation stages described in the paper. Source-Anchored Distribution Matching Distillation compresses the sampling schedule while keeping outputs tied to the original footage, addressing the tendency of few-step distilled editors to drift away from the source. Long-Horizon Autoregressive Distillation targets error accumulation, the failure mode in which causal generation degrades over long sequences as each chunk conditions on its own imperfect output. In the authors' automatic and human evaluations, the model beats existing streaming video editors and stays competitive with slower offline systems on both short and long clips.
Weights are hosted on Hugging Face under jdopensource/JoyAI-Video-Edit, with an online demo and Apache 2.0 licensing on both code and checkpoints. The accompanying paper is arXiv 2608.03974.
Why it matters
Most generative video tooling is still batch-shaped: submit a prompt, wait, review. Real-time editing is a different product category — live streams, virtual production, on-set previsualization, video conferencing — and it has been gated less by quality than by latency and duration limits. A permissively licensed 16B model that clears 30fps on one GPU puts that category within reach of teams that cannot train such a system themselves, and it extends JD's pattern of shipping usable open multimodal models rather than frontier-chasing LLMs. It also raises the bar in an area where closed AIGC vendors have marketed real-time capability without exposing the underlying weights.