JoyAI-Video-Edit

JoyAI-Video-Edit edits video in real time from a natural-language instruction: point it at a live camera stream, tell it to swap the background or restyle the subject, and it edits frames as they arrive rather than waiting for the full clip. It's from JD.com's research arm, and both the code and the model weights are Apache 2.0, which is rarer than it should be.

Now the hardware reality: the headline real-time number comes from a single NVIDIA B200, a data-center accelerator costing tens of thousands of dollars. The README doesn't say this; the paper does. Consumer GPU support is on the roadmap, not in the release. Setup is conda, CUDA wrangling, and external model dependencies that aren't bundled.

Teams with data-center GPU access building live video pipelines, and researchers working on autoregressive diffusion, are the real audience. Runway and Pika are the hosted comparisons; Wan and HunyuanVideo are the open peers, though none target causal real-time editing.

The catch: training pipelines aren't released, so you can run the model but not reproduce it, and a stronger checkpoint is already announced, making this one a waypoint. Free in license, expensive in silicon.