What's this about?
A 16B-parameter autoregressive-diffusion model that edits video in real time as it streams — no access to future frames, no fixed clip length — while preserving source fidelity and long-term consistency (56↑, day's top paper).
What exactly is "autoregressive diffusion" for video?
Diffusion generates by denoising; autoregressive means it produces the video chunk-by-chunk in order, each new piece conditioned on what came before. Combining them lets the model stream output causally (like live video) instead of rendering a whole fixed clip at once — the difference between a live filter and an offline render.
Why does it matter?
Real-time, unbounded video editing is the hard version of video generation — it's the capability behind live filters, streaming effects, and interactive video, not just "make me a clip." Open work at this latency is a meaningful step.