Self-Forcing demonstrates real-time autoregressive video
Video2026-03-30
Source: self-forcing.github.io
A new training recipe generates video frame-by-frame in real time, a key step toward interactive generative video on a single GPU.
Diffusion video models denoise a whole clip at once, which is why they're slow and hard to make interactive. Self-Forcing takes an autoregressive route — generating one frame after another — and trains the model on its own outputs so errors don't compound. The payoff is real-time generation on a single GPU.
It's research code today, but it points directly at the live, steerable video the video tools are racing toward.