Self-Forcing

Why we included this project

Self Forcing is the official code for a NeurIPS 2025 Spotlight paper that tackles a long-standing mismatch in autoregressive video diffusion: models are trained one way but run another, and quality suffers. The fix is to simulate the rollout during training with KV caching, so the model behaves at inference the way it learned. The payoff is practical: it streams video in real time on a single RTX 4090 instead of rendering a full clip first. You get pretrained checkpoints and a GUI demo, so you can try it without wrestling with the math. Worth a look if you build streaming or interactive video tools and want a working reference for how to bridge that gap.