What is Diffusion Forcing?

A generative method blending next-token (autoregressive) prediction with diffusion: each element of a sequence is denoised at its own noise level. Enables flexible-length, more stable sequence generation (video, action planning).