Causal-Forcing

Why we included this project

Teams trying to get diffusion video generation down to interactive or streaming latency will find this repo a more useful starting point than the paper alone, because the code actually runs. It distills an autoregressive diffusion video model into one or two sampling steps, so text-to-video and image-to-video clips render fast enough for real-time use, and the same scheme extends to minute-level sequences. The codebase ships chunk-wise and frame-wise checkpoints built on Wan2.1, with CLI inference commands, per-stage training pipelines, and pretrained weights on Hugging Face, which means reproducing or adapting results does not require reimplementing them from scratch. Frame-wise models fold text-to-video and image-to-video into one setup, and the consistency-distillation variant here is a clearly different approach from earlier autoregressive forcing schemes.