diffusion-forcing-transformer

Why we included this project

Diffusion Forcing Transformer (DFoT) is the official ICML 2025 codebase for History-Guided Video Diffusion, and it tackles a genuinely awkward problem in video generation: how many context frames to condition on. Because it accepts any number of them, the same model can animate a single still photo, extend a short clip into a longer one, or interpolate between two distant frames. The History Guidance methods that ship with it keep temporal consistency and motion quality stable even when rolling out very long sequences, which is where most video models fall apart. Pretrained checkpoints and a detailed wiki make results reproducible, and a browser demo on HuggingFace Spaces lets you try it before committing a GPU. It is a research codebase rather than a polished product, so plan on your own training setup, though running inference with the provided weights is straightforward.