diffusion-forcing

A Minimalist, Batteries-included Repository for Advancing World Model Science.

[ICLR 2026] 🐻 Uniform Discrete Diffusion with Metric Path for Video Generation

An open attempt at reproducing a simplified version of Google's Genie World Model API on a single H100.

nanoWM: a pose-conditioned causal Diffusion Transformer world model at nano scale, built to measure which training tricks (REPA, Muon, muP, diffusion forcing) pay off in 20 ticketed GPU-hours on Kaggle T4s. The deliverable is the measurement, not the model.

Train video world models with diffusion-forcing using this minimalist framework.