Phi-0

PhiZero (Phi-0) is a research world model that learns a compact discrete representation of world-state transitions, called physical language, from unlabeled in-the-wild videos using self-supervision. It follows a reason-then-render approach: an autoregressive vision-language model predicts a physical-language sequence from the current frame and a text action intent, and a diffusion decoder renders that sequence into video.

The tokenizer uses a transition-level Q-Former with FSQ discretization, a Wan2.2 VAE encoder and a Wan2.2-5B diffusion decoder. The reasoner is initialized from Qwen3-VL-4B with a vocabulary extended by 25K symbols. A four-second, 33-frame clip is encoded with 256 discrete symbols, versus 44,800 continuous tokens for the Wan2.2 VAE. Training used a 50K-hour video pool filtered to 10K hours for tokenizer pretraining, then 5M clips and 1M motion-rich clips.

The project page reports leading scores on IQ-Score, Physics Score, Physics Adherence and IntPhys2 Overall among the compared models. Demos cover physically realistic video generation, action-conditioned simulation, interactive image-to-world exploration, driving and robotics world models, and zero-shot motion transfer across objects, embodiments and domains. These results come from the authors. A YouTube review says code is "coming soon", and no code, weights or paper link were found on the page.