✦
Summary
PhiZero is a research world model built around physical language, a compact discrete representation of world-state transitions learned from in-the-wild video. It separates physical reasoning from video rendering so the system can reason about how a scene evolves before producing visual output.
Key Features
Builds a video world model around discrete physical-language tokens.
Separates physical reasoning from the final video rendering process.
Learns world-state transitions from in-the-wild video.
Provides a physical language tokenizer for compact dynamics representation.
Uses a physical language reasoner for motion and interaction prediction.
Supports demonstrations of dynamics transfer and physically coherent generation.
Links to a technical paper for deeper evaluation and method details.
Targets world-model research rather than a consumer editing interface.
About PhiZero
The approach uses a physical language tokenizer and a physical language reasoner to convert visual experience into reusable transition tokens. This lets the model describe motion-rich interactions, transfer body or object dynamics, and preserve physical coherence across generation and understanding tasks.
PhiZero is useful for researchers working on video world models, robotics-adjacent representation learning, and controllable video generation. Its explicit physical-language layer creates an interface for understanding, reusing, and rendering learned dynamics rather than treating video generation as a purely pixel-level process.