Understand
Build representations that connect video observations with objects, geometry, actions, and state.
SREAL AI · WORLD MODELS
Teaching video models to understand the physical world.
We develop physically grounded video models that learn consistent dynamics—building the foundations for general-purpose world models.
PHYSICAL CONSISTENCY MODEL TRACE 01
01Observed
02Predicted
03Verified
From visual patterns to learnable physical laws
2workshop best paper awards
5core world-model capabilities
Demovalidated with industry pilots
OUR FOCUS
From pixels to physical laws.
Today’s video models can generate compelling motion. Our work asks a harder question: can they learn why the world moves the way it does?
Reason
Verify whether predicted changes are visually plausible, physically consistent, and reachable by action.
Generate
Create coherent, controllable futures across multiple agents, viewpoints, and interactive scenes.
GENERATIVE WORLD ENGINE
A foundation for interactive, physically intelligent worlds.
Our generative world engine brings video understanding, physical reasoning, scene generation, multi-agent interaction, and real-time simulation into one research-driven platform.
Built forGamesInteractive contentEmbodied intelligence
01Video understanding
02Physical reasoning
03Scene generation
04Multi-agent interaction
05Real-time world simulation
SELECTED RESEARCH
Evidence, not promises.
Open research is how we make progress legible. Explore the work, methods, demos, code, and results behind our technical direction.
MultiWorld
Scalable Multi-Agent Multi-View Video World Models
A scalable multi-agent, multi-view video world model for action-controllable and view-consistent generation.
World Action Verifier
Self-Improving World Models via Forward–Inverse Asymmetry
A verification framework that helps world models identify prediction errors and self-improve through forward–inverse asymmetry.
Research contributed by SREAL AI Research Intern Weifeng Lu.
FOUNDER
Yingtian Zou
Founder of SREAL AI · Assistant Professor, School of Artificial Intelligence, Shanghai Jiao Tong University
Yingtian’s research explores the theory of foundational world models, self-improving systems, and consistent, interactive, efficient generative world models.
Academic profileBUILD THE NEXT WORLD MODEL
Research with us. Build with us.
We welcome conversations with researchers, builders, game studios, and AI companies exploring physically intelligent generative systems.