Accelerating the search for antimicrobial peptides with Google Antigravity
deepmind.google·2h ago
AI Research
arXiv CS.AI·1d ago·5 min read
World Models Need a Physical Registry: Learning Internal Predictive States for Physically Plausible World Models
TL;DR
A new approach called the Physical Registry enhances video world models by explicitly integrating predictive dynamics, leading to significant performance improvements in generating physically plausible futures.
✦ Why It Matters
Implement the Physical Registry in your video generation models to improve predictive accuracy and physical realism.
Key Takeaways
- Integrate the Physical Registry for better predictive dynamics in video models.
- Leverage the 9.9 point improvement in I2V tasks for enhanced performance.
- Consider using smaller models like Cosmos 3 Nano for efficiency.
- Focus on explicit state representation for improved physical plausibility.
How It Works
The Physical Registry consists of slots that are temporally grounded, allowing the model to learn and represent predictive dynamics explicitly. These slots are initialized from forecasts based on observations and evolve alongside video tokens, receiving supervision from future representations.
This co-evolution enables the model to maintain a clear understanding of how scenes should evolve over time.