VGGT-Omega: Scaling Laws Meet 3D Reconstruction
VGGT-Omega demonstrates that the scaling laws proven in language and vision models also apply to 3D geometry. By introducing register attention and a streamlined architecture, the model trains on 15 times more data than prior work, scales to 10 billion parameters, and achieves state-of-the-art reconstruction on both static and dynamic scenes while running 50 times faster than optimization-based methods. The learned register tokens serve as compact scene representations that align naturally with language and robotic control tasks.Script
Scaling laws have transformed language and vision models, but do they work for 3D geometry? VGGT-Omega is the first feed-forward reconstruction system to show that more data and bigger models systematically improve 3D accuracy across orders of magnitude.
The architecture replaces expensive global attention with register attention. A small set of learned register tokens per frame exchanges information across the sequence, cutting training memory by 70 percent and enabling models with over 10 billion parameters.
Training uses 3 million sequences, filtered through multi-stage annotation that combines motion masking, feature tracking, and multi-view optimization. Strict consistency checks ensure high precision, even at the cost of discarding ambiguous data.
On the Sintel benchmark, VGGT-Omega improves camera accuracy by 77 percent and depth accuracy by 26 percent over the leading optimization method, while running 50 times faster. It handles extreme motion, sparse views, and dynamic content where prior methods break down.
Register tokens emerge as compact, reusable scene representations. Without architectural changes, they align with language embeddings for retrieval, improve robotic control when injected into policy models, and cluster to isolate motion, all from geometric training alone.
VGGT-Omega shows that scaling laws extend into 3D spatial reasoning, and that feed-forward models now rival traditional pipelines in both speed and accuracy. To explore the full paper and create your own video summaries, visit EmergentMind.com.
