Universal Physical Intelligence.
Bagel Labs is a physical AI research lab based in San Francisco and Toronto. It has developed an architecture that uses dramatically less data and compute to train while delivering a state-of-the-art model-size/performance tradeoff.
Decentralized Diffusion Models.
DDM (Decentralized Diffusion Models) is a training architecture that replaces one monolithic diffusion model with independently trained experts and a lightweight router. Each expert trains on a partition of the data without gradient synchronization, making training more data- and compute-efficient.
Paris-1.
Paris 1.0 demonstrates the training advantage for image generation. At matched aggregate compute, it improves FID by 24 percent over a monolithic baseline, paving the way for Paris 2.0.
| Inference Strategy | FID-50K ↓ |
|---|---|
| Monolithic (single) | 29.64 |
| Top-1 | 30.60 |
| Top-2 | 22.60 |
| Full Ensemble | 47.89 |
| Improvement | 7.04 |
Paris-2.
Paris 2.0 extends the same decentralized training architecture to video. Using the same data and matched total compute, it cuts FVD roughly in half compared with a monolithic baseline, paving the way for physical AI world models.
