World Labs' Atlas rebuilds a full room from just 3 photos, a 50-100x capture reduction
drfeifei · x · 2026-09-05
World Labs co-founders Fei-Fei Li, Justin Johnson, and Ben Mildenhall sat down with a16z's Martin Casado to discuss Atlas, their world model for spatial intelligence.
- Core idea: LLMs are built on next-token prediction and video models on next-frame prediction; Atlas is built on new-view prediction, and is claimed to be the first model unifying pixel generation and pixel reconstruction—two tracks computer vision has kept separate for half a century.
- Practical result: Dense reconstruction of a room previously required 100-300 photos; Atlas works from just three, a 50-100x reduction in capture cost.
- Implication: That flip changes what imagery is worth reconstructing—old photos sitting in your camera roll can now be turned into 3D scenes. Mildenhall explains the key is finally reconciling data-driven priors with brute-force dense reconstruction.
More from Models
- GPT-6 Astra and Claude Fable 5.1 write $10,000 research papers — experts reject all — sayashk · 2026-10-21
- GitHub Copilot's HydraFusion Orchestrates Multiple Models Per Task — WirelessLife · 2026-10-21
- OpenVLHarness turns any MLLM into a visual reasoner, +9.9 points across 22 datasets — yining_hong · 2026-10-21
- User claims Gemini analyzed his tennis match video and improved his game — zacharynado · 2026-10-21
- Opus 5.5 tops OpenRouter spending in 20 of 29 task categories — mmurph · 2026-10-21
- Meta team builds open-source 30B LLM from scratch in months, hails 'second era' — TimDarcet · 2026-10-21