Atlas (World Labs) — multimodal world model announcement
- Type
- model
- Venue
- World Labs blog / X announcement
- Year
- 2026
- Source
- x
- Access
- public
- Language
- en
- Added
- 2026-09-29
- Verified
- 2026-09-29
Summary
World Labs introduced Atlas, described as the world's first multimodal world model pretrained from scratch: a multimodal autoregressive diffusion transformer operating natively on text, images, video, camera poses, and 3D depth maps within a shared spatial context. It generates image and video frames with pixel-perfect camera control (up to 1 minute of 1440p video from 1-6 reference images plus a camera path), reconstructs scenes in 3D (point clouds, 3D Gaussian splats), reframes multi-camera footage, and supports Real-to-Sim robotics workflows. In company evaluations it reported 75-94% win rates on camera-controlled generation vs MiniMax H3 and Gemini Omni Flash, and sparse-view reconstruction error of 25.3 per mille vs Pi3X and Depth Anything 3.
Keywords
world models · spatial intelligence · video generation · 3D reconstruction · robotics · World Labs
Topics
world models, spatial intelligence, video generation, 3D reconstruction, robotics
Research notes
- Discovery: Announcement X post from World Labs; underlying work is the World Labs Atlas launch blog post of 2026-09-01.
- Method: Omni model pretrained from scratch; every image and depth map anchored to an explicit camera pose in 3D space forming a spatial context, analogous to how an LLM encodes token context but over physical space.
- Key findings: Single architecture natively handles text, images, camera poses, depth maps; video represented as image sequences; Four task families: camera-controlled generation, spatial reconstruction, space-time simulation, image/360-panorama generation; Intended to power future versions of Marble and other World Labs products
- Limitations: Launch claims are company-reported with no independent benchmark reproduction as of early September 2026; Atlas in selective early access only, absent from the public World API model list, with no public price, model card, latency disclosure, or named partner list. 'World's first' framing is a company claim, not independently established.
- X post login-walled; substance verified via secondary coverage (SiliconANGLE, kingy.ai, startupfortune, dev.to, aoyutech) of the 2026-09-01 launch. Separately, unite.ai reported (~2026-09-28) AMD acquiring World Labs in an $8.2B all-stock transaction.