World Labs Atlas Omni World Model
Announced by World Labs on 1 September 2026 and described by the company as an omni world model pretrained from scratch to natively operate on text, images, video and 3D
Tool Overview
Features, steps and FAQ below
Features
- ✓ Camera-controlled generation: generates images and video from one or more reference images with pixel-perfect camera control, outputting up to 1 minute of 1440p video
- ✓ Spatial reconstruction: rebuilds real-world scenes from one to dozens of images, typically giving faithful reconstructions from as few as two or three, and World Labs says it outperforms models trained only for 3D reconstruction
- ✓ Space-time simulation: reconstructs and reframes video from a few cameras for bullet-time VFX and Real-to-Sim robotics navigation and manipulation
- ✓ Explicit 3D output: natively works on 2D frames and 3D depth maps to produce point clouds or 3D Gaussian splats, the same representation used in Marble
- ✓ Image generation: generates images and 360 panoramas from text, following complex prompts, rendering text and producing a wide range of visual styles
How to Use
- Request early access from the World Labs page
- Prepare one to six reference images, or more for reconstruction, or a short phone video as input
- Specify a camera path or position to generate new views, or run sparse-view 3D reconstruction
- Export point clouds or 3D Gaussian splats when you need explicit geometry for robotics or design workflows
FAQ
What is Atlas?
World Labs' next-generation world model announced on 1 September 2026, which the company describes as an omni model pretrained from scratch to natively operate on text, images, video and 3D.
How is it different from ordinary video generators?
World Labs says Atlas offers pixel-perfect camera control, generating views from any angle from a single input image while staying 3D-consistent, rather than only following a text prompt.
How long and sharp can its video be?
In World Labs' examples it generates 1-minute, 1440p video from a small number of reference images with a hand-designed camera path that keeps the scene coherent.
How many photos does reconstruction need?
World Labs says two to three images typically give a faithful reconstruction and it can use over a hundred inputs; the more it sees, the less it imagines.
Can it output usable 3D assets?
Yes. World Labs says it works natively on 2D frames and 3D depth maps to output point clouds or 3D Gaussian splats for gaming, design, VFX and robotics.
Is it available now?
Atlas is entering early access with select partners and you can request access on the World Labs page; the company says it will power future versions of Marble and other products.