World Labs · 2026-09-01 · major
Atlas — World Labs' omni model for text, image, video and 3D
Atlas is World Labs' new world model, pretrained from scratch to work on text, images, video and 3D at once. It makes up to one minute of 1440p video with exact camera control and rebuilds 3D scenes from a handful of photos.
World Labs' Atlas is one model for text, images, video and 3D, with pixel-level control of the camera.
Key specs
| Human preference vs rivals | 75–94% |
|---|---|
| 3 d reconstruction error | 25.3 (next best 28.7) |
Quick facts
| Maker | World Labs |
|---|---|
| Model type | Multimodal autoregressive diffusion transformer |
| Inputs | Text, images, camera poses, 3D depth maps |
| Video output | Up to 1 minute at 1440p |
| Availability | Early access with select partners |
| Weights and code | Not released |
What is it?
Atlas makes video you can steer with a virtual camera — up to one minute at 1440p — and rebuilds real 3D scenes from as few as one photo. World Labs pretrained Atlas from scratch as what it calls an omni model, working natively on text, images, video and 3D instead of bolting 3D onto a video generator. It follows Marble, the company's first product, which builds persistent 3D worlds from images, video, text and 3D layouts.
How does it work?
A multimodal autoregressive diffusion transformer sits at the core: it reads text, images, camera poses and 3D depth maps in one sequence, then generates using rectified flow diffusion. Because the camera pose is a real input rather than a hint buried in a text prompt, Atlas can follow an exact camera path instead of guessing one. The same backbone covers all four jobs — camera-controlled generation, spatial reconstruction from sparse images, space-time simulation, and text-to-image or 360 panorama.
Why does it matter?
VFX artists and robotics teams normally chain a video generator to a separate 3D reconstruction tool; Atlas does both in one model and returns explicit 3D geometry, not just frames. Human raters preferred it over five rival models in 75% to 94% of camera-controlled comparisons, and its mean absolute-relative reconstruction error of 25.3 beats the next-best baseline at 28.7. The catch is access: Atlas is early access with select partners, and World Labs has published no weights, code or paper.
Who is it for?
VFX studios, robotics teams, 3D developers
Frequently asked questions
- Is Atlas open source?
- Atlas is not open source. World Labs has not published weights, code or a technical paper alongside the September 1, 2026 announcement. The company is instead taking early-access requests from select partners through a form, so teams that want to build on Atlas today have to apply rather than download anything.
- How do I get access to Atlas?
- World Labs takes Atlas early-access requests through a form at form.typeform.com/to/zHFR4r3A, and says the model is entering early access with select partners. That program is separate from Marble, World Labs' existing 3D world product, and from the company's developer platform, both of which you can use without an Atlas invitation.
- How does Atlas compare to Seedance 2.5, FLUX 3 and other video models?
- On camera-controlled generation, human raters preferred Atlas over five named rivals: 75% of the time against MiniMax H3, 81% against Gemini Omni Flash, 86% against Happy Horse 1.1, 93% against FLUX 3 and 94% against Seedance 2.5. Atlas also differs in kind, since it outputs explicit 3D geometry rather than video frames alone.
- What is the difference between Atlas and Marble?
- Marble is World Labs' first product and builds persistent, high-fidelity 3D worlds from images, video, text and 3D layouts. Atlas is the newer general-purpose model announced September 1, 2026, pretrained from scratch to handle text, images, video and 3D in a single system that covers generation, reconstruction and simulation.
Try it
Request early access at https://form.typeform.com/to/zHFR4r3A