World Labs' Atlas Unifies 3D Reconstruction and Video Generation in One Model

World Labs, the spatial intelligence startup cofounded by Fei-Fei Li, has unveiled Atlas, a foundation model that handles world generation, 3D reconstruction, and simulation within a single architecture. The system is entering early access with select partners.

World Labs calls Atlas an omni model: pretrained from scratch to natively process text, images, video, and 3D data through a multimodal autoregressive diffusion transformer. Every input is folded into a shared spatial context, with each image anchored to an explicit 3D position. That design choice is what separates Atlas from the video and 3D specialist models currently dominating the field.

One model, four jobs

Atlas spans four capability families rather than stitching together a pipeline of specialists:

• Camera-controlled generation: Camera geometry is a native input type, returning up to one minute of video at 1440p from one or more reference images.
• Spatial reconstruction: Faithful 3D reconstruction from as few as two or three images, with the spatial context able to hold over a hundred.
• Space-time simulation: Reframes video for visual effects and enables Real-to-Sim workflows for robotics.
• Image generation: Produces images and 360° panoramas from text, follows complex prompts, renders text, and covers a wide range of visual styles.

Camera geometry as a first-class input

Standard video models receive a text prompt like "slow crane shot over a garden" and approximate the camera movement through training. Atlas takes the camera pose as a direct numeric input. Each image in the spatial context is grounded at a specific 3D position, so the model knows not just what it has seen but where each observation sits in space.

Drop two unrelated reference images into the context at different 3D positions, and Atlas will hallucinate the corridors and transitions connecting them. Add a third real photograph of the same scene and the invented geometry collapses into faithful reconstruction. World Labs demonstrated this on Stanford's Main Quad, generating aerial camera paths from two to twenty-five ground-level photographs.

Under the hood

The architecture is a rectified flow model combining autoregressive and diffusion mechanics. Autoregressive structure lets it reuse KV caching and disaggregated serving optimizations; the diffusion side lets it trade compute for quality at inference time by varying denoising steps. Videos are represented internally as image sequences, each conditioned on an explicit camera pose, with previously generated frames accumulated into the shared spatial context to maintain 3D consistency. World Labs says a real-time version for interactive exploration has also been optimized.

Benchmark numbers

On camera-controlled generation, third-party human raters preferred Atlas over competing video models in 75 to 94 percent of head-to-head trials. The breakdown by competitor: 75% over MiniMax H3, 81% over Gemini Omni Flash, 86% over Happy Horse 1.1, 93% over FLUX 3, and 94% over Seedance 2.5.

World Labs flags a meaningful caveat: baseline models received the camera path as a text description because they have no native camera input, an asymmetry that skews the comparison. On 3D reconstruction, Atlas posts lower absolute-relative pointmap error than open-source specialists including VGGT, Depth Anything 3, MapAnything, and Pi3X across DTU, ETH3D, KITTI, ScanNet, and Tanks and Temples benchmarks. One of those comparison models had flagged possible benchmark contamination in the weeks before this release, a detail critics have noted.

What is missing

This release is a research post, not a product launch. No paper, arXiv entry, model card, evaluation code, dataset splits, or model outputs have been published. No pricing, API, or public release date has been announced, and access requires submitting a request form. Atlas will eventually power future versions of Marble, the company's existing tool for building explorable 3D worlds that exports to Gaussian splats and meshes.

Real-to-sim on a phone budget

Capturing a real environment for robotic simulation traditionally requires LiDAR rigs and photogrammetry pipelines. Atlas reconstructs full environments from roughly 24 frames of phone video, then generates the RGB and depth streams a simulated robot's body-mounted cameras would see while navigating that reconstruction. Object interactions, positions, lighting, and backgrounds can all be swapped to produce diverse training data.

The same reconstruction capability drives a VFX demo. Three to five phone angles plus action cameras are enough to freeze time and reframe a scene from any virtual camera position.

The architectural bet

Atlas enters a field that now includes Google DeepMind's Genie 3 and Tencent's HunyuanWorld. The pitch is architectural: if generation and reconstruction are facets of the same problem, maintaining separate NeRF or Gaussian splat pipelines, separate video diffusion models, and separate depth estimators starts to look like unnecessary duplication. A few implications worth watching:

1. Text prompts for camera moves may be a dead end. Native pose conditioning produces cleaner control and, in these benchmarks, meaningfully better results.
2. The 3D specialist gap may be closing. A generalist trained on enough multimodal spatial data can match or beat purpose-built reconstruction networks.
3. Real-to-sim gets cheap. A few minutes of phone video may be sufficient to build a simulator with matching sensor streams, which reshapes the robotics data pipeline.

Atlas is currently a curated demo reel behind a request form. Latency, cost, and out-of-distribution quality will only become visible once it ships inside Marble and reaches production workloads.