World Labs' Atlas Beats Specialized 3D Models With One Omni Model

World Labs, the spatial intelligence company founded by Fei-Fei Li, pulled the covers off Atlas, a foundation model that treats world generation, 3D reconstruction, and simulation as facets of the same problem. Atlas is pretrained from scratch to natively operate on text, images, video, and 3D as a multimodal autoregressive diffusion transformer, with all inputs combined into a shared spatial context. A single model can now be steered with camera geometry rather than fuzzy text prompts, and it outputs both pixels and explicit 3D geometry.

Directing the camera, not spinning the slot machine

The headline capability is precise camera control. Atlas takes camera geometry as a native input type, going beyond coarse text-based instructions and letting you frame every shot and control every motion. In one demo, the team hand-designed a camera trajectory through a scene, and Atlas generated a one-minute video at 1440p from just seven reference images.

The underlying trick is what World Labs calls a spatial context. Similar to an LLM, Atlas first encodes its inputs into a context, then generates outputs conditioned on that context, but each image is grounded at a 3D position in space. Drop two unrelated reference images into that context at different positions and the model will invent the hallways, doorways, and terrain that connect them.

What it can actually do

Atlas is pitched as an omni model that spans four task families:

• Camera-Controlled Generation: generates images and videos from one or more images with pixel-perfect camera control, outputting up to 1 minute of video at 1440p.
• Spatial Reconstruction: reconstructs real-world scenes from one to dozens of input images, generating both novel-view frames and explicit 3D outputs, outperforming state-of-the-art models specialized for 3D reconstruction.
• Space-Time Simulation: models space and time from input videos, reframing footage for dramatic visual effects and enabling real-to-sim workflows for robotics.
• Image Generation: text-to-image plus 360 panorama generation with complex prompt following and text rendering.

The spatial reconstruction story is worth pausing on. Passing more input images gives Atlas more context: the more it sees, the less it imagines. It typically produces faithful reconstructions with as few as two or three images and scales to over a hundred input images in its spatial context. Outputs are not limited to 2D frames either. Atlas natively operates on both 2D image frames and 3D depth maps, so it can emit worlds as point clouds or 3D Gaussian splats.

The architecture, unpacked

The design blends two lineages that have mostly evolved separately, drawing on ideas from modern LLMs and video models so it can benefit from architectural, algorithmic, and systems advances used in both. Each of the four words in the name pulls its weight:

• Multimodal: it operates on text, images, camera poses, and 3D depth maps, with videos represented as sequences of images and each image and depth map conditioned on an explicit camera pose.
• Autoregressive: outputs are generated one element at a time, so any task becomes a different kind of sequence where inputs are followed by outputs.
• Diffusion: it is a rectified flow model that generates outputs by gradually denoising them, naturally trading off speed and quality by varying the number of denoising steps.
• Transformer: the standard backbone lets Atlas inherit LLM-serving tricks like KV-caching, cache-aware routing, and disaggregated serving, plus video-model tricks like diffusion distillation and shifted noise schedules.

The benchmark numbers

World Labs ran head-to-head evaluations on the two tasks with clean ground truth. On camera-controlled generation, third-party human raters compared Atlas against recent video models given a single input image and a sequence of cinematic camera motions. Raters preferred Atlas over MiniMax H3 in 75% of trials, Gemini Omni Flash in 81%, Happy Horse 1.1 in 86%, FLUX 3 in 93%, and Seedance 2.5 in 94%. The gap grows with camera trajectory complexity, the expected failure mode when other models can only ingest camera moves through text prompts.

On sparse-view 3D reconstruction, Atlas is compared to specialist models like VGGT, Depth Anything 3, MapAnything, and Pi3. Measured as mean absolute-relative pointmap error (lower is better), Atlas averages 25.3 across benchmarks versus 28.7 for the next best, with particularly large margins on DTU (8.6 vs 11.1) and T&T (12.4 vs 15.7). The comparison is against open-source baselines, not closed commercial reconstruction systems.

Where the real leverage lives: robotics and VFX

Two use cases feel like the actual point of the release. The first is bullet-time video effects on a shoestring: with footage from as few as three cameras, Atlas can freeze time and reframe shots, letting you view events from impossible angles, filmed by a few engineers with ordinary cell phones on tripods and clamps that fit in a backpack.

The second is real-to-sim for robots. Reconstructing a warehouse or kitchen usually requires LiDAR rigs or dense photogrammetry captures. With Atlas, the pitch is that you walk through with a phone. As a simulated robot moves through space, Atlas also generates the RGB and depth data its sensors would observe along the way, so the world and the robot's view of it come from the same model. That single-model property matters because it removes the seams between reconstruction, rendering, and sensor simulation that usually plague synthetic training pipelines.

Availability and where it fits

Atlas is not a public API yet. It is entering early access with select partners and will power future versions of Marble and other World Labs products. The early access form is live for teams that want to build on it.

For context on the trajectory, World Labs previously made Marble generally available as a massively multimodal world model that creates 3D worlds from text, images, video, or coarse 3D layouts, with editable outputs exportable as Gaussian splats, meshes, or videos. Atlas is the foundation model that will sit underneath the next generation of that product, and it slots into a crowded field that now includes Google DeepMind's Genie 3 and Tencent's HunyuanWorld.

Priors worth updating

A few assumptions worth revisiting after this release:

1. Camera control as a first-class input. Text prompts for camera moves have been a losing abstraction. Native pose conditioning is likely the direction the whole video-model stack moves toward.
2. Generation and reconstruction are the same task. Atlas beating specialized 3D reconstruction models with a general omni model is the same story that played out with LLMs versus task-specific NLP systems.
3. Real-to-sim is getting cheap. If a couple of phone videos can produce navigable, sensor-accurate environments, robotics data pipelines that rely on Gaussian splat capture rigs or Unreal Engine hand-authoring have a much cheaper competitor to benchmark against.