World Labs drops Atlas, an omni world model for 3D and video

Most video generation models still feel like slot machines. You type a prompt, pull the lever, and pray the camera pan doesn't turn your background into melted plastic.

World Labs, the startup founded by AI pioneer Fei-Fei Li, is taking a fundamentally different approach. On September 1, 2026, the company officially introduced Atlas, a next-generation omni world model built for spatial intelligence.

Instead of treating video as just another stack of flat pixels, Atlas was pretrained from scratch across text, images, video, and 3D data. Everything lives inside a shared spatial context where the model actually understands geometry and depth.

Pixel-perfect camera control

The marquee feature here is how Atlas handles motion. Most video models take vague text prompts like "pan left." Atlas takes actual, precise camera geometry as a native input.

That means you can hand-draw a camera path through a scene and let Atlas generate up to one full minute of coherent video at 1440p resolution. In one demo, the model took a single ground-level snapshot of a robot and imagined what was behind it, correctly figuring out the unseen backside and placing a plausible lawn next to the pool.

You can also drop two completely unrelated photos into different 3D coordinates. Atlas will literally build the world between them, hallucinating believable hallways, doors, and nooks that smoothly connect the two frames.

From a few phone snaps to 3D Gaussian splats

Atlas is also built to turn flat images into usable 3D environments. Traditional 3D reconstruction usually demands hundreds of dense photos or specialized LiDAR gear. Atlas pulls it off with just a handful of casual shots.

Give it two or three pictures, and it synthesizes novel angles while outperforming dedicated reconstruction models. In a test using Stanford's Main Quad, Atlas took a couple dozen ground-level snaps and generated smooth aerial flythroughs high above the campus.

Crucially, Atlas doesn't just export flat video. It natively outputs point clouds and 3D Gaussian splats that render on-device at high resolutions and frame rates. In one VFX experiment, engineers mounted a few ordinary smartphones on tripods to capture scenes from three to five angles; Atlas reconstructed the moment into a freeze-frame bullet-time rig.

For robotics, Atlas powers Real-to-Sim workflows. Using a 24-frame video sweep from a smartphone, it reconstructed a full room and generated the exact RGB and depth data a roving robot's sensors would see.

Under the hood, Atlas runs on a multimodal autoregressive diffusion transformer using rectified flow. World Labs notes that the model scales reliably with training compute and will power upcoming versions of its Marble platform. Early access applications are currently open.

Why it matters

While the rest of the industry is obsessed with prettier pixels, World Labs is solving the hard part: spatial consistency.

For filmmakers and 3D artists, pixel-perfect camera paths move generative AI from a lottery into an actual production tool. You are directing shots rather than hoping the prompt sticks. Real estate and VFX creators can turn casual phone clips into explorable Gaussian splats without hauling around thousands of dollars in capture rigs.

For robotics teams, the implications are huge. Gathering real-world training environments has always been a massive bottleneck. Being able to casually film a space with a phone and simulate robot navigation and manipulation data at scale cuts costs dramatically.

We think Atlas's biggest strength is treating 3D geometry and video as two sides of the same coin. The obvious catch is that when inputs are sparse, the model is still hallucinating what it cannot see, which won't cut it for precise engineering measurements. But for creative tools and synthetic simulation, it sets a brand new benchmark.