Fei-Fei Li's World Labs Unveils Atlas, Dubbed First Multimodal World Model

Its most direct feature is camera-controlled generation. Instead of relying on text prompts to approximate camera movement, Atlas takes camera pose parameters as native input. Given one to six ordinary photographs and a defined camera trajectory, it can generate footage of up to one minute at 1440p resolution, with consistent spatial geometry throughout the shot. Ifanr noted that when the camera moves into areas not covered by the original images, the model still guesses from prior knowledge, so wider viewpoint changes can expose local geometry drift, object shape changes, or texture shimmer. The result is a visually reasonable 3D world, not necessarily an exact digital replica of the original scene.

Atlas achieves this by binding each image to three-dimensional coordinates and depth, which ifanr describes as spatial context. When supplied with multiple pictures or videos from different angles, it fuses them into a continuous, navigable 3D scene after estimating their relative positions. Even two seemingly unrelated photos can be stitched together with generated corridors, doors, and rooms to create one space.

The model's second key capability is spatial reconstruction. In a Stanford Main Quad demo, 2 to 25 ground-level tourist photos were enough to regenerate lawns, arcades, and the exterior details of the memorial church, after which the camera could rise into an aerial view. On sparse-view reconstruction benchmarks including DTU, ETH3D, and ScanNet, Atlas scored 25.3 in absolute relative error, lower being better, compared with 28.7 for Pi3X, 39.3 for Depth Anything 3, and 47.7 for MapAnything. The output can be plugged into point clouds and 3D Gaussian Splatting, and connected to World Labs' Marble platform for high-frame-rate rendering.

Atlas can also produce reframing video, a kind of bullet-time effect. Instead of a Hollywood ring of cameras, a few consumer phones or action cameras placed around a scene are enough to let Atlas switch viewpoints in virtual space.

World Labs' larger goal for Atlas is embodied AI and robotics, according to the report. Building simulation environments for robot training traditionally requires labor-intensive 3D modeling of scenes, materials, and objects, while physical data collection is slow and expensive. Some industry estimates put the cost of conventionally collecting robot training data at up to $100 trillion. Atlas offers a real-to-sim path: a cellphone video of a factory or room can be turned into a high-precision 3D physical space where robots can repeatedly practice. It can render the red-green-blue image and depth map that an onboard camera would see, and simulate contact forces when a robotic arm touches rigid boxes, hinged cabinet doors, or soft sponges.

Technically, Atlas is built on a multimodal autoregressive diffusion Transformer. It combines autoregressive prediction of spatial states, rectified-flow diffusion for high-fidelity continuous signals, and a Transformer backbone. This architecture allows Atlas to benefit from techniques in both large language models and diffusion models, such as KV cache, distributed inference, sampling distillation, and prior guidance.

In blind user tests for camera-motion control, ifanr reported that Atlas was preferred over MiniMax H3 75% of the time, over Gemini Omni Flash 81%, over FLUX 3 93%, and over Seedance 2.5 94%. Atlas also has image generation ability and can create 360-degree panoramas from text or images, a task that demands stronger spatial understanding than ordinary text-to-image generation.

The company emphasized scaling as another keyword. According to its announcement, increasing parameter size and compute leads to emergent abilities similar to those seen in large language models. Atlas has entered early access with selected partners and is expected to become the underlying model for Marble and other World Labs products.

Ifanr placed Atlas in the arc of Li's career from ImageNet to vision and robot learning in her Stanford laboratory. Li has compared language models to text-savvy humans living in the dark: they can talk about reality but lack experience of being in it. Atlas represents an attempt to add world experience, covering geometry, physics, time, and action, to language intelligence, and could become the ticket for machines moving from the digital world to the physical one.