World Labs releases Atlas world model, unifying generation, reconstruction, and spatiotemporal simulation.

World Labs has released Atlas, a new generation of world modeling that integrates generation, reconstruction, and spatiotemporal simulation capabilities into a single architecture.Atlas is pre-trained from scratch and can natively process text, images, videos, and 3D data; its core is the multimodal autoregressive diffusion Transformer, which combines different inputs into a shared spatial context.

The tasks officially showcased include: generating videos with controllable camera paths from one to six reference images, reconstructing scenes and explicit 3D results from sparse images, modeling space and time from video, and generating images and 360-degree panoramas from text. Performance conclusions are currently primarily based on World Labs' own evaluations.

refer to:World Labs Release Page