What do we know about Atlas, the AI model created by Fei-Fei Li’s World Lab?

The model can also combine different photographs, position them virtually within a space and imagine corridors, doors or passageways connecting them. It is a powerful feature, but also the point at which reconstruction and generation risk becoming confused: when information is lacking, Atlas does not uncover the hidden reality, but rather constructs a plausible solution.

The company states this quite openly. With just a single image, the system has to work out much of what it cannot see; by adding photographs, the reconstruction becomes more accurate. In some cases, two or three views are sufficient, whilst to reproduce a location with greater precision, Atlas can process more than a hundred.

The result can be output as a video, a point cloud or using 3D Gaussian splats – a technique that represents a scene using millions of coloured three-dimensional elements and allows it to be explored in real time. The most obvious applications are in cinema, video games, design, special effects and virtual environments.

World Labs claims that Atlas outperforms certain video models in its ability to track complex camera movements, as well as various open-source systems specialising in 3D reconstruction. The comparisons were carried out, respectively, using human evaluators and academic benchmarks based on reconstruction error.

However, these are results presented by the company itself. No full scientific paper has been published setting out all the details of the model, the training data and the evaluation procedures, nor is Atlas yet freely available: it is entering an early access phase with a select group of partners.

Atlas is not alone. Google DeepMind has developed Genie 3, which is capable of creating navigable environments in real time at 720p and 24 frames per second from a single prompt, whilst maintaining them consistently for several minutes: virtual worlds in which to train and evaluate autonomous agents. Microsoft Research, with Muse — based on the World and Human Action Model (WHAM) — generates gameplay sequences by predicting how the scenario will change in response to the player’s commands; for now, it is primarily aimed at video game design, rather than robotics. Nvidia Cosmos, on the other hand, offers models and tools specifically designed for ‘physical AI’, with which to produce synthetic videos and simulations for robots and autonomous vehicles. Meta’s V-JEPA models are working in the same direction, learning a representation of the world by predicting how a visual scene will evolve. The difference lies in the starting point: Genie focuses on generated interactive worlds, Muse on the dynamics of video games, Cosmos on data for autonomous machines; Atlas focuses primarily on the controllable spatial reconstruction of the real world.