2 September 2026

World Labs releases Atlas model for 3D scene generation

• Atlas takes a few smartphone photos and generates full 3D scenes viewable from any camera angle, outperforming specialized 3D reconstruction models in testing.
• The model processes text, images, video, and 3D data together in a shared spatial context rather than as separate sequences, anchoring everything to positions in 3D space.
• Atlas can simulate robot training by reconstructing rooms from photos and generating varied sensor data, eliminating need to capture every real-world scenario.
• The model outputs up to one minute of video at 1440p with direct geometric camera control rather than text prompts, and produces actual 3D data formats like point clouds.

Where they differ

All three newsletters accurately reported the core announcement.

• TLDR AI

  focused on the underlying architecture.

• The Neuron

  emphasized the phone video to 3D reconstruction angle.

• The Rundown AI

  stressed early access availability. None significantly diverged on the facts.

What each one reported

TLDR AITLDR editorial team

Atlas is a world generation model pretrained to operate natively on text, images, video, and 3D data, combining inputs into shared spatial context. The model performs broad tasks spanning world generation, reconstruction, and simulation, with performance improving as training compute increases.

The NeuronPete Huang & Grant Harvey

World Labs, led by Dr. Fei-Fei Li, released Atlas, a world model that reconstructs 3D space from ordinary phone clips and generates new camera angles that were never filmed. The model can create new shots, 3D geometry, or training environments for robots from the same footage.

The Rundown AIRowan Cheung

Fei-Fei Li's World Labs introduced Atlas in early access, a world model that converts a few smartphone photos into full 3D scenes or video with camera control.

Reported by The Decoder, SiliconANGLE