World Labs has introduced Atlas, a world model that reconstructs real-world scenes from as few as one image and turns them into explicit 3D outputs like point clouds and 3D Gaussian splats. The model is an omni model pretrained from scratch, built as a multimodal autoregressive diffusion transformer. It can also generate images and 360-degree panoramas from text prompts, follow complex instructions, and render text.
For video, Atlas can generate clips from one to six input images, and it can reframe existing video using as few as three cameras. It models both space and time from input videos, predicts the depth of every frame, and fills in regions no camera ever saw. The model can generate a one-minute video at 1440p resolution.
Atlas encodes inputs into a spatial context, using precise camera geometry as native input, and can generate views from any angle from a single image. It can also interpolate between unrelated image pairs and produce many different trajectories through the same scene. For reconstruction, it typically delivers faithful results with just two or three images, though it can use more than a hundred input images when needed.
The model is designed for Real-to-Sim workflows in robotics, letting users turn real-world captures into simulated environments. It supports manually designed camera paths and can generate aerial views from ground-level photos, as demonstrated by a reconstruction of Stanford's Main Quad from ground-level images. The company says the resulting Gaussian splat scenes render on a device at high frame rates and resolution, using the same representation found in World Labs' Marble.
World Labs says Atlas is built to scale with training compute, with performance improving as more compute is added, and that it handles a broad range of scene types and styles. The company claims it outperforms state-of-the-art 3D reconstruction models. No pricing, availability, or release date for Atlas was announced. The unveiling follows World Labs' earlier work on spatial intelligence, which the company has described as AI's next frontier.
