World labs unveils atlas, the first multimodal world model with pixel-level camera control

On September 2, 2026, World Labs, the startup founded by AI pioneer Fei-Fei Li, launched Atlas, the first multimodal world model capable of generating images and video frames with pixel-level camera control. The model can reconstruct 3D scenes from input images, offering precise control over camera movements and enabling effects like 'bullet time'. Atlas uses a from-scratch pre-training approach to process multimodal inputs, including camera motion, and generate 3D views. It can create images and videos up to 1440p with a duration of one minute, and reconstruct real-world scenes from one to dozens of input images, outperforming existing open-source models. The model also supports text-to-image generation and 360-degree panoramas, adapting to complex prompts and various visual styles. World Labs has already granted early access to some partners, who will receive the model in the coming weeks.