World Labs unveils Atlas, a multimodal world model with pixel-level camera control

World Labs announced Atlas, which it describes as the first multimodal world model capable of generating image and video frames with pixel-precise camera control and reconstructing them in 3D space. The company said, “Atlas is the first multimodal world model of its kind.”

According to the company, this design combines the strengths of modern LLMs (LLM) and video models, benefiting from recent architectural, algorithmic and systems advances in both fields.

The core capability of Atlas is to accept multimodal input—including a defined camera motion—and translate it into spatially grounded 3D views. Users can steer viewing angles, reconstruct real spaces from their own captures, and build explorable worlds. In a published example, a hand-designed camera trajectory produced a one-minute video at 1440p resolution from only seven reference images. The model integrates multiple input views within its spatial context, enabling frame creation with precise control.

Beyond video generation, Atlas extracts an explicit 3D representation from a variable number of input images, ranging from a single picture to many. The company claims the performance surpasses that of leading open-source reconstruction models, though no benchmarks, model names, or testing protocols were disclosed. The principle is that supplying more images enriches the spatial context, reducing geometric hallucinations.

Atlas is also engineered to comprehend both the spatial layout of a scene and its temporal evolution. This allows a small set of ordinary cameras to function as a multi-angle bullet-time capture rig, freezing dynamic moments in time. For robotic simulation, the model can reconstruct an environment from few images and generate photorealistic RGB and depth data as a robot’s sensors would perceive along any trajectory. The practical implication is that robots can now train and be evaluated across a far broader range of environments without the need for dedicated, expensive scanning of each space.

World Labs said early-access to Atlas will be opened in the coming weeks through a dedicated sign-up form. At this stage no information was provided about the business model, licensing restrictions, hardware requirements for local execution, or a timeline for releasing model weights or code. The announcement focused on demonstration capabilities and the underlying architecture, without revealing training data, parameter count, or compute costs.