World Labs releases Atlas multimodal world model

• Atlas is a from-scratch multimodal world model, as drfeifei's launch post describes it, with generation, scene reconstruction, video reframing, and 3D output in scope.
• Generation and reconstruction share one model, according to jcjohnss's model description, rather than appearing as separate product pipelines.
• For robotics, eerac's robotics post lists image registration, 3D reconstruction, novel-view generation, and native RGB-plus-depth operation.
• A three-iPhone capture can yield a smooth synthesized camera move, as shown in eerac's juggling demo.

The technical post makes camera poses a native part of Atlas' data model: every image and depth map is paired with an explicit 3D position, while video is represented as an image sequence. The multiview pipeline sends raw footage and a target trajectory to direct frame generation, according to jcjohnss's reply, without a Gaussian-splat intermediary. World Labs also says the same context can join unrelated reference images by imagining the transitions between them.

World Labs' architecture description calls Atlas a multimodal autoregressive diffusion transformer. Its defining interface is camera geometry, not only a cinematic text instruction.

• Multimodal: text, images, camera poses, and 3D depth maps are native inputs; video is a sequence of images.
• Autoregressive: it produces one multimodal sequence element at a time, conditioned on prior elements.
• Diffusion: a rectified-flow process denoises visual outputs, with an inference-time quality and speed tradeoff.
• Transformer: the shared backbone operates on the multimodal sequence and spatial context.

The release post says Atlas can take one to six reference images and a manually designed camera path, then render new positions and angles. World Labs lists output up to one minute at 1440p.

The input recipe for the phone demo is unusually concrete. According to eerac's capture reply, it needs N frames captured at the same moment and a trajectory that starts and ends at one camera position, whose pose Atlas estimates.

The capture workflow still has a practical boundary: eerac wrote in a reply that freezing the action before moving the camera works better, while unfrozen shots work best with smaller, slower motion.

World Labs' reconstruction examples treat input count as a fidelity control. One image lets Atlas extrapolate unobserved geometry; with two or three images, the company says it typically produces faithful reconstruction; its spatial context can also use more than 100 inputs.

The output path can be scanned as three deliverables:

• Novel-view 2D images and videos.
• Depth predictions and point clouds.
• Complete 3D Gaussian splat scenes for high-resolution, on-device rendering.

The same post says unseen areas are filled with plausible predictions. The result is a reconstructed-and-imagined scene, not solely an observed capture.

Atlas uses the same model to reconstruct a space and generate the RGB and depth observations a robot-mounted camera would receive along a simulated route, according to World Labs' robotics section. Its examples reconstruct two environments from 24 frames each of cell-phone video.

For manipulation, the company says casual recordings can seed simulations of rigid, articulated, and deformable objects, then vary objects, pose, robot motion, lighting, and background to generate alternate scenes.

World Labs' camera-control evaluation uses third-party human raters to choose the output that better follows a target path from a single input image. The company reports Atlas was preferred in:

• 75% of pairwise trials against MiniMax H3.
• 81% against Gemini Omni Flash.
• 86% against Happy Horse 1.1.
• 93% against FLUX 3.
• 94% against Seedance 2.5.

The protocol gives Atlas the path in its native camera representation, while competing models receive the camera motion as cinematic text prompts. World Labs notes that more sophisticated prompting could improve those baselines, so the chart measures camera control through each model's available interface.

World Labs says in its availability note that Atlas is in early access with select partners, with a request form for prospective builders. The company also says Atlas will power future versions of Marble and other World Labs products.

World Labs calls image generation a secondary Atlas capability, but says it supports complex text prompts, text rendering, varied visual styles, and 360-degree panoramas in its launch post. Asked whether an Atlas black-hole clip was physically accurate, jcjohnss said in a reply that it probably was not.