project
Atlas - World Labs launches the world's first multimodal world model
Atlas, developed by World Labs founded by Fei-Fei Li, is the world's first multimodal world model. The model natively understands text, images, videos, and 3D spatial information, anchoring the scene in three-dimensional coordinates through spatial context...
What is Atlas?
Atlas, launched by World Labs, founded by Fei-Fei Li, is the world's first multimodal world model. The model natively understands text, images, videos, and 3D spatial information, anchoring images in three-dimensional coordinates through spatial context. This enables pixel-level precise camera control generation, sparse photo 3D reconstruction, and spatiotemporal simulation. With just a few ordinary photos, the model can generate long videos with arbitrary camera movements and reconstruct realistic scenes, providing a simulation training environment for robots.
Atlas's main functions
- Camera-controlled generationInput 1–6 photos and specify the camera pose to generate a 1-minute, 1440p video with consistent spatial geometry.
- Space Reconstruction: 3D point cloud or Gaussian splash model of a real scene can be reconstructed using only 2–25 ordinary photos, without the need for professional scanning equipment.
- Spacetime SimulationIt enables understanding of spatial structure and world evolution, supports "bullet time" multi-camera reconstruction, and provides a Real-to-Sim simulation training environment for robots.
- Image generationGenerate high-quality images and 360° panoramas based on text or image prompts, supporting complex prompts and multiple visual styles.
Atlas's technical principles
- Multimodal Autoregressive Diffusion TransformerAtlas pre-trains a unified architecture from scratch, encoding text, images, video frames, camera poses, and 3D depth maps into the same "spatial context" sequence, and then generating output element by element in an autoregressive manner; at the same time, it uses a diffusion mechanism to progressively denoise high-dimensional continuous data (images/videos), balancing the sequence flexibility of LLM with the high-quality rendering capabilities of video generation models.
- Spatial ContextUnlike LLM's textual context, Atlas binds explicit 3D coordinates and depth information to each input image, constructing a spatially constrained 3D scene representation within the model. When generating new views, the model extrapolates and completes within this established spatial framework, ensuring geometric consistency and physical plausibility across viewpoints.
How to use Atlas
- Visit the official websiteRequest Atlas access: https://form.typeform.com/to/zHFR4r3A.
-
Camera control generates videoUpload 1–6 reference images, set the camera pose and motion trajectory, and Atlas will generate a cinematic video of up to 1 minute in length.
-
3D spatial reconstructionUpload 2–25 ordinary photos; no professional equipment is required. Atlas will automatically reconstruct them into a roamable 3D point cloud or Gaussian splash scene.
-
"Bullet Time" effectUse 3–5 ordinary mobile phones to shoot the same video from different angles. After reconstruction by Atlas, it can be played back freely from any angle.
-
Robot simulation trainingShoot short videos of real-world environments, and Atlas automatically generates high-precision digital twin scenes for robots to run and test.
-
Wensheng Photo/Panoramic ViewEnter text prompts or reference images to directly generate high-quality images or 360° panoramas.
Atlas's core advantages
-
Pixel-level camera controlIt uses camera pose as the native input, rather than relying on vague text descriptions, to achieve director-level precision in camera movement.
-
Extreme sparse reconstructionHigh-precision 3D scenes can be reconstructed using only a few ordinary mobile phone photos, far exceeding traditional scanning solutions.
-
Spatial geometric consistencyBased on Spatial Context, images are anchored in 3D coordinates, generating images across different viewpoints without distortion or flickering.
-
Multi-task unified architectureA single model can simultaneously cover generation, reconstruction, simulation, and image generation without the need to switch to dedicated tools.
-
Real-to-Sim Closed LoopIt can directly convert real-world videos into robot simulation environments, significantly reducing the cost of embodied intelligence training.
-
Continuous scalingIt was pre-trained from scratch and its performance has been proven to continuously improve with increasing computing power, indicating extremely high potential for future growth.
Atlas project address
- Project official website：https://www.worldlabs.ai/blog/atlas
Comparison of Atlas with similar competing products
| Comparison Dimensions | Atlas（World Labs） | Genie 3（Google DeepMind） |
|---|---|---|
| position | Omni World Model for Spatial Intelligence | General Real-Time Interactive World Model |
| Core Input | Text, images, videos, camera poses, 3D depth maps | Images, action commands, and state history |
| Camera control | Pixel-level precision, natively supports camera pose parameter input | Control is primarily achieved indirectly through text/action descriptions. |
| 3D Reconstruction | Point clouds can be reconstructed from 2–25 ordinary photos/Gaussian splash. | Focus on interactive world generation, non-sparse reconstruction expertise |
| Video generation | Longest runtime 1 minute, 1440p, with extremely strong spatial geometric consistency. | Emphasis is placed on real-time interaction, with relatively short durations per segment. |
| Physical simulation | Support for generating robot training environments via Real-to-Sim | Emphasizing a self-learning physics engine and real-time interactive feedback |
| Output format | Images, videos, point clouds, 3D Gaussian Splatting | Interactive world state, video frame |
Application scenarios of Atlas
- Film and VFX ProductionThe director provides a few scene photos and designs the camera movement trajectory. Atlas can generate a high-resolution 1440p film with precise camera movement, and can achieve "bullet time" effects with just a few mobile phones, saving the need for hundreds of professional camera arrays.
- Game and Virtual Scene ConstructionThe art team uploads a small number of real-world photos, which Atlas automatically reconstructs into roamable 3D Gaussian splash scenes, directly importing them into the game engine and significantly shortening the production cycle of open-world or VR scenes.
- Robotics and Embodied Intelligence TrainingBy shooting short videos of real-world environments with a mobile phone, Atlas automatically generates high-precision digital twin scenes, allowing robots to repeatedly run through the virtual space and try out different scenarios, thus solving the core pain points of slow and costly real-world data collection.
- Architecture and Real Estate ShowcaseIt can reconstruct a complete 3D model of a building or interior space from just a few on-site photos, supporting roaming at any angle and generating aerial perspectives, which can be used for pre-sale display of real estate projects or digital archiving of historical sites.
- Advertising and e-commerce content generationInput a product image or text prompt, and quickly output a 360° panoramic display video or a stylized advertisement. Pixel-level camera control ensures that the product's geometry and lighting remain consistent under different camera movements.