Atlas – The World's First Multimodal World Model from World Labs

1. What is Atlas

Atlas is the world's first multimodal world model introduced by World Labs, founded by Fei-Fei Li. This model natively understands text, images, videos, and 3D spatial information. By anchoring visual content in a 3D coordinate system through a spatial context mechanism, it enables pixel-level precise camera control generation, sparse photo-based 3D reconstruction, and spatiotemporal simulation. Users only need to provide a few ordinary photos to generate long videos with arbitrary camera movements, reconstruct real-world scenes, and build simulated training environments for robots.

Image source: Official article
Image source: official article

Technical positioning and domain: Atlas belongs to the field of multimodal world models (World Models), situated at the intersection of generative AI and 3D vision. Unlike conventional video generation models, Atlas treats spatial understanding as a first principle, with 3D scene representation as its core output goal. It covers four major tasks: image generation, video generation, 3D reconstruction, and physical simulation, establishing a differentiated position in the field of spatial intelligence.

Development background: This model was developed by World Labs, founded by Professor Fei-Fei Li of Stanford University. The team consists of experienced researchers in the fields of computer vision, graphics, and machine learning. The motivation behind its development was to overcome the bottleneck of existing generative models lacking spatial consistency, pushing AI from "generating pixels" toward "understanding space," and providing a unified spatial computing foundation for content creation and embodied intelligence.

Core value: Traditional video generation models rely on text descriptions to indirectly control the visual output, making it difficult to ensure geometric consistency across different perspectives. Atlas treats camera pose as a native input, addressing issues such as perspective drift and structural collapse at the architectural level. At the same time, it reduces the threshold for sparse reconstruction from professional equipment to ordinary smartphone photos, significantly lowering the cost of 3D content production and embodied intelligence training.

Technical features: The model employs a unified architecture based on a multmodal autoregressive diffusion Transformer, pre-trained from scratch. It encodes text, images, video frames, camera poses, and 3D depth maps into a single spatial context sequence, combining the sequence flexibility of language models with the high-quality rendering capabilities of diffusion models. When generating new views, the model extrapolates and completes within the established spatial framework, ensuring geometric consistency and physical plausibility across different perspectives.

2. Key Features

• Pixel-level Camera Control Generation: Input 1–6 photos and specify camera poses to generate a video with smooth camera movement up to 1 minute long and with a resolution of up to 1440p. Camera parameters are input as native data rather than textual descriptions, ensuring high geometric consistency in the image space. This is ideal for applications such as film pre-visualization and advertising production that require precise camera motion.
• Sparse Photo 3D Space Reconstruction: Reconstruct a 3D point cloud or Gaussian splatting model of a real scene using only 2–25 ordinary smartphone photos, without the need for professional scanning equipment. This capability significantly lowers the barrier for 3D scene acquisition, providing a low-cost solution for game development, architectural visualization, and digital archiving.
• Temporal-Spatial Simulation and Multi-Camera Reconstruction: The model understands spatial structure and the rules of world evolution, supporting multi-camera reconstruction in a "bullet time" style. Users can shoot the same video segment from different angles using 3–5 smartphones and then freely replay it from any angle, replacing traditional professional camera array solutions.
• Real-to-Sim Robot Simulation Environment Generation: After capturing short videos of real environments, Atlas automatically generates a high-precision digital twin scene, allowing robots to repeatedly test and learn in a virtual space. This capability greatly reduces the cost of real-world data collection, providing scalable simulation data for embodied intelligence training.
• High-Quality Image and 360° Panorama Generation: Supports generating high-quality images and 360° panoramas based on text or image prompts, capable of handling complex prompts and various visual styles. This is suitable for applications such as e-commerce product displays and VR content creation.
• Unified Multimodal Spatial Context Architecture: The model encodes text, images, video frames, camera poses, and depth maps as a unified spatial context sequence internally, performing extrapolation and completion within an established spatial framework. This ensures geometric consistency and physical plausibility across different perspectives.

3. How to Use

1. Apply for Access: Visit the Atlas release page on the World Labs official website and submit an application via the official form. Atlas is currently open on an application basis, and once approved, you will gain access to the usage portal.
2. Camera-Controlled Video Generation: Upload 1–6 reference photos, set the camera poses and movement trajectories, and Atlas will generate a video with a maximum length of 1 minute and a resolution of 1440p. It is recommended to cover multiple perspectives of the scene with the reference photos to improve spatial consistency in the generated video.
3. 3D Spatial Reconstruction: Upload 2–25 regular photos, and the system will automatically reconstruct them into a navigable 3D point cloud or Gaussian splatting scene. When taking photos, ensure sufficient overlap between adjacent images and avoid overexposure or motion blur to improve reconstruction accuracy.
4. "Bullet Time" Effect Production: Use 3–5 regular smartphones to synchronously capture the same video segment from different angles. After uploading, Atlas will reconstruct it into a spatiotemporal simulation result that can be replayed from any angle, without the need for a professional camera array.
5. Robot Simulation Training: Capture short videos of real environments, and Atlas will automatically generate high-precision digital twin scenes for robots to perform simulation training and strategy validation in a virtual space.
6. Text-to-Image and Panorama Generation: Input text prompts or reference images to directly generate high-quality images or 360° panoramas, supporting complex prompts and multiple visual styles.
7. Notes: Atlas has not yet made its local deployment plan or API documentation public. Specific computational requirements and the schedule for open-sourcing model weights are pending official announcements. When applying for access, it is recommended to use an institutional email address and clearly describe your intended use case to increase the likelihood of approval.

4. Pros and Cons Analysis

5. Comparative Analysis with Similar Tools

Selection Recommendations: For film and visual effects production teams, Atlas's pixel-level camera control and sparse reconstruction capabilities make it the most suitable choice currently, especially for pre-visualization stages requiring precise camera movement and scene reconstruction. If the focus is on real-time interactive experiences, such as environment generation for games, Genie 3's real-time interactive capabilities offer a stronger advantage.

For embodied AI and robotics research institutions, NVIDIA Cosmos provides open-source models and a simulation toolchain tailored for physical AI, with a relatively mature ecosystem. Atlas's Real-to-Sim capabilities perform exceptionally well under sparse input conditions, but the maturity of its toolchain remains to be observed. Content creators who only require high-quality video generation will still find Sora a reliable choice in terms of visual quality and text understanding, though they must accept its limitations in spatial control.

6. Editor's Summary

The release of Atlas marks a significant shift in generative AI from pixel generation to spatial understanding. From a technical architecture perspective, the multimodal autoregressive diffusion Transformer unifies the sequence modeling capabilities of language models with the continuous data generation capabilities of diffusion models within a single framework. The introduction of spatial context addresses cross-perspective consistency issues at a mechanistic level, making this design approach valuable for future world model research.

The model is trained from scratch and has been verified to continuously improve performance with increased computational power, leaving room for scalability.

In terms of practical value, Atlas reduces the threshold for sparse reconstruction to ordinary smartphone photos, directly lowering the cost of 3D content production and embodied intelligence training. It has clear application potential in scenarios such as film previsualization, game development, and architectural presentations. Its Real-to-Sim capability provides an automated pathway from the real world to simulated environments for robot training, helping to alleviate the bottleneck of scarce real-world data in the field of embodied intelligence.

Currently, Atlas is in a limited-access open phase, with model weights, API, and local deployment solutions not yet released. The ecosystem toolchain still needs to be refined. For professional teams engaged in spatial intelligence research, 3D content production, and robot training, Atlas is definitely worth close attention. Ordinary creators may wait for the official public beta before evaluating its compatibility with their workflow. As computational power expands and the ecosystem develops, the future potential of Atlas in the field of spatial intelligence is worth continuous observation.

7. Application Scenarios

• Film and VFX Production: Directors provide a few scene photos and design camera movement paths, and Atlas generates precise camera movement videos with a resolution of up to 1440p; using 3–5 smartphones can achieve the "bullet time" effect, replacing traditional professional camera arrays and significantly reducing the cost of special effects production.
• Game and Virtual Scene Development: Art teams upload a small number of real photos, and Atlas automatically reconstructs them into navigable 3D Gaussian splatting scenes, which can be directly imported into game engines, shortening the production cycle for open-world or VR environments.
• Robotics and Embodied AI Training: By filming short videos of real environments with a smartphone, Atlas automatically generates high-precision digital twin scenes, allowing robots to repeatedly perform simulation training and strategy validation in a virtual space, addressing the pain points of slow and costly real-world data collection.
• Architecture and Real Estate Display: With just a few on-site photos, Atlas can reconstruct complete 3D models of buildings or interior spaces, supporting navigation from any angle and the generation of aerial views, which can be used for property pre-sales presentations or digital archiving of historical sites.
• Advertising and E-commerce Content Generation: Input product images or text prompts to quickly generate 360° panoramic display videos or stylized advertisements, ensuring geometric and lighting consistency of the product across different camera movements through pixel-level camera control.

8. FAQ

Q: What is the fundamental difference between Atlas and video generation models like Sora?
A: Models like Sora generate videos based on text or images as conditions, but they lack explicit understanding of three-dimensional spatial structures, making it difficult to ensure consistency across different perspectives. Atlas, on the other hand, takes camera poses and depth information as native inputs, constructing a 3D scene representation with spatial constraints within the model. This mechanism ensures geometric consistency in the generated visuals and also supports 3D reconstruction and simulation environment generation.

Q: What shooting conditions are required for 3D reconstruction?
A: Only 2–25 regular smartphone photos are needed, without the requirement for professional scanning equipment. It is recommended to ensure sufficient overlap between adjacent photos, avoid overexposure and motion blur, and cover as many perspectives of the scene as possible to improve the accuracy and completeness of the reconstructed point cloud or Gaussian splatting model.

Q: How is the precision of camera control achieved?
A: Atlas encodes the camera pose as part of the spatial context sequence, unifying it with images, video frames, and depth maps in the modeling process. When generating new views, the model extrapolates and completes within the established 3D coordinate framework, enabling pixel-level camera control rather than relying on vague textual descriptions.

Q: Does Atlas support Chinese input?
A: The official release information does not explicitly mention support for Chinese. Considering that the model uses a unified multimodal encoding architecture, its text input capabilities depend on the language coverage of the pre-training data. Specific performance with Chinese input will require official disclosure or practical testing for confirmation.

Q: How can one obtain access to Atlas?
A: Currently, Atlas is open on an application basis. Users must submit an application via the form provided on the official release page. It is recommended to use an institutional email and clearly describe the application scenario. Upon approval, access will be granted. The model weights, API, and local deployment solutions have not yet been released.

Q: Can Atlas be used for real-time interactive generation?
A: Based on the current release information, Atlas primarily operates as an offline task and has not demonstrated real-time interactive generation capabilities. If real-time interactive world models are required, one might consider solutions like Google DeepMind's Genie 3, which focus on real-time responsiveness.

9. Project Links

1. Official Release Page: https://www.worldlabs.ai/blog/atlas — Introduction and technical documentation of the Atlas model published on the World Labs official blog.
2. Access Application: https://form.typeform.com/to/zHFR4r3A — Official Atlas access application form provided by the company.