World Labs just removed the velvet rope. As of November 12, Atlas — the spatial intelligence world model from Fei-Fei Li’s lab — is open to every developer, no partner agreement required. Four production API endpoints are live. Point them at a handful of smartphone photos and get back a navigable 3D world: Gaussian splats, polygon meshes, depth-aware reconstructions, video keyframes. The spatial internet finally has an API.
This Is Not a 3D Image Generator #
That is the most important thing to understand about Atlas. Tools like Stable Diffusion 3D and Luma Dream Machine generate images that look three-dimensional. Atlas generates actual 3D geometry that you can navigate, re-light, physics-simulate, and export into game engines. The difference is architectural.
Atlas is built on a multimodal autoregressive diffusion transformer trained simultaneously on text, images, video, and 3D datasets. Camera pose is a first-class input — not a prompt hack, not an LoRA add-on. The model understands where the camera is, what it is looking at, and what the geometry of the scene should be. That is what makes it a world model rather than a view synthesis trick. For comparison, Google’s Genie 4 world model generates interactive scenes from images but does not produce explicit 3D geometry or reconstruction from real photos.
The team behind Atlas is not a generalist crew riding the AI wave. Fei-Fei Li built ImageNet and ran Google Cloud AI. Co-founder Ben Mildenhall co-invented NeRF. Justin Johnson did the visual reasoning research that became CLEVR. Christoph Lassner spent years on 3D human digitization. These four built the academic foundations of spatial AI before World Labs existed.
Four Atlas API Endpoints, Each with a Real Use Case #
The Atlas API is live at api.atlas-beta.worldlabs.ai/api/v2. Four endpoints cover the main workflows:
- atlasGenerate — Single image in, navigable 3D world out. This is the generative path: upload one photo, get back a full spatial scene.
- atlasChisel — Layout-first generation. Sketch your geometry constraints, Atlas fills in the world. Useful when you need a specific spatial structure rather than a purely generated one.
- images2PosedRGBD — Reconstruction from up to 64 input images. Outputs positioned RGB+depth data with camera calibration baked in.
- splats2Mesh — Converts a 3D Gaussian splat into a polygon mesh (GLB format), making it editable in Blender, Houdini, or any standard 3D DCC tool.
Two web apps ship alongside the API. Viewfinder handles image-to-world reconstruction. Chisel handles the geometry-first workflow. Both are high-level wrappers over the same endpoints.
The Gaussian Splatting Pipeline Is Production-Ready #
This is where Atlas gets practical for teams building real products. The output chain maps directly to game engine requirements. Upload your photos, run images2PosedRGBD, export as SPZ or PLY splat (photorealistic rendering) plus a collider GLB (an invisible watertight mesh for physics) plus a polygon mesh (for DCC editing). Drop all three into Unity or Unreal. Done.
The collider mesh deserves a callout: it is what makes this workflow usable in production games rather than just impressive in demos. You can render the photorealistic Gaussian splat on top while the physics engine interacts with the invisible collider mesh underneath. No manual mesh cleanup required. Blender 5.3 added native Gaussian splatting support last week, so the DCC side of this pipeline just became significantly easier.
Real-to-Sim: The Spatial Intelligence Use Case That Matters Most #
The highest-value application is not games — it is robotics simulation. The workflow: record a physical environment with a smartphone, feed up to 64 frames to images2PosedRGBD, get back a simulation environment with synchronized RGB and depth. Train your robot in the simulation, deploy in the real space.
Real-to-Sim workflows have existed before Atlas, but they required expensive capture rigs, manual preprocessing, and significant engineering overhead. An API call changes that equation. This is what Fei-Fei Li has described as the core thesis behind World Labs: AI that understands space, not just text. It is now callable with four lines of Python.
It is worth noting that robot actions are not yet a native input — Atlas produces scene geometry, not physics simulations. For teams building full robot agent workflows, something like NVIDIA GR00T 2 handles the robot policy side. Atlas handles the environment reconstruction side. Used together, they close most of the Real-to-Sim gap.
AMD Is Buying This for $8.2 Billion — Here Is Why #
AMD announced the acquisition of World Labs on September 28 for $8.2 billion in an all-stock deal expected to close by end of 2026. Post-close, Li joins AMD as EVP and Chief Scientist under CEO Lisa Su. The strategic logic: AMD wants to build silicon optimized for spatial intelligence workloads, the same way Nvidia built CUDA as the substrate for language model training. If spatial AI becomes the next wave, AMD wants to own both the hardware and the foundational model. AMD’s official announcement frames it as designing chips around world models rather than retrofitting world models onto general-purpose chips.
That framing makes the Atlas developer API launch look less like a product release and more like ecosystem seeding. World Labs is building the developer community while AMD prepares the hardware platform.
Getting Access to the Atlas World Model API #
Atlas is available now at atlas.worldlabs.ai. No partner agreement, no waitlist. Developer documentation is at docs.worldlabs.ai. Pricing has not been publicly disclosed. A community ComfyUI plugin already exists with nodes for Create World, Render View, Expand World, and Splat to Mesh operations — a reasonable starting point if ComfyUI is already in your workflow. Install it through ComfyUI Manager by searching for “WorldLabs Atlas” in the Comfy Registry.