World Labs Streams Atlas-Generated 3D Worlds to Browsers at 15 fps
World Labs just put Atlas, their autoregressive world model, into a browser demo that streams generated frames from cloud GPUs at up to 15 fps.
- World Labs released Frameboy, a browser demo streaming Atlas world model frames at up to 15 fps.
- Runs a speed-optimized variant of Atlas on cloud GPUs, preserving novel view synthesis quality.
- Scenes can be seeded from a single image or multiview captures of real locations.
- Integrated into Viewfinder web app for frame-by-frame scene building, available via waitlist.
- Builds on last year's RTFM prototype but scales to much larger model and dataset.
- Full technical writeup details the latency-cost argument behind the release.
World Labs has opened Frameboy, a browser demo that lets users move a virtual camera through scenes generated by a speed-optimized version of Atlas. Cloud GPUs run the model and stream frames to the browser at up to 15 frames per second.
Atlas is an autoregressive world model, which means it generates each new frame using the scene context and preceding output. The model supports precise camera-pose controls and novel-view synthesis, allowing it to depict a scene from angles absent from the original input.
Atlas enters the camera loop
A conventional Atlas workflow requires users to define key camera positions, start generation, and wait for a video render. Frameboy places generation inside the navigation loop: as the camera moves, Atlas produces the next view and the server sends it back continuously.
World Labs describes latency and inference cost as the key constraints on this interaction model. Once generation becomes fast enough, users can test camera paths continuously instead of committing to each path before rendering. Failed directions consume seconds rather than an entire generation job.
| Component | Role | Access |
|---|---|---|
| Atlas | Generates scene views from images, captions, or multiview context | Underlying world model |
| Frameboy | Streams interactive views from a speed-optimized Atlas variant | Public demo with curated scenes |
| Viewfinder | Builds explorable scenes from user-provided inputs | Closed beta |
Two routes into a generated scene
The optimized Atlas variant preserves novel-view synthesis while reducing inference time. World Labs demonstrates two input workflows in its technical post:
- Single-image seeding: Atlas starts with one image, generates supporting viewpoints, and adds them to Frameboy’s spatial context. The user can then move beyond the original camera position.
- Multiview capture: Atlas receives photographs of a real location from several angles and reconstructs an explorable representation, including demonstrations based on well-known landmarks.
Compared with the earlier RTFM prototype, Atlas uses a larger model trained on orders of magnitude more data, according to World Labs. RTFM primarily interpolated between existing viewpoints. Atlas can generate a scene with sparse context, including a single image or text caption.
Viewfinder adds an authoring workflow
Frameboy’s public demo limits navigation to scenes selected by World Labs. Viewfinder extends the same system into a creation tool: users provide source material, move a virtual camera through the emerging scene, and generate new frames at chosen viewpoints.
Each Viewfinder generation includes color and depth output, giving users a visual representation of the scene’s spatial structure. After building enough context, they can enable the Frameboy overlay and navigate through the result as a continuous stream. Viewfinder remains in closed beta through a waitlist.
Cloud costs and model drift remain
Current performance depends on powerful cloud GPUs, which makes Frameboy more expensive than local rasterized rendering. Its peak rate of 15 fps also falls below the frame rates expected from conventional interactive graphics systems.
Autoregressive generation introduces another constraint. Because each frame depends partly on earlier output, visual or geometric errors can accumulate across long camera paths. The scene may gradually drift from its initial structure, particularly when the camera explores areas weakly supported by the source images.
The public release also separates exploration from custom creation. Anyone can navigate the curated Frameboy scenes, while user-supplied inputs and scene authoring remain gated behind the Viewfinder beta.
What lower latency unlocks
Traditional renderers draw explicit geometry, materials, and lighting assembled by developers. Atlas synthesizes the next view from learned scene context, which can fill in unseen regions without a complete hand-authored 3D environment. Production engines retain advantages in consistency, determinism, physics, local execution, and frame rate.
World Labs identifies potential uses in location exploration, virtual film scouting, construction-site inspection, game-level development, and robot simulation. Each application would require stronger temporal consistency, predictable geometry, and lower inference costs than the preview currently demonstrates.
Frameboy tests whether world models can operate within an interactive rendering loop rather than producing isolated clips. Continued reductions in latency and compute cost could make that approach useful as a cloud rendering service, provided Atlas can maintain scene consistency across longer sessions and less curated inputs.
Comments
