World Labs' Frameboy: Real-Time 3D World Generation from Images at 15 FPS

World Labs, founded by Fei-Fei Li, has released Frameboy , an optimized demo of its Atlas world model. Frameboy uses cloud GPU streaming to generate first-person views in real time, achieving up to 15 fps with a response time of approximately 70 ms , bringing the experience close to cloud gaming.

Frameboy Demo: Interactive 3D Exploration

Frameboy currently offers curated scenes for users to explore, including real-world landscapes, cyberpunk spaces, and the Stanford campus. These scenes are constructed in two ways: from a single image where Atlas infers other viewpoints to gradually build an explorable environment, or from multiple photos of the same location where the model generates views between and beyond the provided photos.

The experience link is available at https://frameboy.worldlabs.ai/.

Integration with Viewfinder

World Labs has also integrated Frameboy directly into the beta web application Viewfinder as a streaming overlay render mode. This allows users to build scenes frame by frame from their own input content and then explore and expand the virtual world in real time. Like using a camera, users can navigate from different perspectives to generate 3D-consistent world views. Each frame produced by Atlas contains both color and depth information. Once ready, users can trigger the Frameboy overlay to play back the generated frames live during exploration.

Atlas: From a Single Photo to a 3D World

To understand Frameboy's significance, we must look at the underlying Atlas model. Released on September 1, Atlas is a general world model trained from scratch that processes text, images, video, and 3D spatial information. Unlike video models that primarily generate continuous frames, Atlas tackles a more complex question: if the camera moves to a new position, what should the world look like?

Spatial Context

World Labs demonstrated this with an intuitive example. Researchers gave Atlas a photo of a garden and asked it to generate an aerial view. Since the original photo covered only part of the garden, the model reproduced the visible objects but had to hallucinate the surrounding houses and environment. Adding a second photo of a nearby cottage allowed Atlas to produce a view containing both garden and cottage, though distant buildings were still inferred. A third photo of the main house enabled the model to reconstruct the entire scene more faithfully.

This illustrates Atlas's core mechanism, termed "Spatial Context" . Atlas does not merely recognize objects in an image; it associates each image with the camera's 3D position and orientation. When a user specifies a new viewpoint, the model generates the corresponding frame based on the accumulated spatial information. This differs from typical video generation models, which rely on textual descriptions like "move camera left" or "rotate around subject" to control camera motion. Atlas accepts precise camera poses, letting users explicitly define the viewing angle.

Technical Architecture

Architecturally, Atlas employs a multimodal autoregressive diffusion Transformer . It combines ideas from language models and diffusion generative models: text, images, camera poses, and depth are organized into a unified spatial context, which is then used to progressively generate new frames and spatial data.

Capabilities Beyond Video

In its initial demo, Atlas allowed users to pre-design camera paths to generate videos up to one minute long at 1440p resolution , maintaining consistency across shots. However, Atlas can also simultaneously generate images and depth information , and output point clouds, 3D Gaussian Splatting representations, and other 3D formats. This means the model not only renders a view from a given angle but also produces data usable in 3D rendering and spatial reconstruction pipelines.

For example, researchers used three to five ordinary phones and action cameras to capture a scene from different angles. Atlas reconstructed the space, enabling users to switch viewing angles at the exact same moment in time — achieving a "bullet time" effect similar to The Matrix — without deploying traditional large-scale multi-camera rigs.

Robotics Simulation

Another application direction is robotics simulation . Researchers can record real environments with phone video, reconstruct the space with Atlas, and generate the camera views and depth data that a robot's onboard camera would observe along different trajectories. This provides a new way to build robot simulation environments from sparse real-world recordings.

Impact on Creative Workflows

Frameboy changes how creators work with world models: they can now move and observe, decide the next step, and iterate in real time. This interactive loop holds potential for film location scouting, spatial concept design, and game level ideation .

Reference:

https://www.worldlabs.ai/blog/atlas-realtime