World Labs Releases Frameboy for Real-Time Atlas Exploration

World Labs released Frameboy on November 5, 2026, a research preview that the company says streams Atlas-generated scenes from cloud GPUs to a browser at up to 15 frames per second as users control the camera. The company also integrated the mode into Viewfinder, its closed-beta web app for building 3D scenes.

From Render Requests to Continuous Exploration

World Labs says the optimized Atlas variant behind Frameboy is designed for faster, more efficient inference while retaining novel-view synthesis quality. The announcement’s main distinction is a new interaction loop: instead of requesting a video and waiting for the result, a user receives frames continuously and can change the camera path during exploration.

When a manually triggered Atlas video takes tens of seconds or several minutes, World Labs says users become cautious about building scenes and planning camera paths, with fewer iteration cycles. Continuous output lets them react to an unwanted view and adjust immediately. That folds experimentation into navigation instead of isolating it in a sequence of separate generation jobs.

The potential applications listed for Atlas include scouting virtual viewpoints for film shoots, inspecting progress on construction sites, developing 3D game levels and simulating environments for robots. Those tasks involve choosing or testing viewpoints, which helps explain why World Labs is treating generation latency as a product-level issue rather than only an inference metric.

Viewfinder Builds the Scene Before the Stream

The Frameboy demo offers a curated menu spanning Stanford Quad, Sutro Baths, Palace of Fine Arts, Cyberpunk, Spaceship Interior and Frozen Fortress. World Labs says scenes can be seeded by a single image, with Atlas generating additional viewpoints that populate Frameboy’s spatial context; multiview captures can instead supply real-world locations for navigation.

Viewfinder moves beyond that curated list. Users build a scene frame by frame from their own inputs, generate views from different camera angles and receive both color and depth for each new Atlas frame. A scene built in Viewfinder can then become the input for a real-time Frameboy session, whose overlay streams generated frames during exploration.

Viewfinder is available to test in closed beta through a waitlist; Atlas was also entering early access with select partners when it was introduced. The new streaming workflow sits within those limited access programs.

Atlas Architecture and the RTFM Lineage

World Labs introduced Atlas on September 1, 2026, as an omni world model pretrained from scratch to operate natively on text, images, video and 3D. Its multimodal autoregressive diffusion transformer grounds inputs in 3D space to form a spatial context, then generates sequence elements one at a time conditioned on earlier elements. Text, images, camera poses and 3D depth maps are supported inputs; videos are represented as image sequences, and each image or depth map is conditioned on an explicit camera pose.

World Labs says Atlas can generate up to one minute of 1440p video with camera control and reconstruct real-world scenes from sparse input images. It can also produce point clouds and 3D Gaussian splats, giving the model explicit 3D outputs alongside generated images and video.

Frameboy follows RTFM, a research preview released on October 16, 2025, that targeted interactive frame rates on a single H100 GPU. RTFM was designed as a learned renderer, generating new 2D views without constructing an explicit 3D representation. It used posed frames as spatial memory and retrieved nearby frames for each new view through context juggling, avoiding attention over an ever-growing frame history.

World Labs says Atlas scales up RTFM’s concepts with a much larger model trained on orders of magnitude more data. RTFM was primarily designed to interpolate among existing viewpoints; Atlas can generate scenes from scratch using only one image or a text caption, while retaining novel-view synthesis from collections of images.

Cloud Compute Remains the Constraint

World Labs says Frameboy currently requires a powerful cloud GPU server and is not cheap compared with local rendering methods.

The company expects evolving hardware and software to make the technology faster and cheaper. The preview’s central trade-off is therefore clear: continuous, camera-controlled generation reaches the browser, while the cloud inference required to sustain it remains the practical barrier to routine use.