World Labs, the spatial AI startup founded by researcher Li Feifei, released Frameboy this week, a research demo that brings the Atlas world model into real-time interaction.
The system streams frames generated by cloud GPUs directly to the browser at up to 15 frames per second while the user freely controls the camera — as if exploring a game. The interface recalls a retro handheld, reinforcing the idea of “playing” inside an AI-generated world.
What Atlas and Frameboy do
Atlas is an autoregressive world model introduced in September 2026. It generates frames sequentially with precise camera control and can create or reconstruct consistent 3D scenes from a single image or a handful of views. Frameboy uses a highly optimized version of the model focused on inference speed and efficiency while preserving novel view synthesis quality.
The practical difference is significant: instead of clicking “generate” and waiting tens of seconds or minutes for a video, the user receives frames continuously and can change the camera path at any moment. When the cost of a “bad idea” drops near zero, exploration becomes more natural and iterative.
Why it matters
World models are seen as an important step toward AI with spatial understanding — useful for games, content production, robot simulation, architectural visualization, and more. Frameboy shows it is possible to move beyond the slow “generate and wait” cycle toward something closer to real-time navigation, even if it still relies on expensive cloud infrastructure for now.
The company also integrated the streaming capability into Viewfinder, a web app (currently in closed beta) that lets users build scenes frame by frame and then explore them live. The official announcement notes that what is slow and expensive today tends to become faster and more accessible as hardware and software evolve.
Primary source: official World Labs blog (Frameboy: Atlas in Real-time).
Image credit: World Labs — Source: official site
By Grok
