Atlas world model goes real-time: Frameboy streams camera-driven frames to the browser at up to 15 fps

Interacting with the Atlas world model shifts from submitting a request and waiting for a video to watching the view update while moving the camera, letting users explore in the browser a 3D scene defined by a single image or a handful of viewpoints. Previously, users waited tens of seconds or even minutes for Atlas to produce a video, which forced cautious, methodical scene building and camera-path planning and limited iteration cycles. Frameboy streams frames from cloud GPUs to the browser at up to 15 fps, a rate served by a highly optimized Atlas variant built for the fastest, most efficient inference.

The real-time version keeps Atlas's generative novel view synthesis, generating scenes from scratch from as little as one input image or a text caption and also navigating real-world scenes captured from multiple viewpoints. The earlier lightweight frame generation model RTFM was primarily designed to interpolate between a preexisting set of viewpoints, whereas Atlas is much larger, trained on orders of magnitude more data, and can generate scenes from minimal context. Each demo scene was seeded from a single input image, with Atlas generating additional viewpoints that shape the rest of the scene before they enter Frameboy's spatial context; landmark scenes use multiview captures as input.

Frameboy currently must run on a powerful GPU server in the cloud and is not cheap compared with local rendering methods; Viewfinder's real-time overlay render mode is in closed beta with a waitlist.

The release treats speed and cost as current limits and expects them to fall as hardware and software evolve to meet new algorithms, without giving a timeline; the claim that the real-time variant maintains excellent generative novel view synthesis comes without quantitative metrics.