Virtual scouting becomes interactive with the Frameboy demo

The Frameboy demo from World Labs streams Atlas views at up to fifteen frames per second. The Viewfinder creation tool remains in closed beta.

A single room is used to illustrate two ways of working with a generated set. In the comparison published by World Labs, one interface asks you to place waypoints, start generation, and then wait for the video before correcting the route. In the second, the camera moves and the views update as the exploration unfolds.

The Frameboy research preview gives access to this second experience. An optimized variant of the Atlas World Model generates images on GPUs in the cloud and streams them to the browser, at a reported rate that can reach fifteen frames per second. The user controls the camera while the model produces the next viewpoints. Framing is discovered during exploration

The shift first concerns when decisions are made. In a video generation prepared in advance, the movement must be defined before seeing the result. The interactive flow described by World Labs makes it possible to adjust while navigating: change angle, return to a viewpoint, or continue in a direction suggested by the images that appear.

For virtual location scouting for a shoot or previewing a set, the value lies in being able to find your frame through movement. The company also cites the exploration of game level concepts and the visualization of real environments. These uses are work avenues highlighted in the announcement, without any published assessment of the time saved by creators.

The operation of Atlas gives a precise role to the camera: its position and orientation are part of the model’s inputs. The point of view therefore does not depend solely on an instruction written as “rotate around the object.” The model uses these geometric indications to generate the corresponding image from the already established spatial context. A starting image, then views that build the world

The scenes presented do not all come from the same type of material. Some begin with a single image. Additional viewpoints are first generated with Atlas to define the rest of the space, then added to the context used by Frameboy. The interactive exploration relies on this set of references.

Other scenes use multiple views of a real place. The model then synthesizes the images corresponding to the requested camera positions. The parts missing from the references are completed by generation: the result combines the information visible in the provided images and what the model imagines to extend them.

The browser-accessible demo offers a selection of scenes prepared by the team. It lets users explore the result of this construction, rather than an open tool allowing anyone to immediately import their own images. The personal creation options are grouped in another application. Building your own scenes in Viewfinder

The Viewfinder application is used to compose an environment image by image with Atlas. The user brings their references, moves through the scene, and generates new views from different angles. Each produced image includes color and depth information, in order to visualize the space being built.

The Frameboy mode then acts as a streaming rendering overlay. A scene prepared in Viewfinder can be explored live, with new images generated during navigation. The two modes thus correspond to two moments in the work: establishing the environment’s references, then moving through the viewpoints they make possible.

Access to Viewfinder remains in closed beta, with a waitlist signup. This status concerns the personal scene composition tool; the Frameboy demo is offered separately to discover already prepared environments. Generative rendering that still depends on the cloud

The rate of fifteen frames per second corresponds to the maximum announced for the Atlas variant used in Frameboy. The post does not detail the optimizations applied or the hardware configurations needed to reach this throughput. It describes a model tuned to speed up the generation of new views, without publishing a quantified quality comparison with the initial version.

The browser receives images computed on a remote server. GPU requirements therefore remain on the side of the infrastructure hosting the session. The team specifies that this method remains costly compared with traditional local rendering.