On Sept. 8, 2026, researchers at the University of Illinois Urbana-Champaign and Cornell University submitted a paper describing FIRE3D. The system takes one color photo or an everyday video of an indoor scene, with no need to mark out the objects first. In under about a minute, it turns that into separate 3D objects that can be edited and dropped into a simulation. Games, robotics and other interactive apps need objects they can move and rearrange, not just a 3D picture that looks right.
Rather than slowly refitting a reconstruction for every new room, FIRE3D works in a single pass. It uses an existing tool, Pi3, to estimate depth and camera positions from the footage. It then counts the objects and predicts each one's outline, the box it sits in, its position, its 3D shape and its surface look. The researchers say it needs no outlines, boxes, positions or prompts from the user, and no extra fitting step for each new scene.
The researchers say a new compression step, which shrinks how the system stores each object, lets it build many objects at once on a single data-center chip. That makes generating them more than 5 times faster, they say, with only a small loss in quality. They report results that rival or beat earlier methods on placing objects, completing their shapes and getting their surfaces right, and say FIRE3D is orders of magnitude faster than earlier systems that produce interactive scenes. The project page puts FIRE3D at about one minute, against about eight hours for one such system, HoloScene.
The team trained FIRE3D on about 80,000 scenes, along with 140,000 video clips and 500,000 extra objects. Its code, trained models and example inputs are on GitHub and Hugging Face. FIRE3D's own code is under the MIT license, a license that lets anyone use, change and share it free of charge, while parts that come from two outside projects, DINOv3 and TRELLIS.2, keep their original licenses.
The researchers say FIRE3D works only on still indoor scenes for now. Its results depend on getting the depth, the camera positions and the split into separate objects right. Its 3D objects are not yet guaranteed to be physically stable, to look right under new lighting or to have parts that move. The model's page on Hugging Face warns that quality can drop with poor depth or camera estimates, heavily blocked views, unusual scale, backgrounds that are not rooms, or objects unlike those it was trained on. It says the shapes should be reviewed before any safety-critical or real-world physical use.
The researchers have not shown how often the under-a-minute speed holds outside their own indoor tests, and it is not yet known whether other teams can match their accuracy. Nor is it clear how useful the objects are before physics, lighting and moving parts are added.