Unravel, described in a new arXiv paper, is a training-free framework for recovering an interactive 3D scene from a single image. The central claim is that every object in the scene becomes an independent and movable asset, rather than being locked into a single fixed structure.

The authors position this against existing generators, which they say tend to produce one entangled geometry in which objects cannot be separated or manipulated. The abstract is cut off mid-sentence, so the source does not include implementation details or evaluation results.