Georgia Tech's RCWM Turns Single Images Into Recursive Scene Code

Georgia Tech researchers Zhiqi Li, Yuxuan Liao, and Bo Zhu introduce Recursive Code World Models (RCWM), a framework that reconstructs 3D worlds as executable code from a single reference image, in a 21-page paper posted to arXiv on September 10.

The method pairs a Recursive Scene Program with a solver that follows the same loop at every depth: "establish the whole, recursively reconstruct unresolved parts, and revisit the whole to refine their composition." A vision-language coding agent compares rendered scenes against the reference image to guide edits, and the solver revisits parent nodes when local refinements would otherwise break spatial relations.

On five whole-scene references and five local crops, RCWM reports PSNR scores of 18.7 on a "medieval-village" scene and 19.3 on "japan-island," with LPIPS as low as 0.075 on "park-lake." Ablations find that "deeper recursive solves can improve both local detail and whole-scene fidelity." The abstract concedes that representing worlds as programs "does not determine how to construct a complex world"; recursion is the paper's proposed answer to that gap.

The comparison in the paper is against prior code-based image-to-scene methods; it does not benchmark against neural reconstruction baselines. It arrives in a busy week for world-model work on our tracker, alongside Westlake AGI Lab's 'World in World'.