Recursive Code World Models: Building Complex Worlds through Recursive Scene Programs
RCWM turns a single reference image into a recursive 3D scene-building program.
The paper frames complex world reconstruction as code generation with repeated global-local-global passes. Each recursive call rebuilds unresolved parts, then returns to refine how they fit into the full scene. A vision-language coding agent compares renders against the reference image to guide edits and descent. The authors report stronger results than prior code-based image-to-scene methods, with ablations supporting deeper recursion for finer structures. HF Daily Papers' note
The paper frames complex world reconstruction as code generation with repeated global-local-global passes. Each recursive call rebuilds unresolved parts, then returns to refine how they fit into the full scene. A vision-language coding agent compares renders against the reference image to guide edits and descent. The authors report stronger results than prior code-based image-to-scene methods, with ablations supporting deeper recursion for finer structures. HF Daily Papers' note
score 5