Recursive Code World Models: Building Complex Worlds through Recursive Scene Programs

Georgia Institute of Technology

The same recursive solver reconstructs the whole scene, a village, a house cluster, and a house; the resulting scene program renders from several cameras. — One recursive solver reconstructs the whole, its parts, and an executable 3D world.

Abstract

Code world models represent worlds as executable programs, but this representation alone does not determine how to construct a complex world. We introduce Recursive Code World Models (RCWM), a framework for reconstructing complex 3D worlds in code from a single reference image. RCWM couples a Recursive Scene Program (RSP) representation with a construction solver that recursively calls itself. An RSP represents the executable world as compositional scene code, while each solver call follows the same complete process: establish the whole, recursively reconstruct unresolved parts, and revisit the whole to refine their composition.

Read the full abstract

This global–local–global recursion gives fine-scale structures their own perception-and-editing loops while preserving scene-wide geometry and relationships. Reference-aligned views propagate a shared camera projection across levels, while parent revisitation addresses boundaries, spatial relations, and shared errors that emerge after local refinement. A vision-language coding agent directly compares reference images with scene renders to guide refinement, recursive descent, and return. Across complex scenes, RCWM outperforms prior code-based image-to-scene reconstruction methods. Ablation studies further support the benefits of recursive construction and suggest that deeper calls can improve finer-scale reconstruction. RCWM provides a recursive construction principle for building complex executable worlds from visual evidence.

Read the method and the solver instruction →

The world is a program, built recursively

Pick a part to see what it built itself and what its children returned. Drag to look around.

Geometry

Everything delivered by scene, including its descendants.

Whole scenescene

Building the delivered scene…

Drag to orbit · Scroll or pinch to zoom

Open the root program ↗

Select a node to frame its program. ↑ parent · ↓ child · ← → siblings. On phones, scroll sideways for more.

Whole scene

Depth 0

Reference target received by scene — Target crop

Delivered render of scene — Delivered render

Each part receives a closer view of the reference and returns a scene program.

The brief

Full brief ↗

Explore the full run →

Results at a glance

Compare the reference with our reconstruction, then inspect the scene’s details.

The delivered program's rendered medieval village, with mountain, forests, rivers and buildings — From a single image to an executable 3D scene. Drag to compare · same camera

Input reference: an illustrated medieval village from WorldClaw Figure 9 — From a single image to an executable 3D scene. Drag to compare · same camera

City reconstruction: the whole scene and close-up comparisons with SEIG, VIGA, and img2threejs.

See all comparisons →

BibTeX

@article{li2026rcwm,
  title={Recursive Code World Models: Building Complex Worlds through Recursive Scene Programs},
  author={Li, Zhiqi and Liao, Yuxuan and Zhu, Bo},
  journal={arXiv preprint arXiv:2609.11499},
  year={2026},
  eprint={2609.11499},
  archivePrefix={arXiv}
}