Recursive Code World Models

Results

Comparisons, novel views, and the recursive construction behind them.

The selected example

Reference
Our reconstruction

Explore this run →

From other cameras

Reconstructing the whole and the details

Reference images and delivered scenes, with close-ups of their parts. Open any figure to inspect it at full size.

Medieval village. Full scene and local details across reconstruction methods.
Medieval village. Full scene and local details across reconstruction methods.
City. Full scene and local details across reconstruction methods.
City. Full scene and local details across reconstruction methods.

Beyond the reference view

The delivered scene programs can be rendered from new cameras.

Novel views of the reconstructed worlds across the paper’s scenes.
Novel views of the reconstructed worlds across the paper’s scenes.

Try the live world →

Quantitative comparisons

Scroll sideways to see all metrics →

Reconstruction quality across ten scenes. ↑ higher is better; ↓ lower is better. Bold values follow the paper.
Scene Method PSNR↑ SSIM↑ Edge F1↑ LPIPS↓ CLIP↑
city-full Ours 18.2 0.66 0.94 0.158 0.95
SEIG 14.5 0.51 0.83 0.339 0.90
VIGA 9.5 0.48 0.58 0.670 0.83
img2threejs 15.5 0.55 0.91 0.243 0.93
snow-village Ours 17.9 0.73 0.85 0.237 0.83
SEIG 13.4 0.64 0.37 0.520 0.77
VIGA 13.5 0.61 0.40 0.541 0.75
img2threejs 15.4 0.64 0.65 0.351 0.81
island-harbor Ours 17.4 0.74 0.90 0.186 0.93
SEIG 11.7 0.63 0.57 0.516 0.80
VIGA 10.6 0.65 0.48 0.526 0.83
img2threejs 15.0 0.68 0.81 0.272 0.91
medieval-village Ours 18.7 0.71 0.89 0.200 0.93
SEIG 10.9 0.60 0.50 0.541 0.85
VIGA 10.5 0.58 0.37 0.601 0.83
img2threejs 15.4 0.62 0.77 0.325 0.89
japan-island Ours 19.3 0.73 0.87 0.201 0.96
SEIG 13.3 0.64 0.47 0.513 0.93
VIGA 12.9 0.65 0.43 0.534 0.80
img2threejs 16.6 0.66 0.76 0.302 0.95
school-block Ours 16.4 0.56 0.97 0.171 0.94
SEIG 11.8 0.23 0.80 0.448 0.88
VIGA 11.2 0.19 0.77 0.587 0.85
img2threejs 12.8 0.26 0.88 0.310 0.91
police-corner Ours 16.6 0.44 0.97 0.159 0.91
SEIG 12.7 0.24 0.75 0.415 0.90
VIGA 12.4 0.27 0.82 0.473 0.91
img2threejs 13.7 0.30 0.86 0.285 0.90
park-lake Ours 23.3 0.83 0.99 0.075 0.95
SEIG 12.3 0.48 0.61 0.490 0.90
VIGA 15.1 0.54 0.74 0.330 0.92
img2threejs 18.4 0.66 0.92 0.169 0.93
shop-row Ours 17.4 0.65 0.96 0.120 0.96
SEIG 12.5 0.50 0.79 0.400 0.89
VIGA 15.2 0.54 0.91 0.249 0.91
img2threejs 16.9 0.59 0.94 0.145 0.95
valley-village Ours 14.1 0.38 0.48 0.583 0.87
SEIG 9.7 0.39 0.22 0.742 0.64
VIGA 10.2 0.41 0.20 0.743 0.68
img2threejs 11.3 0.34 0.41 0.726 0.70

Download the source table ↗

A recursive construction process

Establish the whole, reconstruct unresolved parts, then revisit their composition. The same solver repeats this process at each level.

The global–local–global solver and its compositional scene programs.
The global–local–global solver and its compositional scene programs.
Recursion trees across scenes
The recursive scene-program trees for the paper’s reconstructed scenes.
The recursive scene-program trees for the paper’s reconstructed scenes.

Explore one of these runs →