THE FRONTIER OF SPATIAL INTELLIGENCE
From visual evidence.
To executable 3D worlds.
A benchmark for agents that reconstruct Blender scenes from images and design drawings — from individual objects to entire cities.
88reconstruction tasks
6spatial categories
5evaluation metrics
2evaluated models
01 / MODEL PERFORMANCE
The leaderboard
Measured reconstruction quality, averaged equally across evaluated tasks.
Official results
85 of 88 tasks evaluated per modelThree new tasks await evaluation. Unmeasured tasks are excluded from score averages; measured zeros are retained.
View task inventory ↗| RANK | MODEL | OVERALL ↑ | COVERAGE | STATUS |
|---|---|---|---|---|
| Loading official results… | ||||
Scores on a 0–100 scale · higher is betterDownload results JSON ↗
Upcoming evaluations
Planned models are unranked until measured results are available.
02 / PERFORMANCE BREAKDOWN
Across scales. Across signals.
Six scene categories. Each chart compares overall quality and all five normalized metrics.
03 / MODEL COMPARISON
A closer comparison
Choose two evaluated models. Differences are computed on their shared evaluated tasks.
vs.