SceneActBench - Reconstruction: leaderboard

Metric: Task score (0-100): 100 3D-FRONT rooms from about 11 calibrated views: build the furnished scene from an empty Blender scene, scored by F-score at 5%; every configuration drives one headless Blender through the same MCP tool interface (scene and object inspection, Python execution, rendering) with task step budgets; each native geometric error or F-score is mapped to a 0 to 100 case score against a fixed reference, invalid outputs score 0, cases are averaged per task. Source: arxiv.org. Saturation forecast: Around 2032. 11 models tracked.

Top models

#ModelScore
1GPT-5.4 (High)12.3
2Claude Sonnet 5 (High)10.5
3GPT-5.4 (Medium)10.4
4Claude Opus 4.6 (High)9.8
5MiniMax M3 (High)9.6
6Seed 2.0 Pro (High)8.8
7Step 3.7 Flash (High)8.6
8Kimi K2.6 (Thinking)8.5
9Gemini 3.1 Pro (Preview) (High)7.1

Interactive version: theaggregate.ai/benchmark?slug=sceneactbench-reconstruction · How It Works · Data refreshed daily, snapshot 2026-09-29.