SceneActBench - Layout: leaderboard

Metric: Task score (0-100): 100 3D-FRONT rooms from one view: place N canonical object meshes in their poses, scored by ADD-S error against a 4 m reference; 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 March 2028. 11 models tracked.

Top models

#ModelScore
1GPT-5.4 (High)84.1
2Step 3.7 Flash (High)77.5
3Seed 2.0 Pro (High)77.4
4Claude Opus 4.6 (High)73.5
5GPT-5.4 (Medium)72.7
6Kimi K2.6 (Thinking)70.9
7Gemini 3.1 Pro (Preview) (High)65.4
8MiniMax M3 (High)58.1
9Claude Sonnet 5 (High)51.9

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