PhysEditBench - Normals: leaderboard

Surface normals from image editors prompted to turn indoor photographs into physical maps (PhysEditBench, 2026) averaged over three runs, each prompted with one fixed RGB–normal exemplar pair: the share of pixels within 22.5 degrees of the reference normal, in percent. Higher is better.

Metric: Acc@22.5 times 100 (%): share of pixels whose predicted surface normal is within 22.5 degrees of the reference, averaged over three runs, each prompted with one fixed RGB-normal exemplar pair, source-balanced; prompted image editors on the paper main-axis split (indoor OpenRooms-FF and InteriorVerse scenes; official access settings of Appendix Table 19); task-specific specialist models are reference rows, not published; higher is better. Source: arxiv.org. Saturation forecast: Rough model projection: around 2026. 2 models tracked. Updated 7 Oct 2026.

Top models

#ModelScoreOverall rank
1GPT Image 1.520.4
2doubao-seedream-5-0-26012813.2

Interactive version: theaggregate.ai/benchmark?slug=physeditbench-normals · How It Works · Data refreshed daily, snapshot 2026-11-21.