BENCHMARKS / 3D MODELING
CharacterBench.
Can AI models build an editable 3D character from an artist’s reference images? We’re building the benchmark and working out how to grade it. The first study is out.
THE FIRST STUDY / NOVEMBER 2026
Three characters, three models, nine builds.



- 01 / THE TASK
A folder of images, eight hours
Each model got one character: renders of the artist’s model from four angles, close views of the head and some original design art. It had eight hours in Blender to hand back an editable file and four renders.
- 02 / THE RUNS
Compared with the artist’s model
GPT-6 Astra, Muse Wonder and Gemini 4 Argon each ran once per character at their highest reasoning setting. The artist’s model was never in the folder; we used it only to compare.
- 03 / THE GAP
Shape is close, character isn’t
The best builds get the overall shape close. The face, the joins between parts and the surface detail still differ, and whole-body scores barely register it.