Taxon3D
A blind-comparison benchmark for AI-generated 3D models of living organisms, judged against reference photographs. Bradley-Terry rankings with confidence intervals.
Most benchmarks for generative 3D models use furniture and game props, where looking plausible is good enough. Organisms are harder. A maize plant or a monarch butterfly has branching structure, thin surfaces and a lot of self-occlusion, and whether the model is right is a question of anatomy, not aesthetics.
Taxon3D is a Chatbot-Arena-style site for blind comparison of 3D generators on exactly those subjects. You get a task, two anonymised 3D models, and reference photographs of the real organism above them. Rotate, zoom, and pick the one closer to the photos.
What the board covers
Twenty active tasks across plants, fungi and animals; 488 votable models from 52 entrants, spanning single-image reconstruction, text-to-3D, LLM-authored procedural geometry, and agentic render-critique-revise loops. Those counts are measured from the live board by a script and held there by a test.
What keeps the numbers honest
- Bradley-Terry with bootstrap 95% confidence intervals. Models whose intervals overlap share a rank, so a statistical tie reads as a tie instead of being ordered by noise.
- Pairwise significance by paired bootstrap: the probability that A ranks above B, and whether a model actually beats the next rank down.
- A published bias audit: left-win rate, tie and both-bad rates, and cross-format confounds are measured rather than assumed away.
- Vote integrity: gold-standard attention checks score voter trust, low-trust sessions are excluded from the fit, and a generator is never matched against itself.
- Each board shows how many more votes it would take to separate its next pair of models. At the time of writing that is under twenty per board.
The project was called Bio 3D Arena until August 2026, when I renamed it because the name collided with the bio3d R package and with Arena3D. The old host and repository path still work.