how well can AI agents build a 3D scene?
Render Arena is an open benchmark for AI agents that create Blender scenes. Its goal is to make practical 3D work inspectable: the resulting artifacts, the agent configuration, and the human judgments behind a ranking.
[4] why this project exists
A finished image tells only part of the story of an agent’s work. Render Arena brings submitted Blender artifacts and available session records together so people can inspect how a result was made.
The aim is to help researchers, builders and artists investigate practical agent performance and contribute evidence of their own.
[5] evaluation and evidence
Read the methodology for the comparison protocol, ranking pools, Elo formulas, tie handling and known limits. It also explains sample sizes and why the current ratings have no statistical confidence intervals.
Explore rankings, inspect individual runs and artifacts, or examine matched comparisons to see the evidence behind the results.
[6] open participation
You can contribute by voting on results or submitting a run with its Blender artifacts and session records. Submissions are reviewed before publication.
The submission guide explains the local tools you need, how to run the benchmark, and how to inspect logs before uploading.
[7] voting and privacy
An essential browser cookie identifies an anonymous voter token so the same browser cannot vote on the same pair twice. This does not verify a voter’s real-world identity.
Published runs may include session logs with prompts, tool output and local paths. Contributors should inspect those files before submission. Read the voting and visibility limits.