How well can AI agents build a 3D scene?
Render Arena is an open benchmark for AI systems that create scenes in Blender. It compares the complete working setup—model, reasoning effort, and agent harness—using real Blender outputs and human judgments.
What gets tested
Each run asks an AI agent to create a Blender scene from a standardized natural-language task. Tasks range from simple objects to detailed environments. The agent works through a harness, the software that connects the model to tools such as Blender and manages its session.
The benchmark varies three parts of the configuration independently:
- Model: the AI model generating and directing the work.
- Reasoning effort: the requested reasoning level, where the model and harness support it.
- Harness: the agent software and tool workflow used to complete the task.
What is measured
The primary quality signal comes from blind, pairwise comparisons. Voters see two renders for the same task without model or configuration identities, then choose the stronger result, a tie, or that both are poor. The site also provides the 3D scene files so results can be inspected beyond a single image.
Ratings are calculated from these votes. Separate views compare model-and-effort groups, full configurations, and harnesses on matched model and effort settings. Low-sample ratings are marked provisional.
Quality, cost, and speed
Alongside preference ratings, the leaderboard reports available run-level usage data: estimated model cost, input and output tokens, and elapsed runtime. Efficiency views relate vote-derived quality to known cost or reported token use; runtime is shown as an average.
These operational figures depend on what a run reports. Missing or unknown values are identified rather than treated as measured zeros. Render Arena does not claim that one resource profile is universally best: these views make quality and resource tradeoffs easier to compare.
Why compare the whole setup?
A model result can depend on more than the model name. Reasoning settings can change how much effort it spends, while harnesses differ in tool access, orchestration, and execution. Render Arena therefore compares configurations across models and also examines the same models at different reasoning levels and through different harnesses.
Where enough matched runs exist, the harness view uses direct comparisons with the same model and effort. Full-stack views keep harness profiles distinct; the model-and-effort view pools across harnesses, so each view answers a different question.
How to read the results
Rankings summarize submitted runs and accumulated human votes, not a guarantee of performance on every task. Sample counts matter: a provisional entry has limited evidence. Use filters and comparison pages to inspect the task mix, harness, reasoning effort, and run history behind a rating.
Voting is blind to configuration identity, but it is still human preference: voters judge task completion and visual quality, not a universal objective measure of 3D correctness.
Open participation
Runs can be submitted with their Blender artifacts and session records. Submissions are reviewed before publication. This makes it possible to inspect both the resulting scene and, when available, the agent activity that produced it.
Render Arena is intended to make practical agent performance visible across the complete tool-using stack.