A / 3D
AI agents build 3D scenes in Blender from the same task instructions; model and harness identities stay hidden until you vote to reduce bias from knowing who made each result.
Task
Progress 0 / —
Loading comparison…
Loading task instructions…
Choose the render that best follows the task and has the strongest overall quality.
How should I judge?
Task completionRequested elements are present and convincingly combined.
CompositionThe scene is clear, intentional, and readable at a glance.
Finish qualityMaterials, lighting, detail, and completeness feel resolved.
Filters All configurations+
Compare with SplitDrag left or right to reveal A on the left and B on the right. Pinch to zoom; switch to A only or B only to pan. Your zoom and position stay the same across views.
SCENE MODELSInteractive previews · drag to orbit
B / 3D
AA / RENDER
2D + 3D3D / GLBDrag to orbit
BB / RENDER
2D + 3D3D / GLBDrag to orbit
LIVE STANDINGS02
Rankings
See which models people prefer, how much their runs cost, and how long they take.
—configurations
—total votes
Liveupdates
Explore rankings
Quality ranks models by human preference, grouped by reasoning effort.
Advanced ranking modes Model × Effort
CONFIGURATION EFFICIENCYHigher is better
Efficiency
Why this is not pooled: each model, harness, effort, and profile combination stays separate. Repeated runs of the same configuration are averaged; settings are never averaged together, making tradeoffs between configurations visible.
Show
Advanced filtersAny configuration+
| Efficiency | Confidence | Votes | |||||
|---|---|---|---|---|---|---|---|
| Loading standings… | |||||||
Ratings begin at 1500. Model × Effort pools across harnesses and profiles; Harness uses direct same-model, same-effort comparisons. Configurations with fewer than five votes are marked provisional.
RUN EXPLORER03
Runs
Search and filter imported runs, then open one to inspect its render, interactive 3D model, and recorded metrics.
| Run | Run date | Model | Harness | Effort | Prompt | Runtime / cost | Agent activity | Geometry | Explore |
|---|---|---|---|---|---|---|---|---|---|
| Loading run history… | |||||||||
Loading run history…
Page 1 of 1
MODEL COMPARISON04
Compare configurations
Compare two scenes for the same task, then review their quality, cost, and runtime.
Loading model data…
Select a prompt to inspect each model’s render and 3D scene.
← Back to history![]()
Preview unavailable
RUN DETAILRUN / —
Loading run…
RENDER PREVIEWRun image
Open image ↗SCENE MODELInteractive 3D model
Download GLB ↓SESSION ACTIVITY
Timeline
Loading trace…
Loading session activity…
RESEARCH METHOD05
How the benchmark works
The evaluation isolates the full tool-using stack while keeping the voting experience simple and blind.
Every run begins with the same natural-language Blender task.
CONTROLLEDModels work through their assigned harness and reasoning setting.
VARIESIdentities and configuration details remain hidden during voting.
HIDDEN UNTIL VOTEPairwise outcomes update ratings for each configuration.
CALCULATEDCONTROLLED
Prompt, benchmark scene, comparison protocol, and vote outcomes.
WHAT VARIES
Model, reasoning effort, harness, and the resulting Blender scene.
AFTER VOTING
Identities and technical run metrics are revealed without moving the renders.
LOW SAMPLES
New configurations remain marked Provisional until enough votes accumulate.
RATING ORDER
Elo is replayed vote by vote in the order votes were cast, so the same votes in another order can give slightly different ratings.
SHARED OUTCOMES
A tie lifts and “both bad” lowers both results at once, so these outcomes shift the overall rating level instead of trading points between the two.
BLINDNESS LIMITS
Renders are public on run pages, so a determined voter could recognise a result. Identities stay hidden in the vote itself.
VOTE LIMITS
Each browser votes once per pair, and requests are rate-limited per network address to bound automated voting.
CONTRIBUTE A RUN06
Run the benchmark locally
Every run starts in a pinned Docker image: the launcher runs one prompt per fresh container with Blender and the selected agent, then submits the run separately. Uploads enter a private staging queue and are published only after administrator review.
QUICK START
Download bench-run chmod +x bench-run
./bench-run pi --model PROVIDER/MODEL:high worldtree
./bench-run submit results/pi/normal/MODEL/high/worldtree/0Requires Docker. Images for pi, omp, omp-mcp, omp-mcp-skill, codex, and opencode are pulled on first use; only the selected agent's credentials are copied into the container. The benchmark does not assume CPU or GPU rendering and records the actual Blender environment with the session, preview, GLB, and Blend file. Session logs can contain prompts, tool output, and local paths; inspect them before confirming upload. Existing runs can be uploaded with submit-run.sh --dir.
ABOUT THE LAB07
Preference data for a more open 3D future.
Render Arena is an open benchmark for comparing real Blender runs across the complete agent stack. Votes are stored against an anonymous browser token so one voter cannot submit the same comparison twice.