Blender AI Benchmark Prompt Set v1.0

The standardized prompts every tool in the Blender AI Arena runs each season. 12 prompts, 6 categories, identical rules for every tool. Free to use for your own testing under CC BY 4.0 — attribution: “Blender AI Arena Benchmark Prompt Set” with a link to this page.

Benchmark Rules

  1. Use the prompt text verbatim — no tool-specific rewording or prompt engineering.
  2. Default settings for every tool; no expert parameter tuning.
  3. One generation per prompt. One retry allowed only on technical failure (crash, timeout, empty output).
  4. No manual cleanup afterward: no retopology, material fixes, or deletions.
  5. Export to GLB and render with the standard arena lighting rig for comparison.
  6. Publish every result, including failures.

The Prompts

P01Props

A wooden treasure chest with iron fittings and a curved lid, slightly weathered, closed.

Tests: Basic hard-surface prop with mixed materials and readable silhouette.

P02Props

A desk lamp with an articulated two-joint arm and a round metal shade, in a bent working pose.

Tests: Mechanical structure, joint articulation, and plausible proportions.

P03Characters

A low-poly stylized knight character in idle pose, with helmet, shield on back, T-pose acceptable.

Tests: Character topology, symmetry, and stylization control.

P04Characters

A cute cartoon mushroom creature with big eyes and stubby legs, standing.

Tests: Organic shapes, appeal, and character readability at small scale.

P05Architecture

A small medieval stone watchtower with a wooden door, arrow slits, and a conical roof.

Tests: Architectural massing, repeated elements, and material separation.

P06Architecture

A modern single-story house with flat roof, large windows, and an attached carport.

Tests: Clean planar geometry, boolean-like openings, and real-world proportions.

P07Vehicles

A low-poly pickup truck with visible wheels, cabin, and cargo bed, game-asset style.

Tests: Vehicle proportions, wheel placement, and part separation.

P08Vehicles

A small sailing boat with hull, single mast, boom, and simplified sail.

Tests: Curved hull geometry and thin-structure handling (mast, rigging).

P09Nature

A potted plant: terracotta pot with a leafy monstera-style plant, about six large leaves.

Tests: Organic foliage, thin geometry, and countable-element accuracy.

P10Nature

A low-poly island scene: small rocky island with two palm trees and a sandy beach.

Tests: Simple scene composition and multiple-object arrangement.

P11Abstract

A spiral staircase with twelve steps winding around a central column, with a handrail.

Tests: Parametric/procedural reasoning: counted repetition, rotation, and precision.

P12Abstract

A chess set arrangement: checkered board with the two front rows of white pieces placed correctly.

Tests: Multi-object precision, placement rules, and instruction following.

Why a Fixed Prompt Set?

Most AI 3D comparisons show each tool's best cherry-picked output on different prompts — which measures marketing, not capability. A fixed public prompt set means every tool faces the same tasks, weaknesses are visible, and results can be reproduced by anyone.

The set intentionally mixes tasks that favor mesh generators (organic characters, nature) with tasks that favor code-driven assistants (counted repetition, precise placement), so no tool category gets a free win. Per-category results are published on the leaderboard.

Using this set in your own video, article, or research? We'd love to see it — email hello@blenderai.org.

Want AI That Actually Creates in Blender?

3d-agent is the AI agent that connects directly to Blender via MCP — generating, modifying, and rendering 3D models with natural language.

Try 3D-Agent Free →