Leaderboard / Model
3D-Agent
by 3D-AgentagentFree tier, Pro plans for teams
AI agent platform connecting to Blender via MCP, driving the Python API to build editable scenes rather than importing a baked mesh.
Season 2 · outputs
12 unedited results on the benchmark
Each competitor generated its entries independently — no shared code or outputs. Click a tile to orbit it in 3D. Vote on these blind →
3D-AgentP01 · Props 3D-AgentP02 · Props 3D-AgentP03 · Characters 3D-AgentP04 · Characters 3D-AgentP05 · Architecture 3D-AgentP06 · Architecture 3D-AgentP07 · Vehicles 3D-AgentP08 · Vehicles 3D-AgentP09 · Nature 3D-AgentP10 · Nature 3D-AgentP11 · Abstract 3D-AgentP12 · Abstract
Earlier seasons stay published in the Season 1 archive.
Season 2Ranked
- 1
- Entrants
- 12
- Outputs
- of 12 prompts
- 248
- Ballot appearances
- 3D-AgentProvisional1727 ±66164–62–16
Season 1 final 3D-Agent finished 1= (technical draw) on 1760 ±67. Full Season 1 results →
Every competitor starts each season at the Glicko-2 default of 1500 ±350 and stays unranked until it clears the calibration floors — no rating carries over between seasons. Ratings move only on real head-to-head votes, and every ballot is published. Read the methodology.
What is 3D-Agent?
3D-Agent is an AI agent that connects directly to Blender through the Model Context Protocol (MCP) and Blender's Python API. Instead of generating a mesh on a remote server and handing you a file, it works inside your scene: creating objects, editing modifiers, setting up materials, and rendering — all driven by natural-language instructions.
Because it operates on the actual Blender scene graph, its output is fully editable: real objects with modifier stacks and clean object hierarchies, not a baked-down import. It orchestrates multiple steps (model, texture, light, render) in one session.
It entered Season 1 on an earlier release and finished in the four-way technical draw for first, on 8 of the 12 prompts. For Season 2 it re-ran the complete prompt set on a rebuilt pipeline with updated underlying models — the entrant is the product, so it competes under the same name, and both runs stay published side by side.
3D-Agent follows the same published benchmark rules as every other tool in the arena. See the methodology page for the evaluation protocol.
Strengths and limitations
Strengths
- Strength:Works natively inside Blender — output is editable scene objects, not a static mesh import
- Strength:Multi-step agent workflow: modeling, materials, lighting, and rendering in one conversation
- Strength:Model-agnostic by design — the underlying models are swapped as stronger ones ship, without changing the workflow
- Strength:Keeps geometry parametric where possible (modifiers, curves), which artists can refine
Limitations
- Limitation:Procedural approach: organic sculpt-like shapes are harder than for dedicated 3D generators
- Limitation:Requires Blender installed and the plugin configured — heavier setup than a web generator
- Limitation:Quality depends on the underlying models and prompt clarity, and both move as the platform updates
How to use 3D-Agent with Blender
Install the 3D-Agent plugin, connect it to Blender, and type instructions in plain language. The agent executes them live in your open scene via MCP and the Python API.
- Best for
- Blender users who want an AI collaborator inside the viewport — scene assembly, hard-surface modeling, repetitive edits, and automation.