Reference image → agent → headless Blender

Watch an agent rebuild images without opening the GUI.

It reads the reference, writes Python through the CLI, renders headlessly, inspects the result, and iterates.

BlenderBench / Run 00127 tasks · 270 rounds
The full run: reference image and generated scene, task by task.The on-screen percentage is CLIP image-embedding cosine similarity—not literal task accuracy.Read the video description and results · Watch/download on Hugging Face
Open community preview

Give your agent a reference image. Let it build the Blender scene.

blender-mcp makes Blender a headless backend for coding agents. The agent uses the image as its goal, drives Blender through a narrow CLI and MCP surface, then renders and inspects its own work as it goes.

Standalone bpy packaging remains work in progress. Track build work in PR #9.

01 / What developers can do

Blender is the backend. Your agent does the work.

Use the same Python runtime directly, or expose a small set of operations to an agent.

Use it from Python

Build render pipelines, tests, and batch jobs around import bpy in a normal Python workflow.

Python API

Let an agent drive it

Give an agent four CLI-friendly MCP tools to write Python, render the scene, inspect the image, and keep working.

Inspect the interface

02 / How it works

Reference → Build → Inspect

The loop is visual, but the backend is headless. No one has to sit in front of Blender.

  1. 01

    Reference

    The agent gets an image and a goal—rebuild what it sees as a real Blender scene.

  2. 02

    Build

    It writes Python through the CLI and runs Blender as a headless backend.

  3. 03

    Inspect

    It renders its own work, looks at the result, and keeps going. The GUI never needs to open.

03 / Initial evidence

Measured in renders, not claims.

One transparent community-preview run on the public BlenderBench dataset.

27
public BlenderBench tasks
270
GPU-rendered and GPU-CLIP-scored checkpoints
4
exactly four MCP tools
10 rounds per taskno VLM judgeno score feedback during generation

Video description and results

Text alternative for the silent compilation

The compilation covers 27 tasks and 270 rounds. Each checkpoint identifies its task and round, places the generated scene beside its target render, and reports the CLIP image-embedding cosine similarity and generated Python token count shown on screen.

Open the detailed per-task results on Hugging Face.

Category results

Mean best distance by difficulty

Category mean best N-CLIP ×100. N-CLIP is derived from CLIP image-embedding cosine similarity; lower is better. Values retain the precision of the recorded aggregate.

Category mean best N-CLIP times 100
CategoryN-CLIP ×100
Camera / Level 10.3738476170433892
Level 21.603147718641493
Level 32.2820386621687145
Best-round generated bedroom scene beside the BlenderBench target render for task 24 of 27, round 7 of 10.
Best-round checkpoint and target from the initial run.Target data: BlenderBench by DietCoke4671 and contributors, dataset revision 203e4d325e9438ca55b29bdfc4f6a90842d74e68, licensed CC BY 4.0.

Open community preview

Test the boundary. Question the benchmark. Shape what ships.