A Claude Code skill turns a concept image into a rigged Blender game asset, and refuses to move on until the render matches the reference

An agent that turns concept art into a 3D game asset can pass every review by looking at the viewport and still hand you a knight whose head is 12 percent too tall. Majid Manzarpour published a Claude Code skill on 24 September 2026 called blender-image-to-3d that refuses to move to the next step until the render against the reference passes a measured check. It is MIT-licensed and has picked up 93 stars in its first four days.

What it does

You give Claude Code one or more reference images (front and side views of a knight, a model sheet, a photo of a stone chapel, three sketches of a sword), tell it the game engine and the size, and it walks through eleven phases inside Blender: brief, calibration, blockout, forms, topology, UVs and baking, rig, secondary motion, animation, export, and acceptance. The output is a Blender file plus a GLB or FBX with an asset-manifest.json and, when animation is asked for, an animation-contract.json.

The skill uses the open Agent Skills format. Claude Code is the only agent the author has tested against.

Why the render is the gate

Manzarpour writes that most wasted work in reference-driven modelling is detail added on top of wrong proportions, because the eye forgives drift that numbers catch. So the skill makes the agent measure instead of judge.

Every constant in the build scripts carries a comment naming the reference view it came from, or it is marked # inferred. Each phase ends with renders from a camera matched to the reference at gameplay pixel size, and a compare sheet that reports two numbers: a silhouette IoU (intersection over union) between render and reference, and a ten-band width profile. The blockout has to reach IoU 0.85 with band widths inside 0.05 and proportions within 5 percent. Forms have to reach 0.90, 0.03, and 2 percent. Anything that misses is written as a measurement ("head 12 percent too tall", "wheelbase 0.3 m short"), fixed in the phase script, and re-rendered until it clears.

For orthographic model sheets there is a second gate, world_gate.py, that scores IoU in world coordinates rather than after a bounding-box crop, so a model that is the wrong height or off its axis is caught rather than hidden by the framing. And at the export step the file is imported back into an empty Blender scene and checked against the manifest, because the modelling viewport is not what the game engine receives.

What it requires

Blender 4.2 or newer, tested on 5.2. Python 3 with Pillow for the one compose-review script that runs outside Blender. No GPU is required for the review renders, which use Blender's Workbench engine; baking uses Cycles on the CPU by default. The skill runs the scripts for you through Claude Code; you only need them directly if you want to rerun a phase by hand or work without an agent.

Manzarpour also lists what the skill will not do: it measures dimensions from orthographic views only, uses perspective concept art for silhouette and surface detail rather than proportions, and labels anything the references do not show as inferred in the brief and the final report.

Small skill, one asset type today, from one author. The point of interest is the discipline in the middle. Every gate makes the agent render its current work against the reference and pass a measured check before it earns permission to add the next layer of detail.

Source

Primary source: majidmanzarpour/blender-game-skills on GitHub.