Gokayfem releases open-source Claude Opus 5.5 skills for Blender sculptures

• An open-source skill turns reference images into measured Blender sculpts, with gokayfem's release linking the reusable toolkit.
• The showcased Opus 5.5 run took three hours at medium effort, according to gokayfem's timing reply.
• Depth maps, normals and masks guide the reconstruction; gokayfem's explanation also credits Nano Banana with generating additional reference angles.

One requested +90° view registered at −123° in the developer's first-case notes. The published skill reserves a generated view that never influences fitting, while the repository records geometry changes for a Blender timelapse.

Gokayfem put the demo at three hours on medium effort and confirmed that Blender was the 3D application. The replayable build is the clever bit: the toolkit saves intermediate geometry, then renders those steps into the sculpting film.

The skill instructions give the coding agent three kinds of front-image evidence:

• Mask: BiRefNet, SAM3 selection or an existing PNG alpha channel defines the subject.
• Depth and camera: MoGe estimates depth and focal length.
• Surface normals: Marigold estimates the directions the visible surface faces.

The original image governs the surfaces it can see. Generated references provide constraints for hidden surfaces.

Nano Banana supplied additional angles for the demo. The default workflow generates nine fitting views plus one held-out view; photographs supplied by the user can replace generated views.

In a separate discussion, gokayfem said Seedream reproduced a style more closely with two or three Midjourney references than with text prompting alone.

Camera registration provides three concrete requested-to-recovered comparisons in the first-case development notes. Expressed on a 0-360° azimuth scale, with shortest-arc corrections:

• Camera azimuth: 45 → 19 degrees, -26 degrees.
• Camera azimuth: 90 → 237 degrees, +147 degrees. The notes report the equivalent angle of -123°.
• Camera azimuth: 270 → 35 degrees, +125 degrees across zero. The requested 270° is equivalent to the notes' -90°.

These comparisons concern generated-view camera alignment. The pipeline also screens views against the hull formed by the others, rejecting inconsistent views below a silhouette-overlap score of 0.80, according to the skill instructions.

The published reconstruction plan breaks the build into recorded operations:

1. Hull: register cameras, screen views and carve a starting volume from silhouettes.
2. Forms → surface → detail: run 38 fitting passes, using stiff solves for outlines and depth, and flexible solves for normals.
3. Gap cutting and cleanup: open visible spaces and remove webs or specks.
4. Polish: apply four finishing passes.

Symmetry acts on fitting targets. Focus regions such as hands or face let the agent add localized passes without rewriting earlier steps, as described in the skill.

Delivery includes:

• Editable geometry: a .blend file and GLB, OBJ and STL exports.
• Process footage: the step-by-step replay, a 360° turntable and renders from the reference cameras.
• Review artifacts: map sheets, a Markdown report and a static review page.

Each case tracks six measurements, with these default acceptance targets in the checking guide:

• Silhouette overlap, or IoU: at least 0.97 for the source image, 0.92 for the fitting-view mean and 0.85 for the held-out view.
• Outline distance: at most 2.5 pixels for the source image.
• Median normal error: at most 10° for the source, 18° for the fitting-view mean and 22° for the held-out view.
• Normal agreement: the share of pixels within 15°, reported without a target.
• Relative roughness: 0.75-1.25 times reference roughness for the fitting-view mean.
• Depth error: at most 15 thousandths of the subject's size on the source image.

A separate end-to-end test uses a synthetic figure with noisy normals, incorrect camera angles and one changed pose. Passing requires the fitted surface to be closer to the known shape than the starting hull, with held-out silhouette IoU above 0.85 and source-view IoU above 0.95.

Individual metrics can improve while the sculpt gets worse elsewhere. During development, silhouette overlap rose while roughness doubled, a +100% roughness increase; normal agreement also improved while a gap closed, according to the first-case notes.

The held-out generated view measures consistency among references; the unseen back remains the image generator's guess, as the skill explicitly states.

The documented scope excludes:

• Scenes, rooms, hair-thin structures, and transparent or mirror materials.
• Rigging, animation, PBR texturing and CAD-accurate parts.

Gokayfem described the experiment as a demonstration made for fun and to show LLM progress, and questioned the practicality of sculpting token by token every time.

After promising a release within hours, gokayfem published the MIT-licensed toolkit. The installation documentation lists these Claude Code commands:

For Codex, Cursor, Gemini CLI, Copilot and other agents, the repository also provides:

The standard API-backed setup uses Python 3.10+, Blender 5.2+, ffmpeg and a fal API key in FAL_KEY. A no-key route accepts a cut-out PNG and supplied views; front-only reconstruction produces an inflated silhouette with guessed depth.

The README estimates about $2 for a typical case's generated views and maps, with a per-case budget controlling paid calls. For the demo, gokayfem's cost breakdown also included three hours of medium-effort Opus 5.5 usage.