SIFT keypoints, not diffusion: Claude rigs your Spine characters

What it does

It’s a Claude agent skill — a single self-contained SKILL.md — that turns separated body-part PNGs into a working Spine 4.2 character. The pipeline positions each part against a reference image, builds the skeleton JSON with bone hierarchy and draw order, packs a texture atlas, and emits a self-contained HTML preview running the official Spine Web Player. It also ships six animation presets, proposes corrections to existing rigs in a reviewable format, and a new Reskin Studio repaints a rigged character from a one-line prompt via Gemini, re-slicing the result onto the live rig. A complete example — a 10-part, 16-bone sombrero character with a breathing idle loop — proves the whole chain works end to end.

The interesting bit

The cleverest engineering here is decades-old computer vision, not a language model: SIFT keypoints matched with FLANN, then a RANSAC-fitted similarity transform per part — deliberately chosen over full homography because stylized game art has too few features for one. Draw order comes from an occlusion graph: for each overlapping pair, pixels in the overlap are compared against the reference, the closer color wins “on top,” and a topological sort turns that into the z-order. Claude is the conductor, not the soloist.

Key highlights

• Similarity-transform placement with a template-matching fallback for parts too small to yield SIFT keypoints
• Draw order derived from per-pixel occlusion analysis rather than hand-tuned layer lists
• Six presets (idle, walk, run, wave, jump, attack) built from overlapping sine waves and bezier easing, with anticipation and follow-through
• The adjustment format records original offset, AI-suggested delta, and final value, so corrections stay reviewable and revertible
• The core pipeline runs locally on OpenCV, Pillow, and numpy; only character-splitting and reskinning call out to Gemini

Caveats

• PolyForm NC 1.0 — a non-commercial license, which matters a lot if you’re shipping a game
• The motion is procedural sine-wave presets, not learned animation; don’t come expecting a motion model
• The README still carries a TODO for the reskin demo gif, so that feature’s showcase is pending

Verdict

Indie devs who already live in Spine and Claude will save real hours on the placement-and-plumbing grind. If you wanted learned motion synthesis, or a license you can ship under, keep looking.