Transformer BumbleBee.

A transformer is the hardest kind of configurator: every panel has to travel between two states — robot and vehicle — without ever intersecting itself along the way. The base model came out of Tripo3D, and the first real work was separation: cutting the generated mesh into parts that each have one job — armor plates, limbs, wheels, the core chassis — so each one can be rigged with a clean path between forms.

With the parts tree in place, I used GPT Astra to drive Blender through the optimisation pass: auditing topology per part, rebuilding where the generated mesh was mush, choosing pivot points and parenting so every plate folds rather than floats, and holding the whole assembly to a polygon budget a browser can carry. Rig order matters more than polygon count here — a transformation is choreography, and the rig decides the choreography.

The scene is a Three.js WebGPU build. The fold itself is a staged deformation timeline — limbs tuck, panels slide, the spine re-couples into the drivetrain — driven through TSL so the same materials hold up in both forms. Dark stage, warm key light, a floor that keeps the reflection honest, and the transformation plays back in the viewport at full rate. What you are watching is the scene, not a baked animation.