Drawn, Not Computed: Measuring Live2D-Style Turn Keyforms from Generated Views of a Single Anime Illustration

A hand-rigged Live2D character turns its head convincingly because a rigger draws, for each part, the shape it takes at the extreme of each turn parameter (a keyform) and the runtime interpolates. Automatic pipelines that start from a single illustration either compute the turn from a depth estimate, which reads as a sliding card beyond small angles, or learn to predict deformations from a corpus of rigged models. We present a third route that needs no training: we ask an off-the-shelf image-editing model to redraw the same illustration with the head turned about 30° right, left, up and down, decompose all five images into semantic layers, and measure each part's keyform by fitting a lattice deformation of the front part onto the same part in each turned view. What a turn uncovers (back hair, ears, the side of the neck) is baked into the front layers from the turned views. Because generated views are unreliable, every view is screened before fitting by heuristic checks calibrated on a small reference set (fringe area, head height, unchanged outfit), and a wrong-direction body turn is replaced by the mirrored fit of the opposite side. The generative model is used once, offline, at import; playback is a deterministic mesh deformation that runs in a browser with no inference. We describe the full system, including a full-body extension with body-turn and weight-shift keys, a skeleton tree built at playback from detected joints, and layer repair by reconciliation against the source illustration. The evidence is small and stated as such. On two bust imports, under one rendering protocol, per-part lattice keys bring the head's outline closer to the generated views (mean 1.5–1.9 px) than a jointly fitted whole-head lattice (2.2–3.3 px), per-part affine maps (2.3–2.8 px) or no motion (2.6–5.0 px). A check of the combined turn-and-nod square found folds and non-converging inversion in the first runtime; with a per-cell lattice solve and composed corner poses neither remains, and the compression left at the corners is what the two axis keys predict together.

This record contains the report, its LaTeX source with reproduction scripts, and the raw archives (source illustration, generated views, decompositions, keys) of the four keyed imports it evaluates. See README.md.