Bunraku: Generating Editable Live2D-Style Rigs from a Single Illustration

An illustrated character becomes an editable animation asset when its painted layers, deformation meshes, and parameter-dependent motion are stored together. Bunraku generates an open Live2D-style rig from one illustration and a requested layer count: layered diffusion generates RGBA slots, surviving alpha masks define triangle meshes, and a shared-frame Transformer predicts named-control vertex offsets. The asset replays in WebGL rather than native Cubism. We assess the representation with artist rigs, generated layers, and same-architecture attention controls. A post hoc 50-character artist-vertex audit favors joint over layer-blocked attention (moving error 0.0493 versus 0.0532), but a separate prospective 12-character one-seed test does not confirm this effect (0.0643 versus 0.0640). On 12 newly generated rigs, joint and blocked motion have source-mean rest-subtracted rendering errors 1.359 and 1.323 versus 1.000 for static layers; joint improves none of the 12 sources over static. A draw-order training repair lowers joint error from 1.495 to 1.196 on six new source matches, yet joint and blocked motion still trail static layers (1.196 and 1.170 versus 1.000). Only one of the original 12 rigs retains the requested number of usable layers. On a frozen 34-asset high-layer subset, after correcting the checkpoint-specific output-order convention, fine-tuning improves artist-matched alpha IoU (0.256 versus 0.173), pair-weighted order (0.623 versus 0.528), and composite PSNR (26.36 versus 20.91 dB), but reduces the count of alpha-filter-eligible layers. The historical 120-image comparison overlaps training names, and missing earlier manifests prevent certifying Stage 1 source disjointness. Bunraku yields editable controls, while reliable generated-rig animation and rights-cleared replication remain open.