Bunraku turns single illustrations into editable Live2D characters

Bunraku, a new system from Junhao Chen and eight collaborators at an undisclosed institution, is the first to generate the full structured asset a Live2D runtime needs directly from a single illustration. Live2D, the dominant format for anime avatars and virtual streamers, normally requires weeks of manual layer separation, occlusion completion, mesh placement, and keyframing. Bunraku automates this with a two-stage pipeline: Stage 1 uses a layered diffusion process with a Live2D-aware organ taxonomy to output an ordered stack of RGBA layers, including hidden-region completion. Stage 2 builds a triangle mesh for each layer purely from its alpha channel, then predicts keypose displacement for all layers jointly—treating every vertex as a token and letting self-attention span layer boundaries.

The joint prediction is the key innovation: scaling the network 112x alone yields no improvement, but sharing context across layers makes the result a coherent character rather than separately plausible parts. On 50 held-out characters, Stage 2 achieves a per-vertex direction cosine of 0.768 (median 0.828) under true generation. Because meshes derive from alpha channels, users can re-texture a clothing layer from a natural-language prompt while reusing the mesh and animation byte-for-byte. The team also released Live2D-Bench, the first standardized benchmark for this task, and a corpus of 8,884 Live2D models with layer and animation supervision, enabling further research and practical use in games, streaming, and interactive media.