HairDeck Turns Images Into Rigged 3D Hair

HairDeck targets the production bottleneck generic image-to-3D systems usually dodge: usable, rigged stylized hair rather than a fused mesh. It looks valuable for game and VTuber workflows, though single-view ambiguity and a demanding setup mean it remains an artist accelerator, not a push-button asset pipeline.

• –A set diffusion transformer generates hair cards jointly, preserving individual card geometry and UVs.
• –A second model generates spring-bone chains, enabling animated GLB and VRM exports.
• –Reported CHARM benchmark results show 1.01-second generation versus 33.7 seconds, with orientation error dropping from 40.0° to 16.8°.
• –Open code, model weights, training scripts, local demos, and Apple Silicon support make the project unusually accessible for research prototyping.
• –Hidden geometry remains inherently ambiguous from one image, and dependencies such as gated SAM 3, DINOv3, and Blender add setup friction.