HairDeck turns one image into animatable hair cards, 33× faster than CHARM
TL;DR
- HairDeck generates stylized hair cards from a single image by treating each card as one token and denoising the whole bundle with a set diffusion transformer.
- On CHARM's test set it reports a 44% drop in Chamfer distance (0.0117 to 0.0066) and card-orientation error of 16.8° versus 40.0°.
- Card generation runs in 1.01 seconds versus CHARM's 33.7 seconds, with VRM avatar export carrying per-card UVs and spring bones.
"The hair of stylized characters in games and avatars is built from hair cards: artist-placed ribbons whose geometry is the visible hair, animated by spring bones." That is the opening framing of a new preprint from a team at Tsinghua, CMU and CUHK, HairDeck, which proposes to automate that whole job from a single input image.
The method represents each card as a single token and runs a set diffusion transformer across the full bundle at once. A second transformer generates the spring-bone chains that drive the hair in-engine, with geometric skinning binding cards to bones. The authors report VRM avatar export with per-card UVs and spring bones included.
The headline comparison is against CHARM, the prior specialized hair-card generator. On CHARM's own test set, HairDeck reports a Chamfer distance of 0.0066 versus 0.0117, a card-orientation error of 16.8° versus 40.0°, and a generation time of 1.01 seconds versus 33.7 seconds, the 33× speedup the authors advertise.
Code and weights are posted on GitHub and Hugging Face. The project page does not state a license.
Originally reported by paper
Read the original article →Original headline: HairDeck: One Image → Game-Ready Hair Cards With Spring Bones, 33× Faster