HairDeck creates animated stylized hair cards from images
HairDeck: Animation-Ready Stylized Hair Cards from a Single Image · arXiv:2610.40599
In plain words
Stylized character hair in games and avatars is often made from many flat cards, with small joint chains that let the hair move. HairDeck is a method that generates both the cards and their animation chains from one image, or generates cards from a point cloud.
On 100 point-cloud test hairstyles, HairDeck had lower average shape distance than CHARM under CHARM's test protocol. The result describes that test set, not every hairstyle.
The research
Can a single image provide the card-based structure needed to make stylized character hair, including the parts that let it move? Existing image methods do not automatically produce this asset, so the paper tests a method that generates both the visible cards and their animation chains.
HairDeck encodes each hair card as a token, generates the cards together as an unordered set, then generates spring-bone chains from them.
On 100 hairstyles from CHARM's point-cloud test, the authors compared HairDeck with CHARM using the same test cases and protocol. The protocol compares generated cards with reference geometry reconstructed from templates; lower Chamfer distance means a closer shape match.
HairDeck's mean shape distance was 0.0066 versus 0.0117. Card orientation is measured separately against the original artist meshes: mean error was 16.8° versus 40.0°. Generation took 1.01 rather than 33.7 seconds per hairstyle.
These averages support improved measured card shape and orientation, with shorter generation time, on this point-cloud test set.
These results cover the stated test populations and stylized VRoid training data. Very short hair and realistic hair cards were outside that training data; the system also generates no textures, can turn visible props into cards, and may produce too few chains for unusual layouts.
What the evidence shows
- The missing card-based asset
HairDeck is a method for generating stylized hair cards and the chains that animate them from a single image or point cloud. Stylized game and avatar hair uses artist-placed cards whose geometry forms the visible hair, but existing image-based methods do not automatically produce this card-based asset. The paper asks whether those cards and their animation structure can be generated together.
- Point-cloud evidence
The CHARM test measures card generation from point clouds on 100 test hairstyles, comparing generated cards with reference geometry and measuring orientation against artist meshes. HairDeck's mean shape distance was 0.0066 versus 0.0117 for CHARM, and mean orientation error was 16.8° versus 40.0°.
It took 1.01 rather than 33.7 seconds per hairstyle. These are averages for this test set and protocol.
- A test of joint generation
A second comparison used the same card-token type, card VAE, data, condition, training steps, and similar parameter counts. One system generated tokens one by one; HairDeck generated them together as an unordered set.
On the 100 CHARM test hairstyles, HairDeck's mean shape distance was 0.0066 versus 0.0078, and it ran 6.8 times faster. The comparison does not establish that their training objectives were identical.
- From image to hairstyle
For single-image generation, HairDeck was compared with four alternatives on front-view renders of 76 test characters. Its mean aligned shape distance was 0.0089, versus 0.0139 to 0.0186 for the alternatives.
This supports the method in that rendered-image setting; it is separate from the point-cloud test and does not include the in-the-wild image results.
- Evidence for animation chains
HairDeck also generates animation chains. On 58 test hairstyles with artist chains, its chain generator had higher reported recall and precision than retrieval, and its simulated hair motion correlated more closely with artist-chain motion than retrieval or a procedural rig.
The test supports the chain-generation component on that subset. The training data and reported failures limit what can be concluded for other hair types and chain layouts.
Deep
- Original paper authors
- Yuze He, Yixiang Wei, Wang Zhao, Chuhao Chen, Yanning Zhou
- Submitted to arXiv
- License
- arXiv non-exclusive