HairDeck: Animation-Ready Stylized Hair Cards from a Single Image

HairDeck generates stylized hair cards and spring-bone chains from a single image or point cloud by encoding each card as one token and denoising all tokens together with a set diffusion transformer.

Stylized game and avatar hair is built from artist-placed hair cards animated by spring bones, but artists must place every card and rig every chain by hand. Existing hair generators output realistic strands, card extraction targets realistic alpha-textured cards, and image-to-3D generators output a single fused mesh with no cards to animate. CHARM, the only generator of stylized hair cards, decodes a hairstyle autoregressively as a long sequence of control points.

HairDeck describes each card by control points for position, width and thickness, plus its cross-section type and root normal; a card VAE encodes variable-length control points into a latent, and one token holds this latent with the card's root position, root normal, cross-section type and existence flag. A set diffusion transformer without positional embedding denoises all card tokens together with flow matching, conditioned on a point cloud or an image, and existence flags on empty slots decide how many cards appear. For single images, HairDeck predicts cards in a face-anchored head frame and widens training renders with synthetic style augmentation and renders restyled by an image-editing model. A second set diffusion transformer, the chain DiT, generates all spring-bone chains of a hairstyle together from the card tokens, with each chain token holding its path, joint count and stiffness. Fixed geometric skinning attaches cards to chains, producing an animatable asset that exports to downstream formats.

• On CHARM's 100 test hairstyles from point clouds, HairDeck lowers Chamfer distance and orientation error while running much faster than CHARM. §4.2
• A card-level autoregressive model using the same tokens is less accurate and slower than the set formulation, and it generates fewer cards on average. §4.2
• On single images, HairDeck has lower aligned CD and failure rates and takes less time than CHARM, image-to-3D generators and converted DiffLocks strands. §4.3
• Removing the root normal from the card token raises orientation error in the point-cloud ablation. §4.4
• On artist cards, the chain DiT recovers artist spring-bone chains better than retrieval and a procedural rig in recall, precision and motion correlation; scores are lower from generated cards and single images. §4.5

• Appendix F states failure cases on in-the-wild illustrations, where objects inside the matte turn into cards.
• Experiments leave open lower spring-bone chain quality from single images: chain CD is 0.407 and recall is 0.390, versus 0.279 and 0.625 on artist cards.
• Experiments leave open nonzero image failure rates: 3.5% on test characters and 10% on in-the-wild images.
• Chain evaluation covers only the 58 test hairstyles that have chains.

• Stylized hair cards are ribbons with rhombic or triangular cross-section whose width and thickness vary along the card, animated by spring-bone chains. §1
• Each card becomes one continuous token holding a card VAE latent, root position, root normal, cross-section type and existence flag. §1
• The set DiT without positional embedding denoises all card tokens together with flow matching, conditioned on a point cloud or an image. §3
• Empty slots and existence flags let the model decide how many cards to generate. §3
• Single-image generation uses a face-anchored head frame, synthetic style augmentation and renders restyled by an image-editing model. §1
• The chain DiT generates spring-bone chains as a second set conditioned on card tokens, and fixed geometric skinning attaches cards to them. §3.4
• The set DiT has 12 layers of width 768 with 181.3M parameters; the chain DiT has 8 layers of width 384 with 30.5M parameters. §3.5