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

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

Creating hair for a game character or avatar is not only a modeling problem. It is also a representation and rigging problem. Stylized hair is often built from dozens of visible ribbons called hair cards, with each card shaped and oriented by an artist. Spring-bone chains then provide the secondary motion that makes the hairstyle sway when the character moves. HairDeck, from Yuze He, Yixiang Wei, Wang Zhao, Chuhao Chen, Yanning Zhou, and Yong-Jin Liu, targets this production format directly from a single image.

The authors argue that existing approaches usually produce the wrong kind of asset. Strand-based systems model realistic hair as many individual strands, while general image-to-3D systems commonly return a fused surface. Those outputs may look plausible, but they do not give an artist a collection of stylized cards that can be edited and animated. CHARM, the prior system designed for stylized hair cards, generates control points autoregressively, one token at a time. HairDeck instead treats a hairstyle as an unordered set of artist-defined primitives.

One token per card

HairDeck describes each card using its control points, including position, width, and thickness. It also preserves two details that the authors say are important for visual fidelity: the card’s cross-section type and its root normal, which determines the width direction at the card’s root. A card may have a rhombic or triangular cross-section, and its orientation is propagated along the curve with a rotation-minimizing frame.

Because cards can contain different numbers of control points, the system first trains a card variational autoencoder. The encoder compresses a variable-length card into a continuous latent representation, while the decoder reconstructs the card by evaluating it at arc-length positions. This produces a fixed-size token for each card rather than forcing the generator to emit hundreds or thousands of point-level tokens.

The hairstyle generator is a set diffusion transformer. It has no positional embedding for the card slots, so it does not impose an artificial order on the cards. The model denoises all card tokens together and uses an existence flag to decide how many cards should be present. The authors describe this as a general recipe for assets made from unordered primitives: represent each primitive with one token, then generate the collection jointly.

From image to animation-ready asset

For single-image input, HairDeck uses image features from a frozen vision encoder and predicts the cards in a head frame anchored to the face. This is intended to solve a practical ambiguity: a dataset-normalized character frame does not reliably indicate where the head sits or how large it is. The authors also augment training renders with synthetic visual styles and restyled images, aiming to make the model less dependent on a particular rendering appearance.

The system adds a second set transformer for spring-bone chains. Each chain is represented by its path, joint count, stiffness, and existence flag. The chain model is conditioned on the generated card tokens, after which fixed geometric skinning attaches cards to nearby chains. The resulting geometry and rig can be exported to downstream formats, including VRM-oriented workflows described by the authors. This is a significant practical distinction from methods that generate only a static surface.

Reported results

On the 100-hairstyle test set from CHARM, the authors report that HairDeck reduces Chamfer distance from CHARM’s 0.0117 to 0.0066. Its card-orientation error is 16.8 degrees, compared with 40.0 degrees for CHARM, and its reported generation time is 1.01 seconds per hairstyle rather than 33.7 seconds. The paper summarizes this as a 44% reduction in Chamfer distance, more than a halving of orientation error, and a 33-fold speed improvement.

The comparison with an autoregressive model using the same card representation is especially informative. That card-level model reaches a Chamfer distance of 0.0078 and takes 6.91 seconds. Generating the same kind of tokens as an unordered set lowers the reported distance to 0.0066 and is 6.8 times faster. This supports the authors’ claim that both the representation and the generation order matter.

For single images, HairDeck is evaluated on 76 rendered test characters and 40 in-the-wild illustrations. On the rendered characters, it reports an aligned Chamfer distance of 0.0089, an IoU of 0.725, a 3.5% failure rate, and 1.33 seconds per image. The competing methods in the paper report lower performance on these measures, although the comparisons involve different output representations: some produce fused meshes and another converts generated strands into cards.

The spring-bone results are more mixed, but still useful. When conditioned on artist cards, the chain model reaches a motion correlation of 0.702 against the artists’ motion, compared with 0.622 for retrieval and 0.538 for a procedural rig. With generated cards, chain recall falls to 0.542 for point-cloud input and 0.390 for single-image input, while the reported fit of the chains to the cards remains about 0.95. This suggests that the generated rig can attach cleanly even when its layout does not fully reproduce an artist’s rig.

What the ablations reveal

The ablation studies identify several design choices that contribute to the result. Removing the root normal increases orientation error from 16.8 to 26.4 degrees. Arc-length decoding produces smoother cards than index-based decoding, while raw point-control tokens result in visibly more jagged geometry. The head frame substantially improves unaligned placement and overlap with the input hair. Restyled training images improve root precision on restyled inputs, indicating that appearance variation helps the model distinguish hair structure from drawing style.

Limitations and practical meaning

HairDeck is trained on stylized VRoid hairstyles, so the authors explicitly caution that very short hair and realistic alpha-textured cards are outside its training distribution. The system generates card geometry and UV coordinates but not textures. It can also mistake visible accessories, such as bows, hats, horns, or effects retained by an imperfect image matte, for hair cards. Unusual spring-bone layouts, including hairstyles with many chains radiating from the crown, are another known weakness.

For developers and technical artists, the demonstrated value is not simply faster 3D reconstruction. HairDeck aims at a usable intermediate asset: editable stylized cards plus a plausible secondary-motion rig. That could make single-image references useful for prototyping avatar hairstyles, preparing game assets, or generating starting points for artists. Whether it can replace manual cleanup remains unproven. The benchmark is focused on stylized VRoid data, and the paper does not establish equivalent performance across broader character designs, production textures, or complex collision setups. Still, its set-based formulation offers a compelling direction for other structured assets assembled from unordered artist primitives.


Sources

AI-assisted research summary. Check the original paper for complete methods, results, and limitations.

Avatar photo

Tech News

Jeff is a passionate blog writer who shares clear, practical insights on technology, digital trends and AI industries. With a focus on simplicity and real-world experience, his writing helps readers understand complex topics in an accessible way. Through his blog, Jeff aims to inform, educate, and inspire curiosity, always valuing clarity, reliability, and continuous learning.

↑ ↓ navigate Enter open Esc close