HairDeck: single image to rigged game hair cards, 33x faster than CHARM at 44% lower Chamfer error

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

Yuze He, Yixiang Wei, Wang Zhao, Chuhao Chen, Yanning Zhou, Yong-Jin Liu

cs.CV, cs.GR

2026-10-30

HairDeck makes each hair card one token and denoises the set, plus its spring-bone chains, in one pass: 44% lower Chamfer error than CHARM and 33x faster.

What problem this solves

Stylized characters in games and avatars wear hair built from hair cards: a few dozen to a hundred-plus ribbons, each placed by an artist, whose geometry is the visible hair. Spring-bone chains, short bone chains hung from the head and simulated every frame, make the hair sway; Unity's Dynamic Bone, VRChat's PhysBones and UE's Kawaii Physics all ship this. Every card and every chain is placed by hand.

No existing generator produces this asset. Strand-based hair methods, including recent ones that accept anime images, output strands, the representation of realistic hair. Image-to-3D generators output a single fused mesh with nothing to rig. CHARM, the only stylized-hair-card generator, decodes a hairstyle autoregressively as a sequence of hundreds to thousands of discretized control points: slow, and burdened by an order the data never had.

Method

The organizing observation: a card hairstyle is an unordered set of artist-defined primitives, so generate it as a set. One card, one token; everything denoised together.

• Card VAE. Cards have variable numbers of control points (position, width, thickness along a polyline), so a VAE compresses each card into a 32-d continuous latent. The token adds the root position, the root normal (the width direction at the root, read from the artist mesh rather than estimated), the cross-section type (rhombic or triangular) and an existence flag: 40 dims per card, 192 slots per hairstyle, about 70 cards on average. Decoding samples uniformly by arc length, and the output is as smooth as the artists' cards (median turn angle 14.2° vs 14.1°).
• Set DiT. A diffusion transformer with no positional embedding, hence equivariant to permutations of the slots; no order among cards is ever defined. Trained with flow matching, sampled in 50 Euler steps. The existence flag settles together with the shapes, so the card count is not fixed in advance. The condition is a point cloud (CHARM's benchmark input, encoded by Dora-VAE) or an image (frozen DINOv3 features, with SAM 3 masks for in-the-wild images).
• Head frame. The dataset normalizes characters to unit height, which scatters head size and position. Predicting in a frame anchored to the face (face width scaled to the training median, face top moved to a common point) fixes where the hair sits without touching its shape.
• Style coverage. Training renders get halftone, XDoG, flat-color and painterly filters, plus restyles into gacha art, thick paint and semi-realistic painting by an image-editing model.
• Chain DiT. A second set DiT, conditioned on the card tokens, generates up to 48 spring-bone chains (path, joint count, stiffness) in one pass. A fixed geometric skinning binds cards to chains and the VRM spring-bone update simulates them, so the output exports ready to use.

The ablation pins down each choice. With the same tokens generated autoregressively, card by card, CD rises from 0.0066 to 0.0078 and time from 1.01 s to 6.91 s, and the AR variant stops early at 60.8 cards on average against 68.9 for the set model and 67.4 ground truth. One token per card already delivers most of the accuracy; the set formulation adds the rest of it and all of the speed.

Results

Point clouds, CHARM's 100 test hairstyles, its protocol:

CD down 44%, orientation error cut by more than half, 33× faster. Retrieval reaches 0.0119, so CHARM's accuracy amounts to recalling the closest training hairstyle. A 54.6M-parameter model, smaller than CHARM's 61.2M, still reaches 0.0072.

Single images, 76 rendered test characters and 40 in-the-wild illustrations:

Two ablations carry the design: without the root normal, orientation error rises from 16.8° to 26.4°; without the head frame, pre-alignment CD rises from 0.0136 to 0.0189. Spring bones, on the 58 test hairstyles that have them: on artist cards, the chain DiT recovers the artists' chains at 0.625 recall and 0.585 precision, against 0.289 / 0.285 for retrieval, and hair motion correlates 0.702 with the artists' rigs against 0.538 for a procedural rig. On generated cards, recall drops to 0.542 from point clouds and 0.390 from single images.

Why it matters

For anyone generating game and avatar assets, this is the first pipeline that takes a single image and returns a card hairstyle with spring bones attached, simulated under the VRM spec and exportable to engines, at 1.33 s per image, fast enough for interactive use. The work comes from Tsinghua, CMU and CUHK, with code and a project page released. The broader lesson transfers: assets assembled from artist-defined primitives, garment panels or sketch strokes among them, suit the one-primitive-one-token set formulation. Dropping the sequence assumption buys accuracy and speed in equal measure, and the margins are large.

Limitations

• Orientation error is 16.8° against 7.5° for artist meshes rebuilt from root normals, a twofold gap.
• Chain generation degrades on the image path: recall falls from 0.625 on artist cards to 0.390 on image-generated cards, so rigging quality there is the weak link.
• The 10% failure rate on in-the-wild images comes from a 40-image set, so the error bars are wide.
• Training and evaluation are entirely VRoid characters (CHARM's dataset, 36,440 training hairstyles, 100 test); other stylized distributions and other engines' card pipelines go untested.
• Skinning is a fixed geometric rule with two hand-tuned thresholds (0.4 face widths, lower 60% of the centerline) and would need retuning per asset type.
• Style generalization rests on synthetic filters and restyled renders; behavior under extreme styles is shown qualitatively only.

Terms

• Hair card: an artist-placed ribbon whose geometry is the visible hair; a few dozen overlap to form a stylized hairstyle.
• Spring bone: a short chain of bones hung from the head, its tip simulated every frame so the hair sways and settles.
• Flow matching: a diffusion objective that learns the velocity along straight paths between data and noise.
• Permutation equivariance: with no positional embedding, the network's output permutes with its input, a natural fit for unordered sets.
• Chamfer distance: the symmetric mean of nearest-neighbor distances between two point sets, a shape-similarity metric.
• CHARM: the prior stylized-hair-card generator, autoregressive over control points, and the benchmark used here.

Source

• arXiv 2610.40599v1
• PDF
• Hugging Face
• GitHub · https://github.com/hyz317/HairDeck

Related papers

• AskStream: proactive streaming models fail to ask, 7.22% Ask-F1 vs 52.18% for offline MLLMs
• PhysInsert hits 63% text-guided DIoU while adapting pose and lighting
• IntReal: a 100K real-world dataset and 7-dimension benchmark for interleaved text-image generation
• One Shared Encoder, One Schema: Tencent's 5B Youtu-Parsing-Omni Sets OmniDocBench SOTA at 96.96
• Only low-frequency tokens go global: SpecSlot cuts pixel-space image generation cost 38%
• Storing VAE latents inside 3D Gaussians gets single-image world generation to 9 FPS

All paper explainers