CGHair: Compact Gaussian Hair Reconstruction with Card Clustering¶

Conference: CVPR 2026 arXiv: 2604.03716 Code: Project Page Area: 3D Vision Keywords: 3D Gaussian Splatting, hair reconstruction, hair card clustering, compact representation, appearance compression

TL;DR¶

CGHair is proposed, achieving over 200× compression of appearance parameters and 4× acceleration in strand reconstruction while maintaining comparable visual quality, via hair-card-guided hierarchical clustering and a shared Gaussian appearance codebook.

Background & Motivation¶

Background: 3DGS-based hair reconstruction methods (e.g., GaussianHair) model strands via cylindrical Gaussian chains, enabling high-fidelity real-time rendering. However, dense hair modeling requires millions of Gaussian primitives, resulting in prohibitive storage and rendering costs.

Limitations of Prior Work: Methods such as GaussianHair assign independent spherical harmonic coefficients to each strand, ignoring the strong structural and appearance redundancy inherent in a given hairstyle, leading to severe parameter redundancy.

Key Challenge: High fidelity demands a large number of independent parameters, whereas practical deployment requires compact representations.

Goal: To exploit the intrinsic structural similarity of hair and design a compact Gaussian hair representation.

Key Insight: Drawing inspiration from the "hair card" concept widely used in the game and film industries—clustering similar strands into cards and sharing textures among them.

Core Idea: Hierarchical clustering combined with a shared appearance codebook, simultaneously achieving strand-level fidelity and extreme compression.

Method¶

Overall Architecture¶

A four-stage pipeline: (1) efficient strand reconstruction with generative priors; (2) hair card generation (strand clustering + card geometry + texture mapping); (3) compact Gaussian representation via card clustering (shared appearance codebook); (4) multi-view appearance optimization.

Key Designs¶

Generative-Prior-Accelerated Strand Reconstruction: The PERM parametric hair geometry model is adopted, representing hair in scalp UV space as a guide map $\mathbf{G} = \mathcal{D}_{guide}(\alpha)$ and a style map $\mathbf{S} = \mathcal{D}_{style}(\beta)$, decoded via pretrained StyleGAN2 and VAE decoders. The hairstyle encodings $\alpha, \beta$ and the decoders are jointly optimized, with photometric supervision applied via cylindrical Gaussians along strands. Design Motivation: Directly optimizing latent textures is inefficient and requires expensive diffusion regularization. Leveraging PERM's strong geometric priors enables a 4× speedup (vs. GaussianHaircut) and is fully automatic (vs. GaussianHair which requires manual intervention).
Hierarchical Hair Card Generation:
Strand Clustering: Each strand's 3D points are concatenated into a vector, and k-means clustering groups them into $N_c$ clusters (800 for real hair, 400 for synthetic).
Card Geometry Construction: Using the cluster-center strand as the axis, normal vectors are solved via constrained optimization: $\{n_k^*\} = \arg\min \sum_k \sum_{p_i} \|(p_i - \bar{p}_k) \cdot n_k\|$, ensuring all strands lie within the card plane.
Texture Generation: 3D strands are projected onto the card surface by optimizing UV coordinates and normal displacements $\delta_i$, producing anti-aliased texture maps. Design Motivation: Hair cards are the industry-standard simplification for hair, directly encoding geometric and appearance redundancy.
Shared Gaussian Appearance Codebook: The $N_c$ cards are further clustered by texture features into $N_T=64$ groups, each sharing an appearance codebook of $K=10$ entries ($D=64$ dimensions). Each strand selects codebook entries via Gumbel-Softmax soft indexing: $\mathbf{F}_{strand} = \sum_{k=1}^K \pi_k \cdot \mathbf{F}_k$ A globally shared MLP decoder $\phi_{dec}$ then decodes the low-dimensional features into per-Gaussian spherical harmonic coefficients: $[\mathbf{SH}_1, ..., \mathbf{SH}_{|\mathcal{S}|}] = \phi_{dec}(\mathbf{F}_{strand})$ Design Motivation: Directly mapping codebook entries to SH coefficients for all Gaussians requires 4k+ dimensions. Extreme compression is achieved through a low-dimensional latent space combined with a shared decoder.

Loss & Training¶

$$\mathcal{L} = \mathcal{L}_p + \mathcal{L}_a + \mathcal{L}_o$$ - $\mathcal{L}_p$: L1 + D-SSIM photometric loss - $\mathcal{L}_a$: alpha mask supervision loss - $\mathcal{L}_o$: opacity smoothness regularization - Geometric parameters are frozen for the first 7k iterations; thereafter, joint optimization proceeds with a reduced learning rate.

Key Experimental Results¶

Main Results — Strand Reconstruction Quality¶

Metric	CGHair	$\mathcal{D}_{PCA}$ only	Frozen $\alpha,\beta$	$\mathcal{D}_{style}$ only
Pos. Error ↓	0.148	0.162	0.159	0.182
Cur. Error ↓	6.73	9.12	6.98	7.48

Ablation Study — Compact Representation¶

Method	PSNR↑	SSIM↑	LPIPS↓	Size(MB)↓	Compression↑
Unique (per-strand SH)	33.99	0.982	0.019	163.7	1.0
w/o card clustering	28.00	0.938	0.042	22.42	7.30
Single-SH	30.48	0.958	0.037	0.46	355.9
Latents (full)	32.40	0.970	0.026	0.71	230.6

Key Findings¶

The full CGHair achieves PSNR 32.40 with only 0.71 MB of appearance parameters, yielding a 230× compression ratio compared to per-strand SH (163.7 MB) with only a 1.59 dB PSNR drop.
Card clustering is critical: without clustering, globally shared features yield only PSNR 28.00, demonstrating the necessity of card-level structured clustering.
A codebook size of $K=10$ is sufficient; increasing to 90 yields only marginal PSNR improvement at the cost of significantly increased storage.
Latent dimension $D=64$ represents the optimal trade-off.

Highlights & Insights¶

The well-established industrial concept of "hair cards" is elegantly introduced into academic hair reconstruction, unifying structural priors with data-driven reconstruction.
A hierarchical compression strategy is employed: strands → cards → card groups → shared codebook, exploiting redundancy at different granularities.
The design is plug-and-play: the CGHair module can be directly applied on top of reconstructed strands from GaussianHair as a post-processing compression stage.
The Gumbel-Softmax soft indexing enables end-to-end training, avoiding the codebook collapse commonly observed in VQ-based methods.

Limitations & Future Work¶

The cluster counts $N_c$ and $N_T$ are set manually; adaptive determination could be considered.
The method currently supports only static hair and does not handle dynamic or motion scenarios.
The compressed PSNR remains ~1.6 dB below the uncompressed baseline, which may be insufficient for applications demanding extreme detail fidelity.
The method relies on PERM pretrained priors; generalization to unconventional hairstyles has not been thoroughly validated.

General-purpose 3DGS compression methods such as CompGS do not exploit the structural redundancy specific to hair; CGHair serves as a paradigm for domain-specialized compression.
The quantized encoding concept from VQVAE is successfully transferred to 3D hair representation.
The approach offers inspiration for other 3D reconstruction tasks with strong structural redundancy, such as cloth and vegetation.

Rating¶

Novelty: ⭐⭐⭐⭐ The combination of card clustering and shared codebook is novel, though individual technical components are relatively mature.
Experimental Thoroughness: ⭐⭐⭐⭐⭐ Multi-granularity ablations are comprehensive.
Writing Quality: ⭐⭐⭐⭐ The pipeline is clearly presented with rich illustrations.
Value: ⭐⭐⭐⭐ A 200× compression ratio carries significant implications for practical deployment.