Perm: A Parametric Representation for Multi-Style 3D Hair Modeling
Chengan He, Xin Sun, Zhixin Shu, Fujun Luan, S\"oren Pirk, Jorge Alejandro Amador Herrera, Dominik L. Michels, Tuanfeng Y. Wang, Meng Zhang, Holly Rushmeier, Yi Zhou
TL;DR
Perm introduces a novel parametric 3D hair model that disentangles global structure and local curl patterns using PCA in the frequency domain, enabling precise editing and versatile applications.
Contribution
It proposes a new PCA-based strand representation for 3D hair, improving control and editing capabilities over previous joint models.
Findings
Effective disentanglement of hair structure and curl patterns.
Superior performance in hair reconstruction and editing tasks.
Versatile application as a generic prior for various hair-related problems.
Abstract
We present Perm, a learned parametric representation of human 3D hair designed to facilitate various hair-related applications. Unlike previous work that jointly models the global hair structure and local curl patterns, we propose to disentangle them using a PCA-based strand representation in the frequency domain, thereby allowing more precise editing and output control. Specifically, we leverage our strand representation to fit and decompose hair geometry textures into low- to high-frequency hair structures, termed guide textures and residual textures, respectively. These decomposed textures are later parameterized with different generative models, emulating common stages in the hair grooming process. We conduct extensive experiments to validate the architecture design of Perm, and finally deploy the trained model as a generic prior to solve task-agnostic problems, further showcasing…
Peer Reviews
Decision·ICLR 2025 Spotlight
The strengths of the paper include its use of a PCA-based strand representation for 3D hair that efficiently disentangles global and local hair details, enabling precise control and intuitive editing. The PERM model’s architecture is computationally lightweight and achieves high fidelity with reduced memory and training requirements. It showcases versatility by performing well across various applications like single-view reconstruction and hairstyle editing, often matching or surpassing speciali
While the motivation of this paper is clear and the introduction is enjoyable to read, I am extremely frustrated with the forward-referencing (or no-referencing) writing style in Section 3. I often have to hold many questions until much later in the text, making it so easy to lose track of the mathematical flow. For example, Figure 4 starts with StyleGAN, but it is not properly introduced until Section 3.2. The same issue occurs with guide textures, residual textures, and many other components.
- The frequency domain PCA for hair strands is neat. Also, 10 coeffs for coarse strands and 54 coeffs for detailed curls is reasonable and straightforward. - The proposed framework is meticulously designed. The model can be applied to differentiable optimization/rendering. I did not find an obvious fallback of the proposed framework. - The evaluation of the paper is comprehensive, including comparisons with other methods, ablation studies on different key components, and exemplar downstream ta
I am not an expert in hair modeling and I did not find obvious fallbacks of the proposed framework. Below I listed the questions raised after reading the manuscript: - From single view reconstruction results in Fig. 10, I find that the current SOTA is a good fit for the hair in the input image. However, the hair details are still misaligned. Is this restricted by model capacity given the current parameterization? (I understand hair modeling is very difficult). Are there any suggestions to impro
The paper focuses on the hot topic of 3D hair parametric modeling and innovatively proposes a strand representation method based on Principal Component Analysis (PCA) in the frequency domain. The paper is clearly articulated, with a rigorous technical approach and strong innovation. The rendering results presented in the experimental stage effectively demonstrate the effectiveness of the method. The authors have conducted extensive and in-depth experimental validations. At the strand represe
The paper only compared Groom-Gen and presented a limited number of samples, making it difficult to analyze the strengths and weaknesses of the two approaches comprehensively. To more accurately evaluate the methods, it is suggested that the number of samples in the comparison experiments be increased. Furthermore, the paper did not discuss any failure cases of the proposed method. I am particularly interested in the failure cases of the method in random synthesis and single-view reconstruction
Code & Models
Videos
Taxonomy
Topics3D Shape Modeling and Analysis · Generative Adversarial Networks and Image Synthesis · Computer Graphics and Visualization Techniques