TY - JOUR
T1 - Cluster our hairstyles
T2 - A novel deep differentiable clustering algorithm for generating three-dimensional representative hairstyles
AU - Li, Pinyan
AU - Wang, Yapeng
AU - Yang, Xu
AU - Im, Sio Kei
AU - Song, Jucheng
AU - Zhang, Jie
N1 - Publisher Copyright:
© 2026 Elsevier Ltd.
PY - 2026/10/15
Y1 - 2026/10/15
N2 - Highly realistic three-dimensional hair modeling has broad applications in digital media and film production. However, existing three-dimensional hairstyle datasets have limitations such as insufficient hairstyle diversity and a lack of two-dimensional images. To address this, we develop a novel three-dimensional representative hairstyle generation framework and propose a novel deep differentiable clustering algorithm with cross-attention, which integrates cross-attention feature fusion, multi-view feature fusion, differentiable uniform manifold approximation and projection mapping, and a differentiable traditional clustering algorithm variant into a fully differentiable pipeline for parameter optimization. Unlike traditional clustering methods, our algorithm optimizes parameters through differentiable modules, achieving more stable training and effectively clustering hairstyles with similar stylistic characteristics. Additionally, we construct a multi-view two-dimensional hairstyle dataset with 300 hairstyles, leveraging the Chat Generative Pre-Training Transformer to generate prompts and images, covering a broad range of common hairstyle types, including braids, twists, and curls. Using this framework and algorithm, 18 clusters are identified and used to generate three-dimensional representative hairstyles. Comparative experiments demonstrate the effectiveness and superiority of our approach. Our three-dimensional hairstyles can be utilized in digital human, computer graphics, and film production. The three-dimensional data will be publicly released upon acceptance of the paper.
AB - Highly realistic three-dimensional hair modeling has broad applications in digital media and film production. However, existing three-dimensional hairstyle datasets have limitations such as insufficient hairstyle diversity and a lack of two-dimensional images. To address this, we develop a novel three-dimensional representative hairstyle generation framework and propose a novel deep differentiable clustering algorithm with cross-attention, which integrates cross-attention feature fusion, multi-view feature fusion, differentiable uniform manifold approximation and projection mapping, and a differentiable traditional clustering algorithm variant into a fully differentiable pipeline for parameter optimization. Unlike traditional clustering methods, our algorithm optimizes parameters through differentiable modules, achieving more stable training and effectively clustering hairstyles with similar stylistic characteristics. Additionally, we construct a multi-view two-dimensional hairstyle dataset with 300 hairstyles, leveraging the Chat Generative Pre-Training Transformer to generate prompts and images, covering a broad range of common hairstyle types, including braids, twists, and curls. Using this framework and algorithm, 18 clusters are identified and used to generate three-dimensional representative hairstyles. Comparative experiments demonstrate the effectiveness and superiority of our approach. Our three-dimensional hairstyles can be utilized in digital human, computer graphics, and film production. The three-dimensional data will be publicly released upon acceptance of the paper.
KW - Deep differentiable clustering
KW - Hairstyle clustering
KW - Three-dimensional hairstyle generation
UR - https://www.scopus.com/pages/publications/105047530443
U2 - 10.1016/j.engappai.2026.115890
DO - 10.1016/j.engappai.2026.115890
M3 - Article
AN - SCOPUS:105047530443
SN - 0952-1976
VL - 182
JO - Engineering Applications of Artificial Intelligence
JF - Engineering Applications of Artificial Intelligence
M1 - 115890
ER -