We present 3DGH, an unconditional generative model for 3D human heads with composable hair and face components. Unlike previous work that entangles the modeling of hair and face, we propose to separate them using a novel data representation with template-based 3D Gaussian Splatting, in which deformable hair geometry is introduced to capture the geometric variations across different hairstyles. Based on this data representation, we design a 3D GAN-based architecture with dual generators and employ a cross-attention mechanism to model the inherent correlation between hair and face. The model is trained on synthetic renderings using carefully designed objectives to stabilize training and facilitate hair-face separation. We conduct extensive experiments to validate the design choice of 3DGH, and evaluate it both qualitatively and quantitatively by comparing with several state-of-the-art 3D GAN methods, demonstrating its effectiveness in unconditional full-head image synthesis and composable 3D hairstyle editing.
Here we show 4 samples from our model with full-head rendering.
Here we show 3D hairstyle editing examples of our method. We achieve this by adjusting the intermediate latent codes for hair and face. From left to right: the original example, the reference hairstyle, and the edited (composed) result.
Original
Reference Hair
Edited Result
As we explicitly model the hair-face correlation, we can control the influence of face on the composed hairstyles by interpolating the CFG scale factor \(\omega\) introduced in our method. Here we show an example on the effect of hair-face correlation. From left to right: the reference face, the reference hair, intermediate composed results generated by interpolating \(\omega\) from 0 to 1. When \(\omega\) is small, the influence of hair-face correlation is relatively weak, leading to composition results where the hairstyle looks more similar to the reference. With the increase of \(\omega\), the influence of hair-face correlation gradually becomes stronger, thus causing the composed hairstyle to change towards a direction more plausible for the given face (in this case, becomes shorter).
Reference Face
Reference Hair
Interpolation of \(\omega\)
@article{he2025head,
title={3DGH: 3D Head Generation with Composable Hair and Face},
author={He, Chengan and Li, Junxuan and Kirschstein, Tobias and Sevastopolsky, Artem and Saito, Shunsuke and Tan, Qingyang and Romero, Javier and Cao, Chen and Rushmeier, Holly and Nam, Giljoo},
journal={ACM Transactions on Graphics},
volume={44},
number={4},
pages={1--12},
year={2025}
}