Neural 3D Face Stylization

01 — Overview

Overview

Traditional deformation transfer can preserve a person’s facial characteristics in a stylized template, but it is slow and must be optimized again for every new face. This work learns the transfer once and applies it directly to new inputs.

Only one style template is needed for each target look, reducing artist effort and avoiding paired realistic-to-stylized training meshes.

02 — Contributions

Key Contributions

• 01

  Frames 3D face shape stylization as a fast learned deformation-transfer problem.

• 02

  Uses weak supervision so paired source and stylized training data are not required.

• 03

  Introduces template-guided mesh smoothing to preserve the intended structure of each style.

03 — Method

Method

A neural network predicts the deformation from a realistic input face to a chosen style template. Training uses weak supervision and a template-guided mesh-smoothing regularizer that discourages structural artifacts while retaining identity-related facial shape.

04 — Evaluation

Results

The paper reports stylization quality comparable to conventional deformation transfer at roughly 3,000 times the processing speed, with an average Chamfer distance of about 0.01 mm.

05 — Reference

Citation

@Article{Yan_2025_FaceStylization,
  author  = {Yan, Peizhi and Ward, Rabab K. and Tang, Qiang and Du, Shan},
  title   = {Neural 3D Face Shape Stylization Based on Single Style Template via Weakly Supervised Learning},
  journal = {IEEE Transactions on Visualization and Computer Graphics},
  year    = {2025},
  volume  = {31},
  number  = {10},
  pages   = {9522--9529},
  doi     = {10.1109/TVCG.2025.3573690}
}