Few-Shot Face Stylization via GAN Prior Distillation

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Recent advancements in face stylization are hindered by significant performance degradation when training data is limited.
• Existing few-shot learning approaches for face stylization often fail to meet the stringent requirements of very small datasets (less than 10 samples) or yield suboptimal outcomes.

Purpose of the Study:

• To propose GAN Prior Distillation (GPD), an effective method for few-shot face stylization.
• To enable robust face stylization even with extremely limited target domain data.

Main Methods:

• Developed a two-model system: a teacher network with Generative Adversarial Network (GAN) Prior and a student network for end-to-end translation.
• Adapted a large-scale pre-trained teacher network to a target domain using minimal samples.
• Implemented few-shot data augmentation by generating source and target domain images from shared latent codes.
• Introduced an anchor-based knowledge distillation module to transfer knowledge from the teacher to the student network, leveraging differences in augmented data.

Main Results:

• The trained student network demonstrated excellent generalization capabilities by effectively absorbing distilled knowledge.
• Qualitative and quantitative experiments confirmed superior performance compared to state-of-the-art methods in few-shot face stylization scenarios.
• Achieved high-quality stylization results even with a very small number of training samples.

Conclusions:

• GAN Prior Distillation (GPD) provides a powerful solution for few-shot face stylization.
• The proposed knowledge distillation approach effectively transfers knowledge, enabling robust performance with limited data.
• GPD significantly outperforms existing methods in scenarios requiring few-shot learning for face image translation.