Trade-offs in Few-shot Image Generation: Stability, Fidelity, and Adaptability in Data-Efficient GANs

Department of Statistics and Data Science, University of California, Davis, United States

^* Corresponding author: lxiqi@ucdavis.edu

Abstract

This study presents a comparative analysis of data-efficient GANs, evaluating training stability, output fidelity, and cross-domain adaptability under few-shot conditions. Our findings reveal a distinct trade- off between performance and specialization. StyleGAN2-ADA achieves strong stability and fidelity with small datasets. D3T-GAN performs best when the source and target domains share similar structures. WeditGAN and DEff-GAN improve controllability and diversity, but they need careful parameter tuning. Smoothness Regularization increases consistency, though it may reduce variation. LDM-GAN offers semantic richness and diversity, but it often oversmooths and struggles with very limited data. Model choice depends on the task. StyleGAN2-ADA is best for general small-sample fidelity. D3T-GAN is suited for domain transfer. LDM-GAN works well when diversity is the priority. Future work should prioritize hybrid frameworks, combining augmentation and transfer learning for more robust performance. Progress will also depend on better evaluation metrics and reproducible benchmarks. These steps mark the most promising path for robust and practical few-shot image generation.