Noiseless Diffusion-GAN: Scaling-based data augmentation for generative models

This paper explores stable learning methods for generative models designed to facilitate high-quality data generation. Noise injection is a commonly employed technique to enhance learning stability; however, selecting an appropriate noise distribution remains a significant challenge. Diffusion-GAN, a recently proposed approach, addresses this issue by leveraging the diffusion process alongside a timestep-dependent discriminator. In this study, we analyze Diffusion-GAN and identify data scaling as a critical factor for achieving stable learning and high-quality data generation. Based on these insights, we introduce a learning algorithm, termed Scale-GAN, which incorporates data scaling and variance-based regularization. Moreover, we provide a theoretical proof demonstrating that data scaling effectively manages the bias-variance trade-off within the estimation error bound. Experimental evaluations on standard benchmark datasets highlight the proposed method's efficacy in enhancing both stability and accuracy.