.png){kind=icon}
.png){kind=icon}
Return
Noiseless Diffusion-GAN: Scaling-based data augmentation for generative models
.png){kind=icon}
OA
AI
DOI：10.1016/j.neunet.2025.108458.png)
.png){kind=icon}
.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
• We systematically analyzed Diffusion-GAN and found that data scaling plays an important role in stabilizing training and improving generative quality, helping to prevent mode collapse. • We proposed Scale-GAN, which combines data scaling with variance-based regularization to provide a simple alternative to gradient penalties and enhance training stability. • Our method is applicable to a wide class of GAN algorithms without loss of theoretical validity, and we provide theoretical analyses on bias-variance trade-off control, gradient direction invariance, and improved generalization performance. • Extensive experiments on popular datasets demonstrate that Scale-GAN achieves superior training stability and generative quality (FID and recall) compared to existing methods, while maintaining good computational efficiency.
Keywords:
Generative models
Data scaling
Noise injection
Generalization error bounds
07.05.Mh
62M45
AI Summary
Key information extracted from the uploaded paper, including a brief overview, abstract, background, key highlights, visual analysis, and future outlook.