Applying Progressive Frequency Bands to Improve Image Quality and Training Stability of GAN
Generative Adversarial Networks (GANs) have revolutionized the field of image generations, yet their training instability remains a critical challenge that limits their practical applications. This paper introduces Progressive Frequency Band GAN (PFB-GAN), a novel framework that fundamentally reimag...
Saved in:
| Main Authors: | , |
|---|---|
| Format: | Article |
| Language: | English |
| Published: |
IEEE
2025-01-01
|
| Series: | IEEE Access |
| Subjects: | |
| Online Access: | https://ieeexplore.ieee.org/document/11071695/ |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| Summary: | Generative Adversarial Networks (GANs) have revolutionized the field of image generations, yet their training instability remains a critical challenge that limits their practical applications. This paper introduces Progressive Frequency Band GAN (PFB-GAN), a novel framework that fundamentally reimagines GAN training through the lens of frequency domain analysis. Unlike conventional approaches that focus on time domain stabilization, our method leverages the inherent structure of frequency components to enable systematic and stable training. By introducing a progressive learning strategy that gradually incorporates frequency bands from low to high, PFB-GAN achieves remarkable stability while preserving fine details that are often lost in existing methods. Our comprehensive experiments across various datasets demonstrate consistent improvements, with performance metrics showing significant enhancements ranging from 10.55% to 21.03% across FID, Inception Score, Density & Coverage, and Precision & Recall metrics. More importantly, PFB-GAN shows exceptional resilience under extreme learning conditions, maintaining stability even at high learning rates where conventional GANs fail completely. This work not only advances the theoretical understanding of GAN training dynamics but also provides a practical solution for developing more reliable and robust generative models. |
|---|---|
| ISSN: | 2169-3536 |