Introduction
Generative Adversarial Networks (GANs) like StyleGAN2 and BigGAN have set the standard for high-resolution image synthesis. However, their massive parameter counts and high computational requirements (MACs) pose significant challenges for deployment on edge devices. DGL-GAN introduces a streamlined approach to GAN compression by leveraging a pretrained teacher discriminator to provide adversarial supervision, bypassing the pitfalls of traditional generator-to-generator distillation.
The Challenge of GAN Compression
Compressing noise-to-image GANs is difficult due to three primary factors:
- Sensitivity: Pruning generators often leads to immediate quality degradation.
- Lack of Ground Truth: Unlike image-to-image tasks, there is no single "correct" output for a given noise vector, making pixel-wise imitation ineffective.
- Training Instability: The adversarial nature of GANs makes adding complex teacher-student constraints prone to collapse.
Discriminator Guided Learning (DGL)
The core innovation of DGL-GAN is "Discriminator-to-Generator" guidance. Instead of forcing a student generator to match a teacher generator's features, the student is trained against both its own discriminator and a fixed, pretrained teacher discriminator. This teacher discriminator acts as a proxy for the real image distribution, providing a more stable and informative gradient signal.
The training follows a two-stage process:
- Stage I: The narrow student GAN is trained from scratch.
- Stage II: The student is fine-tuned with additional guidance from the teacher discriminator.
Key Results
DGL-GAN achieves state-of-the-art results across multiple benchmarks, including FFHQ, LSUN-church, and ImageNet.
As shown in the figure above, DGL-GAN maintains lower FID scores (better quality) at significantly lower computational costs compared to existing methods. For StyleGAN2 on FFHQ, the model was compressed to 1/3 of its original parameters while maintaining an FID of 3.37, very close to the original 2.84.
Qualitative Quality
Beyond quantitative metrics, DGL-GAN preserves the semantic richness of the latent space. Even at 1/4 width, the model supports smooth latent space editing for attributes like age and gender while maintaining high visual fidelity.
Conclusion
DGL-GAN demonstrates that the discriminator is a superior teacher for GAN compression. By focusing on distribution alignment rather than feature imitation, the method achieves significant speedups and parameter reductions without sacrificing image quality. Furthermore, applying this guidance to uncompressed models has even pushed StyleGAN2 to new state-of-the-art performance levels, achieving an FID of 2.65 on FFHQ.
Summary updated