Fast-Converging Conditional Generative Adversarial Networks for Image Synthesis

Chengcheng Li, Zi Wang, Hairong Qi

Abstract

Conditional Generative Adversarial Networks (cGANs) are powerful tools for class-specific image synthesis, yet they often struggle with slow convergence and "label confusion" during early training. This paper presents FC-GAN (Fast-converging Conditional GAN), an architectural improvement over the standard Auxiliary Classifier GAN (AC-GAN). By introducing a dedicated $(N+1)$-th "fake" class to the discriminator's auxiliary classifier, FC-GAN prevents the model from misattributing synthetic noise to real data categories. Experimental results on MNIST and CIFAR-10 demonstrate that FC-GAN achieves significantly faster convergence and superior image quality compared to traditional AC-GAN architectures.

Introduction

Generative Adversarial Networks (GANs) have revolutionized image synthesis, but providing control over the generated output remains a challenge. While Conditional GANs like AC-GAN address this by using an auxiliary classifier to identify image classes, they suffer from a fundamental logical flaw: the classifier is forced to assign real-world labels to synthetic images that may still be indistinguishable from noise. This "label confusion" provides misleading gradients, slowing down the training process and limiting final image fidelity.

Proposed Architecture: FC-GAN

To resolve label confusion, FC-GAN introduces an advanced auxiliary classifier. In a standard $N$-class problem, the AC-GAN auxiliary classifier has $N$ output nodes. FC-GAN expands this to $N+1$ nodes, where the additional node is reserved exclusively for "fake" data.

FC-GAN Architecture with (N+1) class auxiliary classifier

In this setup, the training dynamics are refined:

  • Discriminator: Learns to map real images to their respective classes ($C_1 \dots C_N$) and all generated images to the new $C_{fake}$ class.
  • Generator: Competes to produce images that the discriminator classifies as both "Real" (source branch) and as a specific target class $C_i$ (auxiliary branch), effectively learning to avoid the $C_{fake}$ designation.

By providing a dedicated "trash can" class for synthetic data, the auxiliary classifier reinforces the primary real/fake detection task, leading to cleaner gradients and more stable training.

Experimental Results

The performance of FC-GAN was evaluated against AC-GAN using the MNIST and CIFAR-10 datasets. The results highlight improvements across both qualitative and quantitative dimensions.

Visual Fidelity and Convergence

Qualitative analysis shows that FC-GAN produces recognizable images much earlier in the training process. On the MNIST dataset, FC-GAN generates clean, distinct digits by epoch 10, a milestone that takes AC-GAN approximately 20 epochs to reach.

Quantitative Metrics

On the more complex CIFAR-10 dataset, FC-GAN demonstrates substantial gains in both diversity and distribution matching:

  • Inception Score (IS): FC-GAN achieved a score of 6.36, significantly outperforming AC-GAN’s 4.19.
  • Parzen Window Estimate: FC-GAN showed a better fit to the true data distribution with a score of 646.9, compared to 581.8 for AC-GAN.

Discussion and Conclusion

FC-GAN's primary advantage lies in its ability to reach Nash Equilibrium faster. By eliminating the requirement for the auxiliary classifier to guess classes for noisy synthetic data, the model stabilizes its source loss significantly earlier—around epoch 13 for MNIST compared to epoch 20 for AC-GAN.

While highly effective for low-resolution benchmarks, future research should investigate the scalability of the $(N+1)$ class strategy for high-resolution datasets like ImageNet. Additionally, further study is needed to determine if this architecture offers inherent protection against mode collapse. Nevertheless, the addition of a "fake" class represents a simple, low-cost, and powerful modification for enhancing conditional image synthesis.

Summary updated

BibTeX
@inproceedings{li2018fast,
  author = {Chengcheng Li and Zi Wang and Hairong Qi},
  title = {{Fast-Converging Conditional Generative Adversarial Networks for Image Synthesis}},
  booktitle = {ICIP},
  year = {2018},
  doi = {10.1109/icip.2018.8451161},
}
Figure