Abstract
Generative Adversarial Networks (GANs) traditionally rely on a direct competition between a generator and a discriminator, a process often plagued by training instability and mode collapse. This paper proposes a novel training paradigm using indirect adversarial losses. By introducing an intermediate distribution as a fixed target, the proposed method provides a stable "anchor" for both models. Utilizing Kernelized Stein Discrepancy (KSD) to measure distances without the need for target-side sampling, this approach significantly reduces training noise and improves convergence. Experimental results on benchmarks such as CIFAR-10 and FFHQ demonstrate that this method consistently outperforms traditional adversarial frameworks in both image quality and diversity.
Introduction
Training GANs is notoriously difficult, often described as a high-stakes game of cat-and-mouse. In the standard framework, the Discriminator ($D$) attempts to distinguish real samples from fake ones, while the Generator ($G$) tries to bridge this gap. However, this direct competition frequently leads to the "attractive problem": if the loss function is not perfectly tuned, gradients can become uninformative, or the model can fall into local minima, resulting in mode collapse.
Previous attempts to mitigate these issues, such as repulsive MMD-GANs, sought to push fake samples away from real ones. However, these methods often suffered from inconsistent learning directions where $G$ and $D$ did not complement each other effectively. Furthermore, many existing methods rely on sampling from distributions, which introduces unwanted stochastic noise into the training loop.
Methodology: The Indirect Adversarial Process
The core innovation of this work is the shift from direct competition to an indirect adversarial process. Instead of the generator chasing the real data distribution directly, both the generator and discriminator interact with a known, fixed intermediate distribution (such as a standard Normal or Uniform distribution).
The Anchor Effect
As illustrated in the figure above, the training is divided into two distinct phases:
- The Discriminator Step: Real samples are pushed toward the intermediate distribution, while fake samples are pushed away.
- The Generator Step: Fake samples are pulled toward the intermediate distribution.
This creates a "dynamic balance" where the intermediate distribution acts as a stable anchor, providing consistent gradients for the generator even when the discriminator is still in the early stages of learning.
Kernelized Stein Discrepancy (KSD)
To measure the distance between the GAN's outputs and this intermediate target, the researchers employ Kernelized Stein Discrepancy (KSD). Unlike the popular Maximum Mean Discrepancy (MMD), KSD is "sample-free" on the target side. This means the loss can be calculated without drawing actual samples from the intermediate distribution, eliminating a major source of training noise and providing a smoother optimization signal.
Experimental Results
The proposed KSD-GAN was evaluated across several standard benchmarks, including CIFAR-10, Tiny-ImageNet, and high-resolution datasets like CelebA and FFHQ.
Quantitative Performance
KSD-GAN consistently achieved superior Fréchet Inception Distance (FID) scores compared to baseline models:
- CIFAR-10: Using a StyleGAN2 backbone, KSD-GAN reached an FID of 8.50, significantly outperforming the Vanilla GAN (12.07) and Repulsive MMD-GAN (19.53).
- FFHQ: On high-resolution human faces, KSD-GAN achieved an FID of 4.82, surpassing the Wasserstein GAN (7.41).
Diversity and Stability
Beyond raw quality, the method maintains high image diversity. Using Learned Perceptual Image Patch Similarity (LPIPS), the authors demonstrated that KSD-GAN avoids the mode collapse common in traditional repulsive methods, particularly when dealing with large, diverse categories.
Discussion and Future Work
The choice of the intermediate distribution is critical. The authors found that Uniform distributions generally yield better image quality (FID) than Normal distributions, likely because they provide a broader feature space for the generator to explore.
While highly effective, the method does introduce some computational overhead due to the KSD calculation, which involves gradients of the score function. Additionally, like most kernel-based methods, performance remains sensitive to hyperparameter choices such as the RBF kernel bandwidth. Future research may explore the use of more complex, learned intermediate distributions or the application of indirect losses to domain transfer and conditional generation tasks.
Summary updated