Abstract
We propose to tackle the mode collapse problem in generative adversarial networks (GANs) by using multiple discriminators and assigning a different portion of each minibatch, called microbatch, to each discriminator. We gradually change each discriminator's task from distinguishing between real and fake samples to discriminating samples coming from inside or outside its assigned microbatch by using a diversity parameter $\alpha$. The generator is then forced to promote variety in each minibatch to make the microbatch discrimination harder to achieve by each discriminator. Thus, all models in our framework benefit from having variety in the generated set to reduce their respective losses. We show evidence that our solution promotes sample diversity since early training stages on multiple datasets.
Introduction
Generative Adversarial Networks (GANs) have revolutionized generative modeling but remain susceptible to mode collapse. This phenomenon occurs when the generator learns to produce a limited set of high-quality samples that successfully deceive the discriminator, failing to capture the full diversity of the underlying data distribution. While multi-adversarial approaches have been explored, they often lack a direct mechanism to incentivize variety. microbatchGAN introduces a novel "microbatch discrimination" task to bake diversity directly into the adversarial objective.
Methodology
The core of microbatchGAN is the division of a standard minibatch of generated samples into $K$ smaller "microbatches," each assigned to one of $K$ discriminators.
Microbatch Discrimination Task
Instead of the traditional binary real/fake classification, each discriminator $D_k$ is tasked with distinguishing its assigned microbatch from a combined set of other microbatches and real data. Specifically, $D_k$ should label samples from its assigned microbatch as "fake" and all other samples (real or from other microbatches) as "real."
The Diversity Parameter ($\alpha$)
A weighting factor, $\alpha$, balances the traditional real/fake objective with the new microbatch-level distinction. To ensure the model first learns basic data features before being pressured for diversity, the authors utilize self-learned $\alpha$ functions (e.g., Sigmoid, Softsign, or Tanh). This allows $\alpha$ to grow dynamically during training, ensuring the model first learns what "realistic" looks like before being forced to diversify.
Experimental Results
The framework was evaluated on toy datasets (8 Gaussian distributions) and standard benchmarks including MNIST, CIFAR-10, and CelebA.
Toy Dataset and Qualitative Analysis
On the 2D Gaussian mixture, microbatchGAN successfully captured all eight modes, whereas standard GANs often collapsed to a single point or a few clusters. Figure 3 illustrates how the choice of $\alpha$ schedule affects the transition from learning the distribution to achieving full diversity. The Sigmoid-based schedule ($\alpha_{sigm}$) proved particularly effective at balancing quality and variety.
Quantitative Performance
On CIFAR-10, microbatchGAN achieved a Mean FID of 103.9 ($K=5$), outperforming the baseline GAN's 125.5. Furthermore, it reached an Inception Score (IS) of 6.77, surpassing other multi-adversarial methods like GMAN (6.00) and standard DCGAN (6.40) under comparable settings. The proposed Intra-FID metric confirmed that the model maintains high variety within the generated sets across different datasets.
Discussion and Conclusion
microbatchGAN demonstrates that diversity can be effectively promoted by framing it as a survival strategy for the generator within a multi-adversarial environment. By forcing discriminators to compete for their "turf" (assigned microbatches), the generator is incentivized to produce unique samples to minimize its own loss.
While the approach introduces some computational overhead due to multiple discriminators and is currently tested primarily on lower-resolution datasets (32x32), it provides a robust framework for mitigating mode collapse. Future work may involve scaling this logic to high-resolution synthesis and integrating it with more advanced backbone architectures.
Summary updated