MCGAN: Enhancing GAN Training with Regression-Based Generator Loss

Baoren Xiao, Hao Ni, Weixin YangAAAI 2025 · DOI

Generative Adversarial Networks (GANs) are powerful tools for data generation but are notoriously difficult to train due to the unstable "cat-and-mouse" game between the generator and the discriminator. In the paper "MCGAN: Enhancing GAN Training with Regression-Based Generator Loss," researchers from UCL and Oxford introduce a novel approach to stabilize this process by reframing the generator's objective as a regression task.

The Challenge of Adversarial Training

The standard GAN framework relies on a min-max game where the generator and discriminator compete. This often leads to undamped oscillations, where the model never settles, or gradient vanishing, where the generator stops learning because the discriminator is too effective. These issues stem from a lack of direct supervision for the generator, which must rely on potentially erratic gradients from the discriminator.

The MCGAN Solution: Regression-Based Loss

The core innovation of MCGAN (Monte Carlo GAN) is the Regression Loss ($L_R$). Instead of simply trying to "fool" the discriminator, the generator is trained to minimize the mean squared error between the discriminator's output on real data and its expected output on fake data.

Mathematically, the loss is defined as: $$L_R(\theta; \phi) := \mathbb{E}_{(x,y)\sim\mu} \left[ |D^\phi(x) - \mathbb{E}_{\hat{x}\sim\nu_\theta(y)}[D^\phi(\hat{x})]| ^2 \right]$$

By targeting an expected value estimated via Monte Carlo sampling, the generator receives a significantly more stable training signal. This approach allows the generator to reach its optimum even with a "weak" discriminator, provided it can distinguish the sign of the difference between distributions.

Training dynamics on the Dirac-GAN toy example. MCGAN (d) shows direct convergence to the equilibrium, while other methods (a-c) exhibit limit cycles or slow spiraling.

Experimental Evaluation

The authors evaluated MCGAN across diverse domains, including high-resolution images (FFHQ, LSUN), video prediction (Moving MNIST), and high-dimensional time series (S&P 500).

Key Results

The empirical findings demonstrate that MCGAN consistently outperforms strong baselines like BigGAN and StyleGAN2. For instance, on CIFAR-10, MCGAN improved the Fréchet Inception Distance (FID) from 4.43 to 3.61 when applied to the BigGAN architecture. On the LSUN Bedroom dataset, it achieved a 35.7% reduction in FID.

Quantitative comparison of FID (lower is better) and IS (higher is better) over training steps. The MCGAN variant (Hinge+MC) shows faster and more stable improvement across both metrics.

In time-series tasks, MCGAN improved metrics like correlation and autocorrelation by 30-40%. For video generation, it produced clearer, more coherent frames with lower Mean Squared Error (MSE).

Insights and Future Directions

MCGAN's success lies in providing stronger, more consistent supervision. By matching statistical expectations rather than chasing a moving target, the training process becomes more robust. Furthermore, the resulting latent space is smoother, leading to more semantically meaningful interpolations between classes.

While highly effective, future work could explore applying MCGAN to more complex tasks like text-to-image generation or combining it with advanced modern discriminative losses. Although the Monte Carlo sampling adds some computational overhead per iteration, the gains in stability and convergence speed often outweigh these costs.

Summary updated

BibTeX
@inproceedings{xiao2025mcgan,
  author = {Baoren Xiao and Hao Ni and Weixin Yang},
  title = {{MCGAN: Enhancing GAN Training with Regression-Based Generator Loss}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {2025},
  doi = {10.1609/aaai.v39i20.35468},
}
Figure