Research map / Image generation / GANs
MCGAN: Enhancing GAN Training with Regression-Based Generator Loss
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes MCGAN, which replaces the standard GAN generator loss with a regression (MSE) loss matching real-data discriminator outputs to the Monte Carlo expected fake-data discriminator output, improving BigGAN/StyleGAN2 image generation, RCGAN time series, and ConvLSTM video prediction.
Problem
GAN generator training is unstable and under-supervised, with oscillation and gradient issues when the discriminator becomes accurate; conditional WGAN additionally needs costly conditional expectation estimation for continuous conditioning variables.
Approach
Reformulate generator training as a least-squares regression: minimize the squared difference between the discriminator output on real samples and the Monte Carlo estimate of the expected discriminator output on generated samples conditioned on the same label, under a weaker discriminability condition instead of an optimal discriminator.
Key results
- On CIFAR-10 with BigGAN, Hinge+DiffAug+MC improves FID from 4.43 to 3.61 and IS from 9.61 to 9.96.
- On CIFAR-10 with StyleGAN2, Hinge+MC+DiffAug achieves FID 2.16 versus 2.25 for the DiffAug baseline.
- On FFHQ256 and ImageNet64 with StyleGAN2-ada, FID improves from 4.51 to 3.77 and from 19.83 to 16.76.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| replace the generator loss with regression loss | BigGAN with Hinge loss and DiffAug · CIFAR-10 | FID: 4.43 to 3.61 | Helped |
| add regression generator loss without DiffAug (ablation) | BigGAN with Hinge loss · CIFAR-10 | FID: 5.31 -> 4.51 | Hurt |
| add DiffAug without regression generator loss (ablation) | BigGAN with Hinge loss · CIFAR-10 | FID: 5.31 -> 4.43 | Helped |
| combine regression generator loss with DiffAug (ablation) | BigGAN with Hinge loss · CIFAR-10 | FID: 4.43 -> 3.61 | Helped |
| add regression generator loss | StyleGAN2 with Hinge loss and DiffAug · CIFAR-10 | FID: 2.25 -> 2.16 | Helped |
| replace the generator loss with regression loss | BigGAN · CIFAR-100 | FID: 7.37 -> 5.77 | Helped |
| replace the generator loss with regression loss | StyleGAN2-ada · FFHQ256 | FID: 4.51 -> 3.77 | Helped |
| replace the generator loss with regression loss | StyleGAN2-ada · ImageNet64 | FID: 19.83 -> 16.76 | Helped |
| replace the generator loss with regression loss | ConvLSTM GAN · Moving MNIST test | MSE: 0.1012 to 0.0840 | Helped |
| replace the generator loss with regression loss | RCGAN with Hinge loss · SPX/DJI | Corr: 0.33790 -> 0.06233 | Helped |
| replace the generator loss with regression loss | RCGAN on SPX/DJI · SPX/DJI | ABS: 0.00868 -> 0.00996 | Hurt |
Surprising findings
- Sensible-looking gains on time series with d=100 reach 30-40% across metrics, with the improvement growing with dimension.
- On stock data, MC with BCE raised the ABS metric (0.01022 to 0.01286), so the baseline was better on that metric.
Limitations
- Theoretical guarantees rely on assumptions (e.g., Assumption on H and discriminability) that are not verified in practice.
- Results for text-to-image and multimodal settings are left to future work.
- Video experiment uses only one-step prediction and a classical GAN baseline.
Experimental setup
Datasets: CIFAR-10 CIFAR-100 FFHQ256 ImageNet64 Moving MNIST VAR synthetic time series S&P 500 and DJI stock data
Benchmarks: CIFAR-10 CIFAR-100 FFHQ256 ImageNet64 Moving MNIST VAR SPX/DJI stock
Metrics: Inception Score FID IFID Weak Accuracy Strong Accuracy Precision Recall MSE ABS Correlation
Baselines: BigGAN StyleGAN2 StyleGAN2-ada RCGAN TimeGAN GMMN SigCWGAN GAN (ConvLSTM video)
Builds on: BigGAN StyleGAN2 DiffAug RCGAN f-GAN
Closest papers by meaning
- MGAN: Training Generative Adversarial Nets with Multiple Generators (ICLR 2018 · similarity 0.91)
- MSG-GAN: Multi-Scale Gradients for Generative Adversarial Networks (CVPR 2020 · similarity 0.91)
- AdaGAN: Boosting Generative Models (NeurIPS 2017 · similarity 0.91)
- Bayesian GAN (NeurIPS 2017 · similarity 0.90)
- Metropolis-Hastings Generative Adversarial Networks (ICML 2019 · similarity 0.90)
- RG-GAN: Dynamic Regenerative Pruning for Data-Efficient Generative Adversarial Networks (AAAI 2024 · similarity 0.90)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.