Research map / Image generation / GANs
Discriminative Forests Improve Generative Diversity for Generative Adversarial Networks
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 Forest-GAN, which replaces the GAN discriminator with a forest of independently trained discriminators on bootstrapped data, and shows improved FID/recall on STL10 (AutoGAN) and LSUN-cat (StyleGAN2-ADA).
Problem
GAN discriminators with high capacity overfit and drive the generator toward mode collapse, limiting generative diversity; capacity and robustness of the discriminator must be balanced.
Approach
Replace the single discriminator with a discriminative forest: K independent discriminators, each trained on a bootstrap sample of the real data, whose outputs are averaged to update the shared generator. A random-forest-style generalization bound (strength and correlation) motivates that large forests add capacity without overfitting.
Key results
- AutoGAN with a 10-discriminator forest (10M total params) reaches FID 19.27 on STL10, versus 30.71 for the original AutoGAN.
- StyleGAN2-ADA with a 5-discriminator forest reaches FID 6.87 on LSUN-cat, versus 9.22 for the original model and 7.45 for PAR-GAN.
- On simulated 2D GMM data, Forest-GAN recovers all 9 modes once K≥20, while K=1 learns only 4.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Replace AutoGAN's single discriminator with a K=10 bootstrapped discriminative forest (10M total parameters) | AutoGAN · STL10 | FID: 30.71 -> 19.27 | Helped |
| Replace StyleGAN2-ADA's single discriminator with a K=5 bootstrapped discriminative forest | StyleGAN2-ADA · LSUN-cat | FID: 9.22 -> 6.87 | Helped |
| Use a K=5 forest instead of the original single discriminator | StyleGAN2-ADA · LSUN-cat | Recall: 0.202 -> 0.279 | Helped |
| Increase K from 1 to 10 while keeping the total discriminator parameters at 2M (ablation) | AutoGAN · STL10 | FID: 29.10 -> 22.72 | Helped |
| Increase K from 1 to 10 while keeping the total discriminator parameters at 5M (ablation) | AutoGAN · STL10 | FID: 27.96 -> 19.82 | Helped |
| Increase K from 1 to 10 while keeping the total discriminator parameters at 10M (ablation) | AutoGAN · STL10 | FID: 31.23 -> 19.27 | Helped |
| Increase discriminator capacity from 5M to 10M with K=1 (ablation) | AutoGAN · STL10 | FID: 27.96 -> 31.23 | Hurt |
| Increase K from 1 to 10 with a fixed-width discriminator architecture (ablation) | AutoGAN · STL10 | FID: 30.71 -> 24.52 | Helped |
| Increase K from 1 to 10 with a fixed-width discriminator architecture (ablation) | AutoGAN · STL10 | Recall: 0.471 -> 0.554 | Helped |
| Increase the number of discriminators from K=1 to K=50 (ablation) | Forest-GAN · 2D-GMM simulated data | Recall: 0.463 -> 0.906 | Helped |
| Replace Forest-GAN with PAR-GAN using K=5 | StyleGAN2-ADA · LSUN-cat | FID: 6.87 -> 7.45 | Hurt |
Surprising findings
- Increasing discriminator parameters at fixed forest size did not improve quality or diversity, while more discriminators at fixed total parameters did.
- PAR-GAN degrades with more discriminators on small datasets, whereas Forest-GAN remains stable because bootstrap datasets keep full size.
Limitations
- Correlation among discriminators is not computed or controlled; the bound cannot be directly calculated in practice.
- Training time grows with K, so only K up to 10 on STL10 and K up to 5 on LSUN-cat were evaluated.
- Results are limited to two image datasets and two base GAN architectures.
Experimental setup
Datasets: STL10 LSUN-cat simulated 2D Gaussian mixture
Benchmarks: STL10 LSUN-cat
Metrics: FID Inception Score precision recall perceptual path length
Baselines: AutoGAN StyleGAN2-ADA PAR-GAN D2GAN GMAN Dropout-GAN MCL-GAN
Builds on: GAN Random Forest PAR-GAN AutoGAN StyleGAN2-ADA
Closest papers by meaning
- Improving Generalization and Stability of Generative Adversarial Networks (ICLR 2019 · similarity 0.90)
- Improving GANs with A Dynamic Discriminator (NeurIPS 2022 · similarity 0.89)
- Training Generative Adversarial Networks with Limited Data (NeurIPS 2020 · similarity 0.89)
- Direct Discriminative Optimization: Your Likelihood-Based Visual Generative Model is Secretly a GAN Discriminator (ICML 2025 · similarity 0.88)
- MCGAN: Enhancing GAN Training with Regression-Based Generator Loss (AAAI 2025 · similarity 0.87)
- Improving the Training of Data-Efficient GANs via Quality Aware Dynamic Discriminator Rejection Sampling (CVPR 2025 · similarity 0.87)
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.