Research map / Image generation / GANs
Top-k Training of GANs: Improving GAN Performance by Throwing Away Bad Samples
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add top-k generator updates (drop lowest-scoring batch elements) | SA-GAN · CIFAR-10 | FID: 19.02 -> 17.93 | Helped |
| add top-k generator updates | DC-GAN · CIFAR-10 | FID: 38.09 -> 35.62 | Helped |
| add top-k generator updates | ICR-BigGAN · CIFAR-10 | FID: 9.21 -> 8.57 | Helped |
| add top-k generator updates | SAGAN · ImageNet | FID: 19.98 -> 18.44 | Helped |
| add top-k generator updates | GAN on mixture of Gaussians (100 modes) · MoG | % high quality samples: 40.3 -> 54.7 | Helped |
| add top-k generator updates | MEG · MNIST anomaly detection | AUPRC: 0.402 -> 0.561 (digit 5) | Helped |
| add top-k generator updates | SAGAN · CIFAR-10 NDB/K | NDB/K: 0.75 -> 0.60 | Helped |
| add top-k updates at batch sizes 64 to 256 | SAGAN · CIFAR-10 | FID: 21.1 -> 19.8 (64); 18.6 -> 17.4 (256) | Helped |
| decay k faster (gamma 0.5 instead of 0.99) (ablation) | Top-k SAGAN · CIFAR-10 | FID: 17.93 -> 25.30 | Hurt |
| decay k slower to minimum 90% of batch (nu=0.9 instead of 0.5) (ablation) | Top-k SAGAN · CIFAR-10 | FID: 17.93 -> 18.47 | Hurt |
| gamma 0.99 vs 0.9 vs 0.999 (ablation) | Top-k SAGAN · CIFAR-10 | FID: 17.93 vs 18.14 vs 18.68 | No clear effect |
| apply top-k to the critic only (ablation) | Top-k SAGAN · CIFAR-10 | FID: 17.93 -> 27.44 | Hurt |
Closest papers by meaning
- GLeaD: Improving GANs With a Generator-Leading Task (CVPR 2023 · similarity 0.90)
- A Large-Scale Study on Regularization and Normalization in GANs (ICML 2019 · similarity 0.90)
- Progressive Growing of GANs for Improved Quality, Stability, and Variation (ICLR 2018 · similarity 0.90)
- Training Generative Adversarial Networks with Limited Data (NeurIPS 2020 · similarity 0.90)
- MCGAN: Enhancing GAN Training with Regression-Based Generator Loss (AAAI 2025 · similarity 0.90)
- Are GANs Created Equal? A Large-Scale Study (NeurIPS 2018 · 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.