Research map / Architectures and efficiency / RNNs and normalization
Balanced Training for Sparse GANs
Yite Wang, Jing Wu, Naira Hovakimyan, Ruoyu Sun
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (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 |
|---|---|---|---|
| adjust discriminator density with balance ratio instead of fixed strong discriminator | SDST-Strong-RigL · BigGAN CIFAR-10, G density 10% | training FLOPs: 83.97\% -> 10.39\% | Improved |
| use ADAPT_relax over second-best baseline | sparse SNGAN · CIFAR-10, G density 10% | FID: 2.3 FID improvement | Improved |
| use ADAPT_relax over second-best baseline | sparse SNGAN · STL-10, G density 10% | FID: 43.08 -> 35.98 | Improved |
| apply DST to both generator and discriminator (DST-bothGD) instead of generator only (ablation) | SDST-Strong-SET · SNGAN CIFAR-10, G density 10% | FID: 16.49 -> 20.57 | Hurt |
| apply DST to both generator and discriminator (RigL) (ablation) | SDST-Strong-RigL · SNGAN CIFAR-10, G density 10% | FID: 17.02 -> 31.95 | Hurt |
| ADAPT_relax instead of post-hoc magnitude pruning | sparse SNGAN · CIFAR-10, G density 10% | FID: 20.89 -> 14.19 | Improved |
| strong discriminator strategy instead of balance strategy | STATIC · SNGAN STL-10, G density 10% | FID: 48.18 -> 52.48 | Hurt |
| RigL growth instead of random SET growth (ablation) | SDST-Strong · SNGAN STL-10, G density 10% | FID: 67.37 -> 53.65 | Improved |
| add dynamic density adjust (DDA) to STATIC | STATIC SNGAN · CIFAR-10 | FID | Improved |
| SDST instead of STATIC when discriminator density is small (10%) | sparse SNGAN · CIFAR-10 | FID | No clear effect |
| ADAPT_strict instead of ADAPT_relax (ablation) | ADAPT · SNGAN STL-10, G density 10% | FID: 35.98 -> 41.18; 10.19 -> 9.29 | Mixed |
| ADAPT_strict instead of SDST-Strong-RigL under d_D<=50% | SNGAN STL-10, G density 10% · STL-10 | FID: 53.74 -> 41.18 | Improved |
Closest papers by meaning
- Don’t Be So Dense: Sparse-to-Sparse GAN Training Without Sacrificing Performance (IJCV 2023 · similarity 0.90)
- RG-GAN: Dynamic Regenerative Pruning for Data-Efficient Generative Adversarial Networks (AAAI 2024 · similarity 0.87)
- Improving GANs with A Dynamic Discriminator (NeurIPS 2022 · similarity 0.87)
- Improving GAN Training with Probability Ratio Clipping and Sample Reweighting (NeurIPS 2020 · similarity 0.87)
- Instability and Local Minima in GAN Training with Kernel Discriminators (NeurIPS 2022 · similarity 0.87)
- Exploring Sparse MoE in GANs for Text-conditioned Image Synthesis (CVPR 2025 · similarity 0.86)
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.
