Research map / Image generation / GANs
Spectral Regularization for Combating Mode Collapse in GANs
Kanglin Liu, Wenming Tang, Fei Zhou, Guoping Qiu
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 |
|---|---|---|---|
| add spectral regularization (compensating singular values) to spectral normalization | SN-GAN · ImageNet E2048-64 | Inception Score: 21.78 -> 31.58 | Improved |
| add spectral regularization | SN-GAN · ImageNet E512-64 | Inception Score: 34.13 -> 40.76 | Improved |
| add spectral regularization | SN-GAN · CIFAR-10 C64-32 (small discriminator) | FID: 85.15 -> 56.11 | Improved |
| add spectral regularization | SN-GAN · CIFAR-10 A16-128 | FID: 22.31 -> 24.67 | Hurt |
| add spectral regularization | SN-GAN · all 26 settings | mode collapse: none happened to SR-GANs in all 10 settings where SN-GANs collapsed | Improved |
| add spectral regularization | SN-GAN · average over settings | IS and FID: IS improved by 13.9% and FID by 21.8% | Improved |
| resume SN-GAN snapshot at 40k iterations with dynamic-compensation SR (ablation) | SN-GAN · ImageNet E2048-64 | IS and FID: IS improved by 44.9% and FID by 9.1% | Improved |
| increase static compensation i from 0.25r to 0.5r (ablation) | SR-GAN · groups A, D, E | Inception Score: 0.25N to 0.5N improved | Improved |
| increase static compensation i from 0.5r to r (ablation) | SR-GAN · groups A, D, E | Inception Score: 0.5N to N deteriorated | Hurt |
| increase static compensation i (ablation) | SR-GAN · group B | Inception Score: performance increases steadily with i | Improved |
| dynamic instead of static compensation (ablation) | SR-GAN · A128-128 and D128-256 | Inception Score: static better | Hurt |
| dynamic instead of static compensation (ablation) | SR-GAN · B16-64, C64-32, E512-64 | Inception Score: dynamic better in the rest tasks | Improved |
Closest papers by meaning
- Why Spectral Normalization Stabilizes GANs: Analysis and Improvements (NeurIPS 2021 · similarity 0.91)
- Mode Regularized Generative Adversarial Networks (ICLR 2017 · similarity 0.90)
- DynGAN: Solving Mode Collapse in GANs With Dynamic Clustering (TPAMI 2024 · similarity 0.90)
- On Computation and Generalization of Generative Adversarial Networks under Spectrum Control (ICLR 2019 · similarity 0.90)
- Spectral Normalization for Generative Adversarial Networks (ICLR 2018 · similarity 0.89)
- The GAN is dead; long live the GAN! A Modern GAN Baseline (NeurIPS 2024 · similarity 0.88)
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.
