Research map / Image generation / GANs
Refining Deep Generative Models via Discriminator Gradient Flow
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 |
|---|---|---|---|
| Refine WGAN-GP samples with KL gradient flow | WGAN-GP · 2D synthetic datasets | KDE Score: -7037 $\pm$ 64 -> -2893 $\pm$ 07 | Helped |
| Refine WGAN-GP samples with KL gradient flow | WGAN-GP · 2D synthetic datasets | % High Quality: 26.5 $\pm$ .8 -> 89.5 $\pm$ .4 | Helped |
| Refine image samples with gradient flow | WGAN-GP · CIFAR10 | FID: 28.37 (.08) -> 23.15 (.07) | Helped |
| Refine image samples with gradient flow | SN-ResNet-GAN · CIFAR10 | FID: 14.10 (.06) -> 9.62 (.03) | Helped |
| Refine image samples with gradient flow | WGAN-GP · STL10 | FID: 51.50 (.15) -> 39.07 (.07) | Helped |
| Use KL instead of JS gradient flow (ablation) | DGflow on WGAN-GP · STL10 | FID: 50.83 (.06) -> 39.07 (.07) | Helped |
| Refine SN-ResNet-GAN with KL gradient flow | SN-ResNet-GAN · CIFAR10 | Inception Score: 8.38 (.03) -> 9.35 (.03) | Helped |
| Refine character-level GAN samples with gradient flow | Character-level GAN · Billion Words | JS-4: 0.224 (.0009) -> 0.186 (.0007) | Helped |
| Use JS rather than KL gradient flow (ablation) | DGflow on character-level GAN · Billion Words | JS-4: 0.212 (.0008) -> 0.186 (.0007) | Helped |
| Use log D rather than KL gradient flow (ablation) | DGflow on character-level GAN · Billion Words | JS-6: 0.512 (.0012) -> 0.506 (.0008) | Helped |
| Remove diffusion by setting gamma to zero (ablation) | DGflow · CIFAR10 and STL10 | FID | No clear effect |
| Add a density-ratio corrector to the surrogate discriminator estimate (ablation) | DGflow for models without corresponding discriminators · Image datasets | sample quality | Helped |
Closest papers by meaning
- Deep Generative Learning via Variational Gradient Flow (ICML 2019 · similarity 0.90)
- Mixed batches and symmetric discriminators for GAN training (ICML 2018 · similarity 0.90)
- Direct Discriminative Optimization: Your Likelihood-Based Visual Generative Model is Secretly a GAN Discriminator (ICML 2025 · similarity 0.89)
- Feature Statistics Mixing Regularization for Generative Adversarial Networks (CVPR 2022 · similarity 0.89)
- Deep MMD Gradient Flow without adversarial training (ICLR 2025 · similarity 0.88)
- MS$^3$D: A RG Flow-Based Regularization for GAN Training with Limited Data (ICML 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.