Research map / Image generation / GANs
How good is my GAN?
Konstantin Shmelkov, Cordelia Schmid, Karteek Alahari
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 |
|---|---|---|---|
| use SNGAN instead of WGAN-GP (10M) | GAN evaluation · CIFAR100 generated images | GAN-train accuracy: 26.7 -> 45.0 | Improved |
| use SNGAN instead of DCGAN | GAN evaluation · CIFAR10 generated images | GAN-test accuracy: 58.2 -> 87.3 | Improved |
| increase WGAN-GP size from 2.5M to 10M parameters (ablation) | WGAN-GP · CIFAR10 generated images | GAN-train accuracy: 76.1 -> 79.5 | Improved |
| increase WGAN-GP size from 2.5M to 10M parameters (ablation) | WGAN-GP · CIFAR100 generated images | GAN-test accuracy: 4.3 -> 40.4 | Improved |
| add salt and pepper noise (1% to 20% of pixels) to CIFAR10 (ablation) | GAN-test measure · CIFAR10 corrupted training set | GAN-test accuracy: 82% -> 15% | Hurt |
| subsample CIFAR10 train set to a random 5k subset with Gaussian noise (ablation) | GAN-train measure · CIFAR10 corrupted training set | GAN-train accuracy: 91% -> 80% | Hurt |
| train SNGAN at 128px instead of 64px | SNGAN · ImageNet generated images | GAN-train top-1 accuracy: 3 -> 9.3 | Improved |
| replace CNN classifier with random forests (ablation) | GAN-train/GAN-test · CIFAR100 generated images | GAN ranking: 15.2%, 19.5% for SNGAN | No clear effect |
| add 50k SNGAN images to 2.5k real images | CIFAR10 classifier · CIFAR10 augmentation, reduced real set | classification accuracy: 73.4 -> 67.0 | Hurt |
| add 50k SNGAN images to 10k real images | CIFAR100 classifier · CIFAR100 augmentation, reduced real set | classification accuracy: 51.5 -> 45.5 | Hurt |
| add 50k real images to 50k GAN images | CIFAR100 classifier · CIFAR100 augmentation | classification accuracy | No clear effect |
| add 2.5k or 5k real images to 50k GAN images | CIFAR10 and CIFAR100 classifier · augmentation with full-data SNGAN | classification accuracy | Improved |
Closest papers by meaning
- Are GANs Created Equal? A Large-Scale Study (NeurIPS 2018 · similarity 0.92)
- Quantitatively Evaluating GANs With Divergences Proposed for Training (ICLR 2018 · similarity 0.91)
- Improved Techniques for Training GANs (NeurIPS 2016 · similarity 0.90)
- Sequential Training of GANs Against GAN-Classifiers Reveals Correlated "Knowledge Gaps" Present Among Independently Trained GAN Instances (CVPR 2023 · similarity 0.89)
- Progressive Growing of GANs for Improved Quality, Stability, and Variation (ICLR 2018 · similarity 0.89)
- Re-GAN: Data-Efficient GANs Training via Architectural Reconfiguration (CVPR 2023 · similarity 0.89)
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.
