Research map / Image generation / GANs
Quantitatively Evaluating GANs With Divergences Proposed for Training
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 |
|---|---|---|---|
| Evaluate GAN frameworks with adaptive divergence metrics instead of relying only on visual assessment or standard metrics | DCGAN, W-DCGAN, and LS-DCGAN · CIFAR10 human comparison | agreement fraction: 0.977 | Helped |
| Train LS-DCGAN rather than DCGAN | GAN framework · MNIST | IW: 7.01 -> 4.50 | Helped |
| Train LS-DCGAN rather than DCGAN | GAN framework · CIFAR10 | LS: -0.0408 -> -0.0535 | Helped |
| Train W-DCGAN rather than DCGAN | GAN framework · LSUN Bedroom | IW: 3.79097 -> 2.91787 | Helped |
| Use LS instead of IW as the test metric | GAN evaluation · LSUN Bedroom | validation-set agreement: 55% -> 88% | Helped |
| Use a larger network architecture (ablation) | GANs · Architecture-size experiments | LS score | Helped |
| Use a large discriminator with a small generator rather than a large generator with a small discriminator (ablation) | LS-DCGAN and W-DCGAN · CIFAR10 | LS: -0.0092 ± 0.0007 -> -0.0372 ± 0.0068 | Helped |
| Use a large generator with a small discriminator rather than a small architecture for both networks (ablation) | GANs · Architecture-size experiments | sample quality | Hurt |
| Set noise dimension to 50 rather than 100 or 150 (ablation) | LS-DCGAN · CIFAR10 | IW: 5.6588 ± 1.47 -> 3.9010 ± 0.60 | Helped |
| Set noise dimension to 150 rather than 50 or 100 (ablation) | W-DCGAN · CIFAR10 | IW: 6.0948 ± 3.21 -> 3.6945 ± 1.33 | Helped |
| Tune the discriminator-to-generator update ratio by dataset and model (ablation) | DCGAN, W-DCGAN, and LS-DCGAN · MNIST and CIFAR10 | LS score | Mixed |
| Increase training examples from 10,000 to 40,000 (ablation) | DCGAN · CIFAR10 | LS score improvement: +0.0124 ± 0.00127 | Helped |
Closest papers by meaning
- Are GANs Created Equal? A Large-Scale Study (NeurIPS 2018 · similarity 0.93)
- Towards GAN Benchmarks Which Require Generalization (ICLR 2019 · similarity 0.92)
- A Domain Agnostic Measure for Monitoring and Evaluating GANs (NeurIPS 2019 · similarity 0.91)
- SMaRt: Improving GANs with Score Matching Regularity (ICML 2024 · similarity 0.89)
- MCGAN: Enhancing GAN Training with Regression-Based Generator Loss (AAAI 2025 · similarity 0.89)
- The GAN is dead; long live the GAN! A Modern GAN Baseline (NeurIPS 2024 · 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.