Research map / Image generation / GANs
Bridging the Gap Between f-GANs and Wasserstein GANs
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 |
|---|---|---|---|
| Reweight generated samples using normalized exponentiated critic scores | WGAN on synthetic datasets · MoG, Banana, Rings, Square, Cosine, Funnel | NLL: 2.65 \pm 0.00 -> 2.54 \pm 0.00; 3.61 \pm 0.02 -> 3.57 \pm 0.00; 4.25 \pm 0.01 -> 4.25 \pm 0.00; 3.73 \pm 0.01 -> 3.72 \p | Mixed |
| Reweight generated samples using normalized exponentiated critic scores | WGAN on synthetic datasets · MoG, Banana, Rings, Square, Cosine, Funnel | MMD (×10^3): 25.45 \pm 7.78 -> 6.51 \pm 3.16; 3.33 \pm 0.59 -> 1.45 \pm 0.12; 2.05 \pm 0.47 -> 1.20 \pm 0.10; 2.42 \pm 0.24 -> 1.10 \ | Mixed |
| Reweight generated samples using normalized exponentiated critic scores | WGAN on UCI datasets · RedWine, WhiteWine, Parkinsons | NLL: 14.55 \pm 0.04 -> 14.41 \pm 0.03; 14.12 \pm 0.02 -> 14.08 \pm 0.02; 20.24 \pm 0.08 -> 20.16 \pm 0.05 | Helped |
| Reweight generated samples using normalized exponentiated critic scores | WGAN on UCI datasets · RedWine, WhiteWine, Parkinsons | MMD (×10^3): 2.61 \pm 0.37 -> 2.55 \pm 0.11; 1.32 \pm 0.10 -> 1.23 \pm 0.17; 1.30 \pm 0.09 -> 0.84 \pm 0.04 | Helped |
| Reweight generated samples using normalized exponentiated critic scores | BigGAN* · CIFAR10 unconditional | Inception score: 8.60 \pm .10 -> 8.66 \pm .09 | Helped |
| Reweight generated samples using normalized exponentiated critic scores | BigGAN* · CIFAR10 unconditional | FID: 16.38 -> 15.23 | Helped |
| Use the KL-WGAN objective in place of BigGAN's original objective | BigGAN* · CIFAR10 conditional | Inception score: 9.08 \pm .11 -> 9.20 \pm .09 | Helped |
| Use the KL-WGAN objective in place of BigGAN's original objective | BigGAN* · CIFAR10 conditional | FID: 9.51 -> 9.17 | Helped |
| Use the KL-WGAN objective in place of BigGAN's original objective | BigGAN · CelebA 64 × 64 | FID: 18.07 \pm 0.47 -> 17.70 \pm 0.32 | Helped |
| Use density-ratio-constrained KL-WGAN rather than WGAN | Distribution modeling · MoG, Square, Cosine | negative divergence estimates across 500 batches: 46, 181 and 55 -> 29, 100 and 7 | Helped |
| Estimate density ratios using KL-WGAN's constrained importance weights | KL-WGAN · Synthetic datasets | density ratio estimation quality | No clear effect |
| Replace the WGAN objective with the original KL f-GAN objective | BigGAN · Image generation | training stability | Hurt |
Closest papers by meaning
- Wasserstein Generative Adversarial Networks (ICML 2017 · similarity 0.92)
- Improving the Improved Training of Wasserstein GANs: A Consistency Term and Its Dual Effect (ICLR 2018 · similarity 0.91)
- Hidden Convexity of Wasserstein GANs: Interpretable Generative Models with Closed-Form Solutions (ICLR 2022 · similarity 0.90)
- Improved Training of Wasserstein GANs (NeurIPS 2017 · similarity 0.90)
- A Primal-Dual link between GANs and Autoencoders (NeurIPS 2019 · similarity 0.90)
- A Two-Step Computation of the Exact GAN Wasserstein Distance (ICML 2018 · similarity 0.90)
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.