Research map / Image generation / GANs
Robust training of implicit generative models for multivariate and heavy-tailed distributions with an invariant statistical loss
José Manuel de Frutos, Manuel A. Vázquez, Pablo M. Olmos, Joaquín Míguez
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 |
|---|---|---|---|
| replace Gaussian latent noise with generalized Pareto noise (Pareto-ISL) (ablation) | ISL · Keystrokes | A_CCDF: 2.7 -> 1.4 | Improved |
| replace Gaussian latent noise with generalized Pareto noise (Pareto-ISL) (ablation) | ISL · Wiki Traffic | A_CCDF: 8.6 -> 1.19 | Improved |
| add Pareto-ISL instead of Pareto GAN | Heavy-tailed generator · Keystrokes | KS: 0.013 -> 0.006 | Improved |
| replace adversarial training with Pareto-ISL | GAN/WGAN/MMD-GAN · Cauchy(1,2) | KSD: 0.03 -> 1.90e-3 | Improved |
| pretrain with ISL-slicing before GAN training | GAN · 2D Circle of Gaussians | KL divergence: 2.24 -> 0.46 | Improved |
| pretrain WGAN with ISL | WGAN · 2D-Ring, batch size 64 | %HQ: 30.01 -> 99.31 | Improved |
| pretrain DCGAN with ISL | DCGAN · MNIST | recall (%): 75.43 -> 96.68 | Improved |
| pretrain DCGAN with ISL | DCGAN · CelebA | recall: 0.834 -> 0.954 | Improved |
| pretrain DCGAN with ISL | DCGAN · CelebA | FID: 30.93 -> 31.64 | No clear effect |
| increase K progressively instead of fixed K=10 (ablation) | ISL · N(4,2) | training time (s): 9016 -> 4149 | Improved |
| increase K progressively instead of fixed K=10 (ablation) | ISL · Pareto(1,1) | KSD: 0.24238 -> 0.08344 | Improved |
| use ISL-slicing instead of marginal ISL (ablation) | ISL · Synthetic D=100 GAN data | execution time: 10x and 5x less execution time | Improved |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 2 on most shared results.
- MMD-GAN — MMD GAN: Towards Deeper Understanding of Moment Matching Network (NeurIPS 2017) · better on 2 of 2 shared results
- MCL-GAN — MCL-GAN: Generative Adversarial Networks with Multiple Specialized Discriminators (NeurIPS 2022) · better on 0 of 1 shared result
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- Pareto GAN: Extending the Representational Power of GANs to Heavy-Tailed Distributions (ICML 2021 · similarity 0.90)
- Statistically Optimal Generative Modeling with Maximum Deviation from the Empirical Distribution (ICML 2024 · similarity 0.88)
- Generative Modeling Using the Sliced Wasserstein Distance (CVPR 2018 · similarity 0.88)
- Taming the Tail in Class-Conditional GANs: Knowledge Sharing via Unconditional Training at Lower Resolutions (CVPR 2024 · similarity 0.87)
- Revisiting Adversarial Training Under Long-Tailed Distributions (CVPR 2024 · similarity 0.86)
- MCGAN: Enhancing GAN Training with Regression-Based Generator Loss (AAAI 2025 · similarity 0.86)
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.
