Research map / Image generation / GANs
TREND: Truncated Generalized Normal Density Estimation of Inception Embeddings for GAN Evaluation
Junghyuk Lee, Jong-Seok Lee
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 truncated generalized normal instead of normal density (ablation) | Inception feature density estimation · ImageNet Inception features | average log-likelihood: -9463 -> 23867 | Improved |
| add truncation at zero to generalized normal (ablation) | Inception feature density estimation · ImageNet Inception features | average log-likelihood: 879 -> 23867 | Improved |
| add sharper-peak shape parameter to normal (untruncated) (ablation) | Inception feature density estimation · ImageNet Inception features | average log-likelihood: -9463 -> 879 | Improved |
| replace FID with TREND when sample count drops from 50000 to 5000 | BigGAN evaluation · BigGAN thresholds 0.4 and 0.6 | metric stability: FID 20.2 vs 21.1 (mean at 0.6 worse); TREND 0.0139 vs 0.0094 | Improved |
| use TREND instead of FID in a toy scenario with non-Gaussian densities | Toy GAN evaluation · Toy scenario 1 | ranking of model 2 vs 1: FID 0.0017 vs 0.0018; TREND 0.0086 vs 0.0022 | Improved |
| use TREND instead of FID when GAN fails to learn truth | Toy GAN evaluation · Toy scenario 2 | score: TREND 0.015 vs FID 1.4e-4 | Improved |
| increase image disturbance level 1 to 3 (Gaussian noise) | TREND · ImageNet val disturbance | TREND: 0.0058 -> 0.0208 | Improved |
| increase random erasing level 1 to 3 | TREND · ImageNet val disturbance | TREND: 0.0369 -> 0.0827 | Improved |
| increase BigGAN truncation threshold 0.2 to infinity | TREND · BigGAN ImageNet | TREND: 0.0165 -> 0.0032 | Improved |
| increase BigGAN truncation threshold 0.2 to infinity | IS · BigGAN ImageNet | IS: 330.4 -> 144.6 | Hurt |
| use ADM-U instead of BigGAN (inf) under KID | KID · ImageNet | KID: 0.0009 vs 0.0015 | Hurt |
Closest papers by meaning
- Projected GANs Converge Faster (NeurIPS 2021 · similarity 0.88)
- A Large-Scale Study on Regularization and Normalization in GANs (ICML 2019 · similarity 0.88)
- On the Robustness of Quality Measures for GANs (ECCV 2022 · similarity 0.88)
- A Domain Agnostic Measure for Monitoring and Evaluating GANs (NeurIPS 2019 · similarity 0.88)
- MIND: Monge Inception Distance for Generative Models Evaluation (NeurIPS 2026 · similarity 0.87)
- 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.
