Research map / Image generation / GANs
Exploring and Exploiting Hubness Priors for High-Quality GAN Latent Sampling
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 |
|---|---|---|---|
| select latents with hubness threshold t=50 instead of random sampling | StyleGAN2 · FFHQ-1 | FID: 35.455 -> 21.955 | Helped |
| select latents with hubness threshold t=50 instead of truncating with ψ=0.7 | StyleGAN2 · FFHQ-1 | FID: 25.097 -> 21.955 | Helped |
| select latents with hubness threshold t=50 instead of truncating with ψ=0.7 | StyleGAN2 · FFHQ-2 | FID: 25.127 -> 23.609 | Helped |
| select latents by hubness instead of truncating with ψ=0.7 | BigGAN · ImageNet ILSVRC 2012 | precision: 0.131 -> 0.147 | Helped |
| select latents by hubness instead of truncating with ψ=0.7 | BigGAN · ImageNet ILSVRC 2012 | recall: 0.264 -> 0.311 | Helped |
| use k=7 instead of k=5 (ablation) | StyleGAN2 hubness sampling · StyleGAN2 W-space, n=10000, t=50 | FID: 22.782 -> 22.720 | No clear effect |
| use k=10 instead of k=5 (ablation) | StyleGAN2 hubness sampling · StyleGAN2 W-space, n=10000, t=50 | FID: 22.782 -> 22.560 | No clear effect |
| use k=1 in the nearest-neighbor algorithm (ablation) | StyleGAN2 hubness sampling · StyleGAN2 | image quality | Hurt |
| raise the hubness threshold from t=50 to t=60 (ablation) | StyleGAN2 hubness sampling · StyleGAN2 W-space, k=5, n=10000 | FID: 22.782 -> 20.749 | Helped |
| lower the hubness threshold from t=50 to t=40 (ablation) | StyleGAN2 hubness sampling · StyleGAN2 W-space, k=5, n=10000 | FID: 22.782 -> 24.517 | Hurt |
| lower the hubness threshold from t=50 to t=35 (ablation) | StyleGAN2 hubness sampling · StyleGAN2 W-space, k=5, n=10000 | FID: 22.782 -> 25.412 | Hurt |
| increase latent sample set size from n=10000 to n=40000 (ablation) | StyleGAN2 hubness sampling · StyleGAN2 W-space, k=5, t=50 | FID: 22.782 -> 19.124 | Helped |
Closest papers by meaning
- Improving Generative Adversarial Networks via Adversarial Learning in Latent Space (NeurIPS 2022 · similarity 0.90)
- Your GAN is Secretly an Energy-based Model and You Should Use Discriminator Driven Latent Sampling (NeurIPS 2020 · similarity 0.89)
- ClusterGAN: Latent Space Clustering in Generative Adversarial Networks (AAAI 2019 · similarity 0.89)
- Complexity Matters: Rethinking the Latent Space for Generative Modeling (NeurIPS 2023 · similarity 0.89)
- Non-Parametric Priors For Generative Adversarial Networks (ICML 2019 · similarity 0.89)
- Unveiling the Latent Space Geometry of Push-Forward Generative Models (ICML 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.