Research map / Image generation / Fast diffusion sampling
Polarity Sampling: Quality and Diversity Control of Pre-Trained Generative Networks via Singular Values
Ahmed Imtiaz Humayun, Randall Balestriero, Richard Baraniuk
Check 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 |
|---|---|---|---|
| Apply polarity sampling at ρ=0.2 | StyleGAN2 · LSUN Church 256×256 | FID: 6.29 -> 6.02 | Improved |
| Apply polarity sampling in pixel space at ρ=-0.001 | StyleGAN2 · LSUN Church 256×256 | FID: 3.97 -> 3.92 | Improved |
| Apply polarity sampling in style space at ρ=0.01 | StyleGAN2 · LSUN Car 512×384 | FID: 2.34 -> 2.27 | Improved |
| Apply polarity sampling in style space at ρ=-0.1 | StyleGAN2 · LSUN Cat 256×256 | FID: 6.49 -> 6.39 | Improved |
| Apply polarity sampling in InceptionV3 space at ρ=0.01 | StyleGAN2-F · FFHQ 1024×1024 | FID: 2.74 -> 2.57 | Improved |
| Apply polarity sampling in VGG space at ρ=-0.001 | StyleGAN3-T · AFHQv2 512×512 | FID: 4.05 -> 3.95 | Improved |
| Apply pixel-space polarity sampling at ρ=-0.01 | ProGAN · CelebAHQ 1024×1024 | FID: 7.37 -> 7.28 | Improved |
| Increase the number of latent samples from N=100K to N=500K (ablation) | StyleGAN2 · FFHQ 1024×1024 | FID (lowest): 2.63 -> 2.58 | Improved |
| Increase the number of latent samples from N=100K to N=500K (ablation) | StyleGAN2 · LSUN Cat 256×256 | FID (lowest): 6.38 -> 6.34 | No clear effect |
| Use K=40 top singular values rather than K=10 (ablation) | StyleGAN2 · FFHQ 1024×1024 | FID (lowest): 2.71 -> 2.57 | Improved |
| Use K=100 rather than K=40 top singular values (ablation) | StyleGAN2 · FFHQ 1024×1024 | FID (lowest): 2.57 -> 2.70 | Hurt |
| Use positive polarity to increase diversity | StyleGAN2 with ψ<1 · FFHQ | FID: {.69, 8.11, 11.1} points | Improved |
Closest papers by meaning
- An Efficient Explorative Sampling Considering the Generative Boundaries of Deep Generative Neural Networks (AAAI 2020 · similarity 0.89)
- Plug & Play Generative Networks: Conditional Iterative Generation of Images in Latent Space (CVPR 2017 · similarity 0.88)
- Efficient Precision and Recall Metrics for Assessing Generative Models using Hubness-aware Sampling (ICML 2024 · similarity 0.86)
- Statistically Optimal Generative Modeling with Maximum Deviation from the Empirical Distribution (ICML 2024 · similarity 0.86)
- Consistency Posterior Sampling for Diverse Image Synthesis (CVPR 2025 · similarity 0.86)
- On the Stability of Iterative Retraining of Generative Models on their own Data (ICLR 2024 · similarity 0.85)
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.
