Research map / Image generation / Autoregressive and latent generation
On Aliased Resizing and Surprising Subtleties in GAN Evaluation
Check 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 |
|---|---|---|---|
| Use aliased resizing rather than antialiased PIL-bicubic for both real and generated images | StyleGAN2 FID evaluation · FFHQ | FID: 2.98 -> 2.68 | Helped |
| Use aliased resizing rather than antialiased PIL-bicubic for both real and generated images | StyleGAN2 FID evaluation · MetFaces | FID: 65.32 -> 55.09 | Helped |
| Use aliased resizing rather than antialiased PIL-bicubic for both real and generated images | StyleGAN2 FID evaluation · AFHQ-Cats and AFHQ-Dogs | FID: 5.13 -> 4.80; 20.16 -> 18.25 | Helped |
| Downsample FFHQ training images with antialiased PIL-bicubic rather than aliased resizing | StyleGAN2 Config-E · FFHQ 1024→256 | FID: 4.82 ± 0.09 -> 6.21 ± 0.23 | Hurt |
| Use aliased TensorFlow-bilinear rather than PIL-bicubic to resize the same real images | FID feature extraction · FFHQ images 1024→299 | FID: 0 -> 4.34 | Hurt |
| Compress generated images with JPEG quality 87 instead of saving as PNG | StyleGAN2 trained on JPEG-75 LSUN Church images · LSUN outdoor Church | FID: 4.00 -> 3.48 | Helped |
| Apply 8-bit quantization to generated images | FID evaluation · Not specified | FID: <0.01 | No clear effect |
| Add JPEG compression to real images during evaluation | FID and KID evaluation · FFHQ at multiple resolutions | FID | Hurt |
Closest papers by meaning
- The Role of ImageNet Classes in Fréchet Inception Distance (ICLR 2023 · similarity 0.88)
- Revisiting the Evaluation of Image Synthesis with GANs (NeurIPS 2023 · similarity 0.87)
- Projected GANs Converge Faster (NeurIPS 2021 · similarity 0.87)
- Representation Fréchet Loss for Visual Generation (NeurIPS 2026 · similarity 0.87)
- Singular Value Scaling: Efficient Generative Model Compression via Pruned Weights Refinement (AAAI 2025 · similarity 0.86)
- Rethinking FID: Towards a Better Evaluation Metric for Image Generation (CVPR 2024 · 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.