Research map / Image generation / GANs
Alias-Free Generative Adversarial Networks
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 |
|---|---|---|---|
| Replace the learned input constant with Fourier features (ablation) | StyleGAN2 generator · FFHQ 256×256 | FID: 5.14 -> 4.79 | Helped |
| Remove per-pixel noise inputs (ablation) | Fourier-feature generator · FFHQ 256×256 | EQ-T: 16.23 -> 15.81 | Hurt |
| Simplify the generator and disable selected StyleGAN2 regularizers and skips (ablation) | Generator without noise inputs · FFHQ 256×256 | EQ-T: 15.81 -> 19.47 | Helped |
| Extend boundaries and use windowed sinc upsampling (ablation) | Simplified generator · FFHQ 256×256 | FID: 5.21 -> 6.02 | Hurt |
| Filter nonlinearities by upsampling before and downsampling afterward (ablation) | Generator with improved boundaries and upsampling · FFHQ 256×256 | EQ-T: 24.62 -> 30.60 | Helped |
| Lower filter cutoff so alias frequencies fall in the stopband (ablation) | Generator with filtered nonlinearities · FFHQ 256×256 | EQ-T: 30.60 -> 43.90 | Helped |
| Learn global translation and rotation of input Fourier features (ablation) | Non-critical-sampling generator · FFHQ 256×256 | FID: 4.78 -> 4.64 | Helped |
| Specialize sampling rates and filter parameters for each layer (ablation) | Transformed-Fourier-feature generator · FFHQ 256×256 | EQ-T: 45.20 -> 63.01 | Helped |
| Use 1×1 convolutions and radial jinc downsampling filters for rotation equivariance (ablation) | StyleGAN3-T · FFHQ 256×256 | FID: 4.62 -> 4.50 | Helped |
| Reduce filter size from n=6 to n=4 (ablation) | StyleGAN3-R · FFHQ 256×256 | EQ-T: 66.65 -> 57.49 | Hurt |
| Reduce nonlinearity upsampling factor from m=2 to m=1 (ablation) | StyleGAN3-R · FFHQ 256×256 | EQ-T: 66.65 -> 39.96 | Hurt |
| Increase the initial stopband target from 2^2.1 to 2^3.1 (ablation) | StyleGAN3-R · FFHQ 256×256 | FID: 4.50 -> 4.68 | Hurt |
Closest papers by meaning
- Large Scale GAN Training for High Fidelity Natural Image Synthesis (ICLR 2019 · similarity 0.88)
- Generative Adversarial Transformers (ICML 2021 · similarity 0.88)
- On the "steerability" of generative adversarial networks (ICLR 2020 · similarity 0.88)
- FIANCEE: Faster Inference of Adversarial Networks via Conditional Early Exits (CVPR 2023 · similarity 0.88)
- Learning Modulated Transformation in GANs (NeurIPS 2023 · similarity 0.87)
- Projected GANs Converge Faster (NeurIPS 2021 · similarity 0.87)
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.