Research map / Image generation / GANs
Self-Attention 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 |
|---|---|---|---|
| Use SAGAN rather than SNGAN-projection | Class-conditional image generation · ImageNet | Inception Score: 36.8 -> 52.52 | Helped |
| Use SAGAN rather than SNGAN-projection | Class-conditional image generation · ImageNet | FID: 27.62 -> 18.65 | Helped |
| Add self-attention at feat_32 instead of feat_8 (ablation) | SAGAN · ImageNet | FID: 22.98 -> 18.28 | Helped |
| Add self-attention at feat_64 instead of feat_8 (ablation) | SAGAN · ImageNet | Inception Score: 43.15 -> 52.52 | Helped |
| Remove self-attention (ablation) | SAGAN · ImageNet | FID: 18.65 -> 22.96 | Hurt |
| Replace self-attention with a residual block at feat_8 (ablation) | SAGAN · ImageNet | FID: 22.98 -> 42.13 | Hurt |
| Replace self-attention with a residual block at feat_32 (ablation) | SAGAN · ImageNet | FID: 18.28 vs 27.33 | Hurt |
| Apply spectral normalization to the generator as well as the discriminator (ablation) | GAN with 1:1 generator/discriminator updates · ImageNet training | Training stability | Helped |
| Use TTUR with spectral normalization on generator and discriminator (ablation) | SAGAN training · ImageNet training | Training stability | Helped |
| Use self-attention for classes with complex geometric structure | SAGAN · ImageNet classes | Intra FID | Helped |
| Use self-attention for classes with few structural constraints | SAGAN · ImageNet texture-dominated classes | Intra FID | Mixed |
Closest papers by meaning
- Your Local GAN: Designing Two Dimensional Local Attention Mechanisms for Generative Models (CVPR 2020 · similarity 0.89)
- Improved Transformer for High-Resolution GANs (NeurIPS 2021 · similarity 0.89)
- Large Scale GAN Training for High Fidelity Natural Image Synthesis (ICLR 2019 · similarity 0.88)
- A Unified Generative Adversarial Network Training via Self-Labeling and Self-Attention (ICML 2021 · similarity 0.88)
- Optimizing the Latent Space of Generative Networks (ICML 2018 · similarity 0.87)
- Dynamically Grown Generative Adversarial Networks (AAAI 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.