Research map / Architectures and efficiency / Attention and transformers
Your Local GAN: Designing Two Dimensional Local Attention Mechanisms for Generative Models
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 |
|---|---|---|---|
| replace dense attention with two-dimensional local sparse attention | SAGAN · ImageNet | FID: 18.65 -> 15.94 | Helped |
| replace dense attention with two-dimensional local sparse attention | SAGAN · ImageNet | Inception score: 52.52 -> 57.22 | Helped |
| increase dense attention from 1 to 8 heads (ablation) | SAGAN · ImageNet | FID: 18.65 -> 20.09 | Hurt |
| increase dense attention from 1 to 8 heads (ablation) | SAGAN · ImageNet | Inception score: 52.52 -> 46.01 | Hurt |
| add ESA two-dimensional locality mapping (ablation) | YLG-SAGAN without ESA · ImageNet | FID: 17.47 -> 15.94 | Helped |
| add ESA two-dimensional locality mapping (ablation) | YLG-SAGAN without ESA · ImageNet | Inception score: 51.09 -> 57.22 | Helped |
| replace LTR and RTL with the Full Information Strided pattern (ablation) | YLG-SAGAN · ImageNet | FID: 15.94 -> 16.64 | Hurt |
| replace LTR and RTL with the Full Information Strided pattern (ablation) | YLG-SAGAN · ImageNet | Inception score: 57.22 -> 55.21 | Hurt |
| combine 4 sparse heads with 4 dense heads (ablation) | SAGAN · ImageNet | FID: 19.21 | Hurt |
| combine 4 sparse heads with 4 dense heads (ablation) | SAGAN · ImageNet | Inception score: 51.23 | Hurt |
| use narrow 2-position and wide 6-position dense heads (ablation) | SAGAN · ImageNet | FID: 19.57 | Hurt |
| use narrow 2-position and wide 6-position dense heads (ablation) | SAGAN · ImageNet | Inception score: 50.93 | Hurt |
Closest papers by meaning
- Self-Attention Generative Adversarial Networks (ICML 2019 · similarity 0.89)
- Improved Transformer for High-Resolution GANs (NeurIPS 2021 · similarity 0.87)
- Local Attention Pyramid for Scene Image Generation (CVPR 2022 · similarity 0.87)
- Global Versus Localized Generative Adversarial Nets (CVPR 2018 · similarity 0.87)
- Improving GAN Equilibrium by Raising Spatial Awareness (CVPR 2022 · similarity 0.86)
- Compressing Image-to-Image Translation GANs Using Local Density Structures on Their Learned Manifold (AAAI 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.