Research map / Architectures and efficiency / Attention and transformers
GFNet: Global Filter Networks for Visual Recognition
Yongming Rao, Wenliang Zhao, Zheng Dong Zhu, Jie Zhou, Jiwen Lu
Read the paperCheck 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 |
|---|---|---|---|
| replace self-attention with global filter layer (GFNet-Ti vs DeiT-Ti) | DeiT-Ti · ImageNet-1k val | top-1 accuracy (%): 72.2 -> 74.6 | Improved |
| replace spatial MLP with global filter (GFNet-Ti vs gMLP-Ti) | gMLP-Ti · ImageNet-1k val | top-1 accuracy (%): 72.0 -> 74.6 | Improved |
| use global filter instead of ResMLP token mixing | ResMLP-12 · ImageNet-1k val | top-1 accuracy (%): 76.6 -> 78.6 | Improved |
| use global filter instead of self-attention at DeiT-S scale | DeiT-S · ImageNet-1k val | top-1 accuracy (%): 79.8 -> 80.0 | No clear effect |
| finetune at 384 resolution for 30 epochs | GFNet-B · ImageNet-1k val | top-1 accuracy (%): 80.7 -> 82.1 | Improved |
| finetune at 384 resolution for 30 epochs | GFNet-XS · ImageNet-1k val | top-1 accuracy (%): 78.6 -> 80.6 | Improved |
| use hierarchical GFNet-H instead of ResNet-50-scale CNN | ResNet-50 · ImageNet-1k val | top-1 accuracy (%): 76.1 -> 81.5 | Improved |
| use GFNet-H-B instead of Swin-S | Swin-S · ImageNet-1k val | top-1 accuracy (%): 83.0 -> 82.9 | No clear effect |
| use GFNet-H-Ti backbone instead of PVT-Ti | PVT-Ti · ADE20K | mIoU: 35.7 -> 41.0 | Improved |
| use GFNet-H-B backbone instead of Swin-S | Swin-S · ADE20K | mIoU: 45.2 -> 44.8 | Hurt |
| use GFNet-S instead of DeiT-S for robustness | DeiT-S · ImageNet FGSM/PGD | accuracy (%): 40.7 -> 42.6; 16.7 -> 21.0 | Improved |
| use GFNet-S instead of DeiT-S on ImageNet-A | DeiT-S · ImageNet-A | top-1 accuracy (%): 18.9 -> 14.3 | Hurt |
Closest papers by meaning
- Global Filter Networks for Image Classification (NeurIPS 2021 · similarity 0.98)
- Learning Versatile Convolution Filters for Efficient Visual Recognition (TPAMI 2021 · similarity 0.90)
- LSNet: See Large, Focus Small (CVPR 2025 · similarity 0.88)
- AsCAN: Asymmetric Convolution-Attention Networks for Efficient Recognition and Generation (NeurIPS 2024 · similarity 0.87)
- Deep Edge Filter: Return of the Human-Crafted Layer in Deep Learning (NeurIPS 2025 · similarity 0.87)
- LookWhere? Efficient Visual Recognition by Learning Where to Look and What to See from Self-Supervision (NeurIPS 2025 · 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.
