Research map / Architectures and efficiency / Attention and transformers
Blending Anti-Aliasing into Vision Transformer
Shengju Qian, Hao Shao, Yi Zhu, Mu Li, Jiaya Jia
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 |
|---|---|---|---|
| Add filter-bank anti-aliasing (ARM) to attention maps in early layers | Swin-T · ImageNet-1k | top-1 accuracy: 81.2 -> 82.0 | Improved |
| Add ARM to other transformer families (DeiT-S, DeiT-B, CoaT, T2T) | DeiT/CoaT/T2T · ImageNet-1k | top-1 accuracy: 79.8 -> 80.7 | Improved |
| Use fixed Gaussian filter (ablation) | Swin-T · ImageNet-1k | top-1 accuracy: 81.2 -> 81.5 | Improved |
| Use learnable convolution filter instead of Gaussian (ablation) | Swin-T · ImageNet-1k | top-1 accuracy: 81.5 -> 81.6 | No clear effect |
| Use pre-defined filter bank (n=8) instead of Gaussian (ablation) | Swin-T · ImageNet-1k | top-1 accuracy: 81.5 -> 82.0 | Improved |
| Apply anti-aliasing after patch embedding (ablation) | Swin-T · ImageNet-1k | top-1 accuracy: 81.2 -> 81.0 | Hurt |
| Apply anti-aliasing to attention maps instead of patch embedding (ablation) | DeiT-S · ImageNet-1k | top-1 accuracy: 79.8 -> 80.7 | Improved |
| Apply anti-aliasing after attention plus shortcut (ablation) | Swin-T · ImageNet-1k | top-1 accuracy: 81.2 -> 79.8 | Hurt |
| Apply anti-aliasing at patch merging (downsampling) (ablation) | Swin-T · ImageNet-1k | top-1 accuracy: 81.2 -> 79.5 | Hurt |
| Filter only layers 1-2 instead of all four layers (ablation) | Swin-T · ImageNet-1k | top-1 accuracy: 81.7 vs 81.3 | Mixed |
| Increase filter bank size from 2 to 8 (ablation) | Swin-T · ImageNet-1k | top-1 accuracy: 81.6 -> 82.0 | Improved |
| Increase filter bank size beyond 8 (16, 24) (ablation) | Swin-T · ImageNet-1k | top-1 accuracy: 82.0 -> 81.8 | No clear effect |
Closest papers by meaning
- Alias-Free ViT: Fractional Shift Invariance via Linear Attention (NeurIPS 2025 · similarity 0.87)
- Frequency-Aware Token Reduction for Efficient Vision Transformer (NeurIPS 2025 · similarity 0.86)
- An Image is Worth More Than 16x16 Patches: Exploring Transformers on Individual Pixels (ICLR 2025 · similarity 0.86)
- Skip-Attention: Improving Vision Transformers by Paying Less Attention (ICLR 2024 · similarity 0.85)
- Vision Transformers Need Registers (ICLR 2024 · similarity 0.85)
- You Only Need Less Attention at Each Stage in Vision Transformers (CVPR 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.
