Research map / Architectures and efficiency / Attention and transformers
Norm$\times$Direction: Restoring the Missing Query Norm in Vision Linear Attention
Weikang Meng, Yadan Luo, Liangyu Huo, Yingjian Li, Yaowei Wang, Xin Li, Zheng Zhang
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 cosine direction similarity to ReLU + constant power linear attention (ablation) | FLatten-Swin-T · ImageNet-1K | Top-1 accuracy (%): 82.1 -> 82.5 | Improved |
| upgrade constant power to query-norm-aware power function (ablation) | FLatten-Swin-T with cosine direction · ImageNet-1K | Top-1 accuracy (%): 82.5 -> 82.9 | Improved |
| replace attention with NaLaFormer in Swin-T | Swin-T · ImageNet-1K | Top-1 accuracy (%): 81.2 -> 82.9 | Improved |
| QK-Norm instead of FLatten baseline (ablation) | FLatten Swin / PVT · ImageNet-1K | Top-1 accuracy (%): -0.04% on Swin, +0.15% on PVT | Mixed |
| NaLaFormer instead of FLatten baseline (ablation) | FLatten Swin / PVT · ImageNet-1K | Top-1 accuracy (%): +0.21% on Swin, +0.50% on PVT | Improved |
| apply NaLa feature map to ESRT for super-resolution | ESRT · Set5/Set14/BSD100/Urban100 x4 | PSNR: Urban100 x4: 25.85 -> 25.83 | No clear effect |
| apply NaLa to DCTLSA for super-resolution | DCTLSA · Urban100 x4 | PSNR: 26.41 -> 26.74 | Improved |
| replace SiLU with query-norm-aware feature map in DeltaNet | DeltaNet 340M · common-sense reasoning average | average accuracy: 43.6 -> 44.3 | Improved |
| replace SiLU with query-norm-aware feature map in Gated DeltaNet | Gated DeltaNet 340M · common-sense reasoning average | average accuracy: 44.3 -> 44.8 | Improved |
| replace attention with NaLaFormer in diffusion transformer | SiT S/2 · ImageNet-1K | FID: 58.61 -> 53.08 | Improved |
Closest papers by meaning
- Bridging the Divide: Reconsidering Softmax and Linear Attention (NeurIPS 2024 · similarity 0.91)
- PolaFormer: Polarity-aware Linear Attention for Vision Transformers (ICLR 2025 · similarity 0.88)
- Attention Alignment Between Humans and Vision-Language Models (NeurIPS 2026 · similarity 0.87)
- Breaking the Low-Rank Dilemma of Linear Attention (CVPR 2025 · similarity 0.87)
- Degrees of Freedom for Linear Attention: Distilling Softmax Attention with Optimal Feature Efficiency (NeurIPS 2025 · similarity 0.87)
- MetaLA: Unified Optimal Linear Approximation to Softmax Attention Map (NeurIPS 2024 · similarity 0.86)
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.
