Research map / Architectures and efficiency / Pruning and efficient tuning
Grid-Attention: Enhancing Computational Efficiency of Large Vision Models without Fine-Tuning
Pengyu Li, biao wang, Tianchu Guo, Xian-Sheng Hua
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 |
|---|---|---|---|
| plug GridAttn into frozen SAM-B without fine-tuning | SAM-B · COCO instance segmentation | AP: 41.02 -> 41.39 AP; GFLOPs 969.6 -> 813.8 | Improved |
| plug GridAttn into frozen SAM-H without fine-tuning | SAM-H · COCO instance segmentation | AP: 46.13 -> 46.12 AP; GFLOPs 5952 -> 5692 | No clear effect |
| use Expedit-SAM token merging instead of GridAttn | SAM-H · COCO instance segmentation | AP: 46.13 -> 42.33 | Hurt |
| add GridAttn on top of Expedit-SAM | Expedit-SAM · COCO instance segmentation (SAM-H) | GFLOPs / AP: 5094 -> 4893 GFLOPs; AP 42.33 -> 42.42 | Improved |
| naive 2x2 sub-image sampling (lower boundary) instead of GridAttn (ablation) | frozen SAM-B · COCO instance segmentation | AP: 41.39 -> 36.50 | Hurt |
| fine-tune SegFormer with GridAttn for 10% iterations instead of naive fine-tuning | SegFormer B0-B5 · ADE20K | mIoU (MS): B1: 43.9 -> 44.1; B4: 51.0 -> 51.3 | Improved |
| use Shuffle strategy instead of Straightforward when training from scratch (ablation) | GridAttn SegFormer · ADE20K from scratch | mIoU (MS): B4: 49.1 -> 50.7; B5: 51.1 -> 51.8 | Improved |
| use Deformable strategy instead of Straightforward when training from scratch (ablation) | GridAttn SegFormer-B5 · ADE20K from scratch | mIoU (MS): 51.1 -> 52.0; 423.7 -> 305.4 GFLOPs vs benchmark | Improved |
| use Deformable strategy instead of Shuffle (ablation) | GridAttn SegFormer-B5 · ADE20K from scratch | GPU latency: 56.2 -> 240 ms | Hurt |
| increase grid stride S (ablation) | frozen SAM with GridAttn · COCO instance segmentation | performance-efficiency tradeoff | Mixed |
| reduce fine-tuning iterations to 1600 (1%) (ablation) | GridAttn SegFormer · ADE20K | mIoU (MS): B3: 50.4 -> 49.5 | Hurt |
| fine-tune more than 8000 iterations (ablation) | GridAttn SegFormer · ADE20K | mIoU | No clear effect |
Closest papers by meaning
- Revisiting the Integration of Convolution and Attention for Vision Backbone (NeurIPS 2024 · similarity 0.88)
- Resource-Efficient Transformer Pruning for Finetuning of Large Models (CVPR 2024 · similarity 0.88)
- LaVin-DiT: Large Vision Diffusion Transformer (CVPR 2025 · similarity 0.88)
- LookHere: Vision Transformers with Directed Attention Generalize and Extrapolate (NeurIPS 2024 · similarity 0.87)
- Large Vision-Language Models Get Lost in Attention (ICML 2026 · similarity 0.87)
- Linear Differential Vision Transformer: Learning Visual Contrasts via Pairwise Differentials (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.
