Research map / Architectures and efficiency / Attention and transformers
Quick ViTs: Speeding up Vision Transformers through Equivariance
David Nordström, Johan Edstedt, Fredrik Kahl, Georg Bökman
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 blocks with octic H8 layers | ViT-H/14 DeiT III · ImageNet-1K | Top-1: 84.6 -> 85.0 | Improved |
| replace blocks with octic I8 layers | ViT-H/14 DeiT III · ImageNet-1K | Top-1: 84.6 -> 84.7 | No clear effect |
| use I8 invariant model | ViT-H/14 DeiT III · ImageNet-1K rotated val | OOD rotation delta accuracy: -12.6 -> 0.0 | Improved |
| octic layers instead of standard for DINOv2 | ViT-H/16 · ImageNet-1K linear | accuracy: 81.7 -> 82.2 | Improved |
| I8 octic layers instead of standard | ViT-H/16 DINOv2 · ADE20K linear | mIoU: 34.7 -> 33.9 | Hurt |
| I8 octic layers instead of standard | ViT-L/16 DINOv2 · VOC2012 k-NN | mIoU: 58.1 -> 59.6 | Improved |
| use equivariant instead of arbitrary patch embedding (ablation) | H8(ViT-B) · ImageNet-1K DeiT III | Top-1: 82.4 -> 83.0 | Improved |
| use power spectrum invariantization instead of triple correlation (ablation) | D8(ViT-L/16) · ImageNet-1K | Top-1: 76.8 -> 79.5 | Improved |
| use power spectrum instead of max filtering (ablation) | D8(ViT-L/16) · ImageNet-1K | Top-1: 77.9 -> 79.5 | Improved |
| use power spectrum instead of linear invariant (ablation) | D8(ViT-L/16) · ImageNet-1K | Top-1: 79.2 -> 79.5 | Improved |
| keep late blocks non-equivariant (H8 vs I8) (ablation) | ViT-H/16 DINOv2 · ImageNet-1K linear | accuracy: 81.9 -> 82.2 | Improved |
| scale width from ViT-L to ViT-22B (ablation) | octic ViTs · speed bf16 | speedup: 1.32x -> 3.54x | Improved |
Closest papers by meaning
- Global Vision Transformer Pruning With Hessian-Aware Saliency (CVPR 2023 · similarity 0.89)
- Thicker and Quicker: The Jumbo Token for Fast Plain Vision Transformers (ICLR 2026 · similarity 0.89)
- Making Vision Transformers Truly Shift-Equivariant (CVPR 2024 · similarity 0.88)
- Your ViT is Secretly an Image Segmentation Model (CVPR 2025 · similarity 0.88)
- Slicing Vision Transformer for Flexible Inference (NeurIPS 2024 · similarity 0.88)
- Linearly Decomposing and Recomposing Vision Transformers for Diverse-Scale Models (NeurIPS 2024 · 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.
