Research map / Architectures and efficiency / Attention and transformers
SPViT: Enabling Faster Vision Transformers via Latency-Aware Soft Token Pruning
Zhenglun Kong, Peiyan Dong, Xiaolong Ma, Xin Meng, Wei Niu, Mengshu Sun, Xuan Shen, Geng Yuan, Bin Ren, Hao Tang, Minghai Qin, Yanzhi Wang
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 |
|---|---|---|---|
| use learned token selector instead of random token removal (ablation) | SPViT DeiT-T · ImageNet-1K at 0.90 GFLOPs | top-1 accuracy (%): 69.87 -> 72.10 | Improved |
| use learned token selector instead of random token removal (ablation) | SPViT DeiT-S · ImageNet-1K at 2.64 GFLOPs | top-1 accuracy (%): 77.25 -> 79.34 | Improved |
| use token selector instead of structure (dimension) pruning (ablation) | SPViT DeiT-T · ImageNet-1K at 0.90 GFLOPs | top-1 accuracy (%): 70.32 -> 72.10 | Improved |
| use token selector instead of structure (dimension) pruning (ablation) | SPViT DeiT-S · ImageNet-1K at 2.64 GFLOPs | top-1 accuracy (%): 77.86 -> 79.34 | Improved |
| place first selector after block 3 instead of block 1 (ablation) | SPViT DeiT-S · ImageNet-1K | top-1 accuracy (%): 76.10 -> 79.34 | Improved |
| place last selector after block 9 instead of block 11 (ablation) | SPViT DeiT-S · ImageNet-1K | top-1 accuracy (%): 78.76 -> 79.34 | Improved |
| use three selector phases (3-6-9) instead of two (6-9) (ablation) | SPViT DeiT-S · ImageNet-1K | top-1 accuracy (%): 78.53 -> 79.34 | Improved |
| use four selectors (3-5-7-9) instead of three (3-6-9) (ablation) | SPViT DeiT-S · ImageNet-1K | top-1 accuracy (%): 79.34 -> 79.34 | No clear effect |
| apply SPViT token pruning to DeiT-T | DeiT-T · ImageNet-1K | top-1 accuracy (%): 72.20 -> 72.10; GFLOPs 31% lower | No clear effect |
| apply SPViT to Swin-T | Swin-T · ImageNet-1K | top-1 accuracy (%): 81.20 -> 80.70; GFLOPs 4.50 -> 3.47 | Mixed |
| apply SPViT on mobile CPU | DeiT-S · Samsung Galaxy S20 latency | latency (ms): 113 -> 60 ms | Improved |
| apply SPViT on FPGA | DeiT-T · Xilinx ZCU102 latency | latency (ms): 8.81 -> 5.60 ms | Improved |
Later papers that compared with this method
2 later accepted papers report SPViT as a baseline in their results tables; 2 of them report a better number on most of the results they share with it.
- PYRA: Parallel Yielding Re-Activation for Training-Inference Efficient Task Adaptation (ECCV 2024 · better on 1 of 1 shared result)
- VTC-LFC: Vision Transformer Compression with Low-Frequency Components (NeurIPS 2022 · better on 2 of 3 shared results)
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- DynamicViT: Efficient Vision Transformers with Dynamic Token Sparsification (NeurIPS 2021 · similarity 0.90)
- SAViT: Structure-Aware Vision Transformer Pruning via Collaborative Optimization (NeurIPS 2022 · similarity 0.89)
- SHViT: Single-Head Vision Transformer with Memory Efficient Macro Design (CVPR 2024 · similarity 0.89)
- Zero-TPrune: Zero-Shot Token Pruning through Leveraging of the Attention Graph in Pre-Trained Transformers (CVPR 2024 · similarity 0.88)
- Token Cropr: Faster ViTs for Quite a Few Tasks (CVPR 2025 · similarity 0.88)
- Multi-criteria Token Fusion with One-step-ahead Attention for Efficient Vision Transformers (CVPR 2024 · similarity 0.88)
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.
