Research map / Vision-language and multimodal / CLIP and zero-shot
Attention to the Burstiness in Visual Prompt Tuning!
Yuzhu Wang, Manni Duan, Shu Kong
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 SPT prompts with whitened bursty prompts (BPT-fWhiten) (ablation) | SPT · ImageNet-1K (ablation, MAE ViT-B shallow) | top-1 accuracy: 69.98 -> 72.09 | Improved |
| tune the whitening matrix (ablation) | BPT-fWhiten · ImageNet-1K (ablation, MAE ViT-B shallow) | top-1 accuracy: 72.09 -> 72.37 | Improved |
| use random-init bilinear prompt instead of tuned whitening (ablation) | BPT-tWhiten · ImageNet-1K (ablation, MAE ViT-B shallow) | top-1 accuracy: 72.37 -> 72.15 | No clear effect |
| increase prompt length from 64 to 196 (ablation) | BPT-bilinear · ImageNet-1K (ablation, MAE ViT-B shallow) | top-1 accuracy: 71.92 -> 72.48 | Improved |
| increase prompt width from 25 to 100 (ablation) | BPT-bilinear · ImageNet-1K (ablation, MAE ViT-B shallow) | top-1 accuracy: 72.06 -> 72.18 | Improved |
| increase prompt width from 25 to 100 (ablation) | BPT-bilinear · CUB-200 (ablation, MAE ViT-B shallow) | top-1 accuracy: 76.54 -> 78.59 | Improved |
| extend shallow to deep prompts in 4 blocks (ablation) | BPT-bilinear · ImageNet-1K (ablation, MAE ViT-B) | top-1 accuracy: 72.15 -> 74.25 | Improved |
| insert prompts in all 12 blocks rather than 4 (ablation) | BPT-Deep · ImageNet-1K (ablation, MAE ViT-B) | top-1 accuracy: 74.25 -> 74.73 | Improved |
| insert prompts in 3 blocks rather than 4 (ablation) | BPT-Deep · ImageNet-1K (ablation, MAE ViT-B) | top-1 accuracy: 74.25 -> 74.07 | Hurt |
| replace SPT-Deep with BPT-Deep | SPT-Deep · ImageNet-1K (ablation, MAE ViT-B) | top-1 accuracy: 73.34 -> 74.25 | Improved |
| replace SPT-S with BPT-S under MAE | SPT-S · FGVC mean (5 datasets, MAE pretrained ViT-B) | mean accuracy: 73.95 -> 80.39 | Improved |
| replace SPT-S with BPT-S under MoCo-v3 | SPT-S · FGVC mean (5 datasets, MoCo-v3 pretrained ViT-B) | mean accuracy: 84.08 -> 85.05 | Improved |
Closest papers by meaning
- DA-VPT: Semantic-Guided Visual Prompt Tuning for Vision Transformers (CVPR 2025 · similarity 0.91)
- Facing the Elephant in the Room: Visual Prompt Tuning or Full finetuning? (ICLR 2024 · similarity 0.91)
- Revisiting the Power of Prompt for Visual Tuning (ICML 2024 · similarity 0.90)
- Visual Fourier Prompt Tuning (NeurIPS 2024 · similarity 0.89)
- Fair-VPT: Fair Visual Prompt Tuning for Image Classification (CVPR 2024 · similarity 0.89)
- VIPAMIN: Visual Prompt Initialization via Embedding Selection and Subspace Expansion (NeurIPS 2025 · similarity 0.89)
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.
