Research map / Architectures and efficiency / Pruning and efficient tuning
Visual Prompt Tuning
Menglin Jia, Luming Tang, Bor-Chun Chen, Claire Cardie, Serge Belongie, Bharath Hariharan, Ser-Nam Lim
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 bias tuning to VPT-Deep (ablation) | VPT-Deep · VTAB-Structured | mean accuracy: 54.98 -> 53.87 | Hurt |
| add bias tuning to VPT-Shallow (ablation) | VPT-Shallow · VTAB-Natural | mean accuracy: 76.81 -> 79.78 | Improved |
| insert prompts in pixel space (Prepend-pixel) instead of latent embedding space (ablation) | VPT-Shallow · VTAB-Natural | mean accuracy: drops 6.9% | Hurt |
| concatenate a new input channel instead of prepending prompt tokens (ablation) | VPT · VTAB-Natural | mean accuracy: 30 accuracy scores drop | Hurt |
| pool final prompt outputs into the classifier input (ablation) | VPT-Deep · VTAB | accuracy: as large as 8 points | Hurt |
| average-pool image patch outputs instead of CLS as final output (ablation) | VPT-Deep · VTAB-Specialized | accuracy: 82.4 vs 82.3 | No clear effect |
| learn the prompt embeddings rather than keeping them frozen (ablation) | VPT · VTAB | accuracy | Improved |
| share one prompt across layers (Shared-inter) instead of per-layer prompts (ablation) | VPT-Deep · VTAB | accuracy: 1.14x vs 1.13x params | No clear effect |
| share the same prompt among and within layers (Shared-all) (ablation) | VPT-Deep · VTAB | accuracy | Hurt |
| initialize prompts with class-average CLS embeddings instead of random (ablation) | VPT · VTAB | accuracy | Hurt |
| insert prompts in earlier layers rather than only later layers (prompt depth) (ablation) | VPT-Deep · VTAB | accuracy | Improved |
| use VPT-Deep instead of full fine-tuning on ViT-B | ViT-B/16 · VTAB-Structured | mean accuracy: 47.64 -> 54.98 | Improved |
Closest papers by meaning
- DA-VPT: Semantic-Guided Visual Prompt Tuning for Vision Transformers (CVPR 2025 · similarity 0.93)
- Facing the Elephant in the Room: Visual Prompt Tuning or Full finetuning? (ICLR 2024 · similarity 0.92)
- Fair-VPT: Fair Visual Prompt Tuning for Image Classification (CVPR 2024 · similarity 0.90)
- Revisiting the Power of Prompt for Visual Tuning (ICML 2024 · similarity 0.89)
- Visual Fourier Prompt Tuning (NeurIPS 2024 · similarity 0.89)
- Skip Tuning: Pre-trained Vision-Language Models are Effective and Efficient Adapters Themselves (CVPR 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.
