Research map / Vision-language and multimodal / CLIP and zero-shot
Meta-Adapter: An Online Few-shot Learner for Vision-Language Model
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (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 Tip-Adapter's searched modulation with a meta-trained learnable adapter | CLIP few-shot classification · Novel classes across UCF101, Caltech101, DTD, and FGVCAircraft | accuracy: 40.09 -> 52.28; 68.33 -> 71.46; 42.92 -> 49.17; 13.96 -> 19.58 | Helped |
| Transfer the ImageNet-trained adapter without target-specific tuning | Meta-Adapter · Seven target datasets from ImageNet (average) | accuracy: 46.82 -> 51.81 | Helped |
| Transfer ImageNet-trained adapter weights rather than retuning on target data | CLIP RN50 · ImageNet-A, ImageNet-R, ImageNet-Sketch | accuracy: 23.88 -> 23.71; 60.54 -> 59.96; 35.45 -> 35.54 | Mixed |
| Integrate Meta-Adapter into ViLD | ViLD · LVIS val rare categories | APr: 18.1 -> 19.1 | Helped |
| Integrate Tip-Adapter into ViLD | ViLD · LVIS val rare categories | APr: 18.1 -> 11.3 | Hurt |
| Add multi-head attention to the Meta-Adapter (ablation) | Meta-Adapter · ImageNet | accuracy: 32.9 -> 40.2 | Helped |
| Add multi-head attention to the Meta-Adapter (ablation) | Meta-Adapter · SUN397 | accuracy: 28.9 -> 52.7 | Helped |
| Add multi-head attention to the Meta-Adapter (ablation) | Meta-Adapter · UCF101 | accuracy: 21.4 -> 51.4 | Helped |
| Add multi-head attention to the Meta-Adapter (ablation) | Meta-Adapter · DTD | accuracy: 10.0 -> 54.6 | Helped |
| Add the learnable gating block (ablation) | Meta-Adapter · ImageNet, SUN397, UCF101, and DTD | accuracy: 39.6 -> 40.2; 51.3 -> 52.7; 49.8 -> 51.4; 53.8 -> 54.6 | Helped |
| Remove value projection from the attention block (ablation) | Meta-Adapter · ImageNet, SUN397, UCF101, and DTD | accuracy: 25.6 -> 40.2; 18.2 -> 52.7; 5.9 -> 51.4; 5.0 -> 54.6 | Helped |
| Widen projection layers by ×2 (ablation) | Meta-Adapter · ImageNet | accuracy: 40.8 -> 40.1 | Hurt |
Closest papers by meaning
- Making Large Vision Language Models to Be Good Few-Shot Learners (AAAI 2025 · similarity 0.90)
- Revisiting Few-Shot Object Detection with Vision-Language Models (NeurIPS 2024 · similarity 0.90)
- Rethinking Few-Shot Adaptation of Vision-Language Models in Two Stages (CVPR 2025 · similarity 0.90)
- ProKeR: A Kernel Perspective on Few-Shot Adaptation of Large Vision-Language Models (CVPR 2025 · similarity 0.90)
- A Hard-to-Beat Baseline for Training-free CLIP-based Adaptation (ICLR 2024 · similarity 0.89)
- Demystifying CLIP Data (ICLR 2024 · 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.