Research map / Vision-language and multimodal / CLIP and zero-shot
Black Box Few-Shot Adaptation for Vision-Language Models
Yassine Ouali, Adrian Bulat, Brais Matinez, Georgios Tzimiropoulos
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 soft-prompt learning (CoOp) with Linear Feature Alignment | CLIP ViT-B/16 · 11 few-shot datasets avg (16-shot) | top-1 accuracy: 79.71 -> 81.21 | Improved |
| replace CoOp with LFA on training cost | ImageNet ViT-B/16 · ImageNet 16-shot | training time: 3h 22min -> 3min 9s | Improved |
| add beta regularization to orthogonal Procrustes (ablation) | LFA · ImageNet 16-shot | top-1 accuracy: 52.5 -> 64.8 | Improved |
| add ARerank refinement on top of beta-Procrustes (ablation) | LFA · ImageNet 16-shot | top-1 accuracy: 64.8 -> 71.7 | Improved |
| use ARerank loss instead of CSLS (ablation) | LFA refinement · ImageNet 16-shot | top-1 accuracy: 71.1 -> 71.7 | Improved |
| use ARerank loss instead of contrastive loss (ablation) | LFA refinement · Aircraft 16-shot | top-1 accuracy: 40.1 -> 45.1 | Improved |
| add beta-Procrustes initialization before refinement (ablation) | LFA · ImageNet 16-shot | top-1 accuracy: 71.6 -> 71.7 | No clear effect |
| vary beta between 0.6 and 0.9 (ablation) | LFA · ImageNet 16-shot | top-1 accuracy: 72.45 -> 72.54 | No clear effect |
| use 5 crops and prompt template (ablation) | LFA · 11 datasets avg (ViT-B/16) | top-1 accuracy: 80.54 -> 81.21 | Improved |
| combine LFA with CoOp soft-prompt prototypes (ablation) | LFA · ImageNet 16-shot (ViT-B/16) | top-1 accuracy: 72.61 -> 73.10 | Improved |
| use two mappings (W and EMA mapping) for base-to-new generalization (ablation) | LFA · ImageNet new classes | top-1 accuracy: 67.1 -> 69.4 | Improved |
| use LFA for unsupervised adaptation (U-LFA, n=5) without labels | CLIP ViT-B/16 · ImageNet | top-1 accuracy: 62.8 -> 68.6 | Improved |
Closest papers by meaning
- Rethinking Few-Shot Adaptation of Vision-Language Models in Two Stages (CVPR 2025 · similarity 0.91)
- Frustratingly Easy Test-Time Adaptation of Vision-Language Models (NeurIPS 2024 · similarity 0.91)
- A Closer Look at the Few-Shot Adaptation of Large Vision-Language Models (CVPR 2024 · similarity 0.91)
- Language Models as Black-Box Optimizers for Vision-Language Models (CVPR 2024 · similarity 0.91)
- On the Test-Time Zero-Shot Generalization of Vision-Language Models: Do We Really Need Prompt Learning? (CVPR 2024 · similarity 0.90)
- Make Prompts Adaptable: Bayesian Modeling for Vision-Language Prompt Learning with Data-Dependent Prior (AAAI 2024 · similarity 0.90)
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.
