Research map / Vision-language and multimodal / CLIP and zero-shot
No Token Left Behind: Explainability-Aided Image Classification and Generation
Roni Paiss, Hila Chefer, Lior Wolf
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 explainability loss to CoOp prompt learning (ablation) | CoOp (CLIP ViT-B/16) · ImageNet 1-shot unified | top-1 accuracy (%): 66.93 -> 67.62 (M=4); 63.66 -> 67.09 (M=16) | Improved |
| add explainability loss to CoOp prompt learning (class-specific) (ablation) | CoOp (CLIP ViT-B/16) · ImageNet 1-shot CSC | top-1 accuracy (%): 46.20 -> 48.74 | Improved |
| add explainability loss to CoOp prompt learning (ResNet-50) (ablation) | CoOp (CLIP ResNet-50) · ImageNet-A 1-shot unified | top-1 accuracy (%): 21.43 -> 22.74 | Improved |
| few-shot CoOp and linear probing instead of manual 0-shot prompts | CLIP ResNet-50 · ImageNet 1-shot unified | top-1 accuracy (%): 58.18 -> 54.45 (CoOp M=16); 58.18 -> 21.70 (LP) | Hurt |
| use class name at end of prompt with explainability loss (ablation) | CoOp (ResNet-50) · ImageNet 1-shot | top-1 accuracy (%): 53.41 -> 57.78 | Improved |
| add explainability loss in 2-shot setting (ablation) | CoOp (ResNet-50) · ImageNet 2-shot | top-1 accuracy (%): 56.70 -> 58.42 | Improved |
| add explainability loss in 4-shot setting (ablation) | CoOp (ViT-B/16) · ImageNet 4-shot | top-1 accuracy (%): 68.20 -> 68.90 | Improved |
| add explainability loss to StyleCLIP latent optimization | StyleCLIP · user study prompt 'A man with a beard' | quality rating (1-5): 2.92 -> 4.28 | Improved |
| add explainability loss to StyleCLIP latent optimization | StyleCLIP · user study prompt 'A man with a beard' | identity preservation rating (1-5): 3.61 -> 2.23 | Hurt |
| add explainability loss to StyleCLIP latent optimization | StyleCLIP · user study prompt 'A person with purple hair' | identity preservation rating (1-5): 4.13 -> 3.51 | Hurt |
| use explainability IoU loss instead of similarity-based masked CLIP loss | CLIP-guided VQGAN · MSCOCO spatial conditioning | AP (%): 8 -> 26.2 | Improved |
| use explainability IoU loss instead of similarity-based loss | CLIP-guided VQGAN · MSCOCO spatial conditioning | precision (%): 46.4 -> 71.7 | Improved |
Closest papers by meaning
- Interpreting and Analysing CLIP's Zero-Shot Image Classification via Mutual Knowledge (NeurIPS 2024 · similarity 0.92)
- Transductive Zero-Shot and Few-Shot CLIP (CVPR 2024 · similarity 0.89)
- PerceptionCLIP: Visual Classification by Inferring and Conditioning on Contexts (ICLR 2024 · similarity 0.88)
- TPR: Topology-Preserving Reservoirs for Generalized Zero-Shot Learning (NeurIPS 2024 · similarity 0.88)
- Seeing What Matters: Empowering CLIP with Patch Generation-to-Selection (CVPR 2025 · similarity 0.88)
- Mining Fine-Grained Image-Text Alignment for Zero-Shot Captioning via Text-Only Training (AAAI 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.
