Research map / Detection and segmentation / CNN features
Improving CLIP Fine-tuning Performance
Yixuan Wei, Han Hu, Zhenda Xie, Ze Liu, Zheng Zhang, Yue Cao, Jianmin Bao, Dong Chen, Baining Guo
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 |
|---|---|---|---|
| feature-distill CLIP ViT-B/16 into a fresh student (FD-CLIP) | CLIP ViT-B/16 · ImageNet-1K fine-tuning | top-1 accuracy (%): 82.9 -> 85.0 | Improved |
| feature-distill CLIP ViT-B/16 | CLIP ViT-B/16 · COCO | APbox: 45.0 -> 48.2 | Improved |
| feature-distill CLIP ViT-B/16 | CLIP ViT-B/16 · NYUv2 | RMSE: 0.416 -> 0.352 | Improved |
| use full feature map as distillation target instead of [CLS] token (ablation) | FD-CLIP · ImageNet-1K fine-tuning | top-1 accuracy (%): 81.9 -> 84.4 | Improved |
| use full feature map instead of GAP feature (ablation) | FD-CLIP · ADE20K | mIoU: 50.3 -> 51.8 | Improved |
| distill full images instead of masked partial images (ablation) | FD-CLIP · ImageNet-1K fine-tuning | top-1 accuracy (%): 84.4 -> 84.4 (75% vs full) | No clear effect |
| standardize teacher feature map instead of none (ablation) | FD-CLIP · ImageNet-1K fine-tuning | top-1 accuracy (%): 83.5 -> 84.4 | Improved |
| standardization instead of l2 norm (ablation) | FD-CLIP · ImageNet-1K fine-tuning | top-1 accuracy (%): 83.9 -> 84.4 | Improved |
| add drop path on teacher branch (ablation) | FD-CLIP · ImageNet-1K fine-tuning | top-1 accuracy (%): 84.4 -> 84.0 | Hurt |
| shared relative position bias instead of APE (ablation) | FD-CLIP · ImageNet-1K fine-tuning | top-1 accuracy (%): 84.0 -> 84.4 | Improved |
| apply feature distillation to DeiT | DeiT ViT-B · NYUv2 | RMSE: 0.403 -> 0.404 | No clear effect |
| apply feature distillation to SwinV2-G | SwinV2-G · COCO | APbox: 63.1 -> 64.2 | Improved |
Closest papers by meaning
- DFD: Distilling the Feature Disparity Differently for Detectors (ICML 2024 · similarity 0.88)
- Dataset Distillation for Pre-Trained Self-Supervised Vision Models (NeurIPS 2025 · similarity 0.86)
- Hybrid Distillation: Connecting Masked Autoencoders with Contrastive Learners (ICLR 2024 · similarity 0.86)
- TPR: Topology-Preserving Reservoirs for Generalized Zero-Shot Learning (NeurIPS 2024 · similarity 0.85)
- Emphasizing Discriminative Features for Dataset Distillation in Complex Scenarios (CVPR 2025 · similarity 0.85)
- Classifier-guided CLIP Distillation for Unsupervised Multi-label Classification (CVPR 2025 · similarity 0.85)
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.
