Research map / Vision-language and multimodal / CLIP and zero-shot
An Inverse Scaling Law for CLIP Training
Xianhang Li, Zeyu Wang, Cihang Xie
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 |
|---|---|---|---|
| use global average pooling instead of class token (ablation) | CLIPA-L/16 (I17,T16) · ImageNet-1k zero-shot | top-1: 66.3 -> 67.8 | Improved |
| add color jitter augmentation (ablation) | CLIPA-L/16 (I17,T16) · ImageNet-1k zero-shot | top-1: 66.3 -> 67.1 | Improved |
| combine GAP and color jitter (ablation) | CLIPA-L/16 (I17,T16) · ImageNet-1k zero-shot | top-1: 66.3 -> 68.1 | Improved |
| fine-tune with batch size 7k instead of 32k (ablation) | CLIPA-L/16 (I17,T16) · ImageNet-1k zero-shot | top-1: 68.1 -> 67.6 | No clear effect |
| fine-tuning batch size sweep 32k to 2k (ablation) | CLIPA-L/16 (I17,T16) · ImageNet-1k zero-shot | top-1: 68.1 -> 66.8 | Hurt |
| scale model from L/16 to H/14 (ablation) | CLIPA · ImageNet-1k zero-shot | top-1: 69.3 -> 72.8 | Improved |
| switch data from LAION-400M to LAION-2B (ablation) | CLIPA-H/14 · ImageNet-1k zero-shot | top-1: 72.8 -> 74.1 | Improved |
| increase pre-training schedule 5x (12.8B samples) (ablation) | CLIPA-H/14 · ImageNet-1k zero-shot | top-1: 74.1 -> 77.9 | Improved |
| 25% random masking during fine-tuning (ablation) | CLIPA-H/14 · ImageNet-1k zero-shot | top-1: 77.9 -> 78.2 | Improved |
| 75% masking during fine-tuning (block) (ablation) | CLIPA-H/14 · ImageNet-1k zero-shot | top-1: 77.9 -> 74.3 | Hurt |
| 4x longer fine-tuning with 30% random masking (ablation) | CLIPA-H/14 · ImageNet-1k zero-shot | top-1: 78.0 -> 78.6 | Improved |
| progressive fine-tuning 224 then 336 vs direct 336 (ablation) | CLIPA-H/14 · ImageNet-1k zero-shot | top-1: 78.9 -> 79.1 | Improved |
Closest papers by meaning
- Demystifying CLIP Data (ICLR 2024 · similarity 0.88)
- un$^2$CLIP: Improving CLIP's Visual Detail Capturing Ability via Inverting unCLIP (NeurIPS 2025 · similarity 0.88)
- TiC-CLIP: Continual Training of CLIP Models (ICLR 2024 · similarity 0.88)
- Does CLIP’s generalization performance mainly stem from high train-test similarity? (ICLR 2024 · similarity 0.87)
- A Hard-to-Beat Baseline for Training-free CLIP-based Adaptation (ICLR 2024 · similarity 0.87)
- Scaling Laws of Synthetic Images for Model Training ... for Now (CVPR 2024 · similarity 0.87)
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.
