Research map / Vision-language and multimodal / CLIP and zero-shot
Wukong: A 100 Million Large-scale Chinese Cross-modal Pre-training Benchmark
Jiaxi Gu, Xiaojun Meng, Guansong Lu, Lu Hou, Minzhe Niu, Xiaodan Liang, Lewei Yao, Runhui Huang, Wei Zhang, Xin Jiang, Chunjing Xu, Hang Xu
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 character-grained instead of word-grained Chinese tokenization (ablation) | Wukong ViT-B · Zero-shot classification, 10 Chinese datasets | avg top-1 accuracy: 62.65 -> 65.65 | Improved |
| lock the image encoder (LiT-tuning) rather than train it (ablation) | Wukong ViT-B · ImageNet zero-shot validation accuracy | validation accuracy | Improved |
| lock the image encoder (LiT-tuning) (ablation) | Wukong ViT-B · training loss | training loss | Hurt |
| add token-wise similarity with reduced-token interaction (Wukong-style) vs FILIP-style on ViT-B | ViT-B model · Zero-shot classification, 10 Chinese datasets | avg top-1 accuracy: 58.09 -> 65.65 | Improved |
| add reduced-token interaction to Swin-L encoder (Wukong vs FILIP style) | Swin-L model · Zero-shot classification, 10 Chinese datasets | avg top-1 accuracy: 70.27 -> 69.12 | Hurt |
| Wukong ViT-L (reduced-token) vs CLIP ViT-L (global similarity) | ViT-L model · Zero-shot classification, 10 Chinese datasets | avg top-1 accuracy: 72.60 -> 73.03 | Improved |
| pre-train on Wukong dataset (CLIP/FILIP/Wukong models) vs BriVL on its own data | Chinese VLP models · Zero-shot classification, 10 Chinese datasets | avg top-1 accuracy: 43.72 -> 73.03 | Improved |
| Wukong ViT-L vs WenLan 2.0 pre-trained on 650M pairs | Wukong ViT-L · AIC-ICC fine-tuned retrieval | mean recall: 58.7 -> 71.6 | Improved |
| global-similarity CLIP ViT-L vs Wukong ViT-L on in-domain Wukong-Test | CLIP ViT-L · Wukong-Test fine-tuned retrieval | mean recall: 85.8 -> 89.6 | Mixed |
Closest papers by meaning
- WikiDO: A New Benchmark Evaluating Cross-Modal Retrieval for Vision-Language Models (NeurIPS 2024 · similarity 0.88)
- CVLUE: A New Benchmark Dataset for Chinese Vision-Language Understanding Evaluation (AAAI 2025 · similarity 0.87)
- Large Multilingual Models Pivot Zero-Shot Multimodal Learning across Languages (ICLR 2024 · similarity 0.87)
- Dynamic Adapter with Semantics Disentangling for Cross-lingual Cross-modal Retrieval (AAAI 2025 · similarity 0.87)
- CL2CM: Improving Cross-Lingual Cross-Modal Retrieval via Cross-Lingual Knowledge Transfer (AAAI 2024 · similarity 0.87)
- Are We on the Right Way for Evaluating Large Vision-Language Models? (NeurIPS 2024 · similarity 0.86)
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.
