Research map / Vision-language and multimodal / CLIP and zero-shot
Language-Image Pre-training with Long Captions
Kecheng Zheng, Yifei Zhang, Wei Wu, Fan Lu, Shuailei Ma, Xin Jin, Wei Chen, Yujun Shen
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 |
|---|---|---|---|
| pre-train with DreamLIP long-caption objective instead of CLIP at the same data scale | CLIP ViT-B/32 · Flickr30k text retrieval, YFCC15M | R@1: 34.9 -> 84.9 | Improved |
| train on 30M DreamLIP pairs versus CLIP on 400M pairs | CLIP ViT-B/32 · Flickr30k text retrieval R@1 | R@1: 78.7 -> 87.2 | Improved |
| train on 30M DreamLIP pairs versus CLIP on 400M pairs | CLIP ViT-B/32 · zero-shot ImageNet | top-1 accuracy: 62.0 vs 55.7 | Hurt |
| train on 30M DreamLIP pairs versus CLIP on 400M pairs | CLIP ViT-B/32 · semantic segmentation avg | mIoU avg: 34.8 -> 37.3 | Improved |
| uniformly sample sub-captions from long captions as multiple positives (ablation) | DreamLIP · CC3M, Flickr30k text retrieval | R@1: 32.6 -> 63.0 | Improved |
| add short captions to sampled sub-captions (ablation) | DreamLIP · CC3M, Flickr30k text retrieval | R@1: 63.0 -> 68.3 | Improved |
| add subcaption-specific grouping loss (ablation) | DreamLIP · CC3M, Flickr30k text retrieval | R@1: 68.3 -> 69.5 | Improved |
| use stronger MLLM captioner (ShareGPT4V instead of InstructBLIP) (ablation) | DreamLIP · CC3M, Flickr30k text retrieval | R@1: 58.7 -> 69.5 | Improved |
| combine long captions from three MLLMs (ablation) | DreamLIP · CC3M, Flickr30k text retrieval | R@1: 69.5 -> 74.4 | Improved |
| increase sampled sub-captions K beyond about 8 (ablation) | DreamLIP · CC3M, Flickr30k text retrieval | R@1: 70.9 (K=8) -> 69.5 (K=10) | No clear effect |
| increase sampled sub-captions K from 3 to 8 (ablation) | DreamLIP · CC3M, Flickr30k text retrieval | R@1: 65.4 -> 70.9 | Improved |
| use DreamLIP encoder in LLaVA-1.5 instead of CLIP at same scale | LLaVA-1.5 vision encoder · ScienceQA-IMG, TextVQA, POPE, MMVP avg | avg score: CC3M avg 51.8 -> 52.4; ScienceQA-IMG 65.7 -> 65.1 | Mixed |
Closest papers by meaning
- LoTLIP: Improving Language-Image Pre-training for Long Text Understanding (NeurIPS 2024 · similarity 0.93)
- FLAME: Frozen Large Language Models Enable Data-Efficient Language-Image Pre-training (CVPR 2025 · similarity 0.91)
- Modeling Caption Diversity in Contrastive Vision-Language Pretraining (ICML 2024 · similarity 0.90)
- Semantic and Expressive Variations in Image Captions Across Languages (CVPR 2025 · similarity 0.90)
- SmartCLIP: Modular Vision-language Alignment with Identification Guarantees (CVPR 2025 · similarity 0.89)
- Probabilistic Language-Image Pre-Training (ICLR 2025 · similarity 0.89)
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.
