# FashionSAP: Symbols and Attributes Prompt for Fine-grained Fashion Vision-Language Pre-training

**Authors:** Yunpeng Han, Lisai Zhang, Qingcai Chen, Zhijian Chen, Zhonghua Li, Jianxin Yang, Zhao Cao  
**arXiv:** [2304.05051](https://arxiv.org/abs/2304.05051)  
**Date:** 2023-04-11  
**Summary page:** https://www.summarizepaper.com/en/paper/2304.05051v1/

## Key points

- Fashion vision-language pre-training models are effective in downstream tasks but often overlook fine-grained domain features
- Introduction of Fashion Symbols and Attributes Prompt (FashionSAP) for fine-grained fashion vision-language pre-training
- FashionSAP captures intricate multi-modal fashion attributes using fashion symbols and attribute prompt method
- Extensive experiments show state-of-the-art performance across four popular fashion tasks with FashionSAP
- Abstract fashion symbols and attribute prompt method help model grasp fine-grained semantics within the fashion domain
- Importance of considering fine-grained domain features in vision-language pre-training models highlighted
- FashionSAP establishes a new benchmark for future research and advances understanding of complex multi-modalities in fashion industry

## Summary

Fashion vision-language pre-training models have shown effectiveness in various downstream tasks. However, they often overlook fine-grained domain features crucial for distinguishing specific domain tasks from general ones. To address this gap, we introduce a novel approach for fine-grained fashion vision-language pre-training known as Fashion Symbols and Attributes Prompt (FashionSAP). This method captures intricate multi-modal fashion attributes and characteristics by incorporating fashion symbols and an attribute prompt method. Extensive experiments on two public fashion benchmarks demonstrate that FashionSAP achieves state-of-the-art performances across four popular fashion tasks. The incorporation of abstract fashion symbols and the attribute prompt method enables the model to efficiently grasp fine-grained semantics within the fashion domain. These results not only establish a new benchmark for future research but also highlight the importance of considering fine-grained domain features in vision-language pre-training models. This refined approach holds promise for advancing advancements in understanding and modeling complex multi-modalities within the realm of fashion.