# Parser-Free Virtual Try-on via Distilling Appearance Flows

**Authors:** Yuying Ge, Yibing Song, Ruimao Zhang, Chongjian Ge, Wei Liu, Ping Luo  
**arXiv:** [2103.04559](https://arxiv.org/abs/2103.04559)  
**Date:** 2021-03-08  
**Summary page:** https://www.summarizepaper.com/en/paper/2103.04559v2/

## Key points

- Traditional methods in the field of garment try-on rely heavily on human parsing, which can lead to unrealistic results with noticeable artifacts if segmentation is inaccurate.
- A recent innovative approach reduces dependence on human parsing by using try-on images generated by a parser-based model to train a "student" network without relying on segmentation.
- The limitation of this approach is that the image quality of the student network is constrained by the performance of the parser-based model.
- The "teacher-tutor-student" technique has been proposed to overcome this limitation and aims to produce highly realistic images without relying on human parsing, offering several advantages over previous approaches.
- This new approach treats fake images generated by the parser-based method as "tutor knowledge" and corrects them using real "teacher knowledge" extracted from actual person images in a self-supervised manner.
- Instead of using real images as direct supervision, the focus is on distilling appearance flows between person and garment images to identify accurate dense correspondences and achieve high-quality results.
- Extensive evaluations have shown significant superiority of this novel approach compared to existing methods, producing more realistic virtual try-on results and offering a robust and accurate solution without heavy reliance on human parsing techniques.

## Summary

In the field of , the goal is to seamlessly fit a garment image onto a person image. Traditional methods rely heavily on , which can lead to unrealistic results with noticeable artifacts if the segmentation is slightly inaccurate. This issue arises because inaccurate parsing can mislead parser-based methods, resulting in visually unrealistic try-on images. To address this challenge, a recent innovative approach utilized  to reduce the dependence on human parsing. In this method, try-on images generated by a parser-based model are used as guidance to train a "student" network without relying on segmentation, enabling the student network to mimic the try-on ability of the parser-based model. However, one limitation of this approach is that the image quality of the student network is constrained by the performance of the parser-based model. To overcome this limitation, a novel technique called "teacher-tutor-student"  has been proposed. This method aims to produce highly realistic images without relying on human parsing and offers several advantages over previous approaches. One key aspect of this new approach is that it treats fake images generated by the parser-based method as "tutor knowledge", which can be corrected using real "teacher knowledge" extracted from actual person images in a self-supervised manner. Additionally, instead of using real images as direct supervision,  in this context focuses on distilling appearance flows between person and garment images. By doing so, accurate dense correspondences between these images can be identified, leading to high-quality results. Extensive evaluations have demonstrated significant superiority of this novel approach compared to existing methods. The refined technique not only produces more realistic virtual try-on results but also offers a more robust and accurate solution for generating visually appealing images without relying heavily on human parsing techniques.