# HairCLIP: Design Your Hair by Text and Reference Image

**Authors:** Tianyi Wei, Dongdong Chen, Wenbo Zhou, Jing Liao, Zhentao Tan, Lu Yuan, Weiming Zhang, Nenghai Yu  
**arXiv:** [2112.05142](https://arxiv.org/abs/2112.05142)  
**Date:** 2021-12-09  
**Summary page:** https://www.summarizepaper.com/en/paper/2112.05142v2/

## Key points

- The paper proposes a new hair editing interaction mode that allows users to manipulate hair attributes based on texts or reference images.
- This approach eliminates the need for well-drawn sketches or masks as conditional inputs for editing, which is required by many existing methods.
- The proposed method encodes image and text conditions in a shared embedding space and leverages the powerful image-text representation capability of the Contrastive Language-Image Pre-Training (CLIP) model to propose a unified hair editing framework.
- The framework performs high-quality hair editing in a disentangled manner through carefully designed network structures and loss functions.
- Extensive experiments were conducted to demonstrate the superiority of their approach in terms of manipulation accuracy, visual realism of editing results, and preservation of irrelevant attributes.
- A project repository is provided for their work.
- This paper addresses an interesting and challenging problem in computer vision and graphics, as hair editing has been traditionally difficult due to its complex structure and texture.
- The proposed method offers an efficient way for users to edit hair attributes without requiring them to have advanced drawing skills or spend significant time on interaction processes.

## Summary

The paper "HairCLIP: Design Your Hair by Text and Reference Image" proposes a new hair editing interaction mode that enables users to manipulate hair attributes individually or jointly based on texts or reference images. This approach frees users from the tedious process of providing well-drawn sketches or masks as conditional inputs for editing, which is required by many existing methods. The proposed method encodes image and text conditions in a shared embedding space and leverages the powerful image-text representation capability of the Contrastive Language-Image Pre-Training (CLIP) model to propose a unified hair editing framework. The framework performs high-quality hair editing in a disentangled manner through carefully designed network structures and loss functions. The authors conducted extensive experiments to demonstrate the superiority of their approach in terms of manipulation accuracy, visual realism of editing results, and preservation of irrelevant attributes. They also provide a project repository for their work. This paper addresses an interesting and challenging problem in computer vision and graphics, as hair editing has been traditionally difficult due to its complex structure and texture. The proposed method offers an efficient way for users to edit hair attributes without requiring them to have advanced drawing skills or spend significant time on interaction processes. Overall, this work contributes towards making hair editing more accessible and user-friendly while achieving high-quality results.