Hair Transfer with Efficient Heuristic Chain of Editing

Yuansheng Ma, Dong Zhang, Suyang Zhu, Shoushan Li

Abstract: The objective of the hair transfer (HT) task is to maintain the texture and style of the reference hair while preserving the facial attributes and background. Previous studies normally try their best to find the optimal latent code in W+ and FS space via thousands of optimization to directly generate a style-swapped image. This schema spends much time (more than 6 mins) for once HT. What’s worse, once these methods produce the final results, they cannot be locally edited to compensate for the obvious flaws in the image after changing the hairstyle. To address the above issues, we approach hair transfer from a new perspective of localized editing and introduce an efficient heuristic chain of editing (HCoE). Our method not only takes only a few seconds to complete the hairstyle transfer, but also can partially modify the primary style-swapped image. Extensive human and automatic evaluations, as well as case analyses, demonstrate the effectiveness of our HCoE.