Addressing pose variance in reference-based hairstyle transfer for human face image editing

Abstract

This dissertation explores innovative hairstyle transfer technologies that effectively reflect personal identity and style. We present novel methods to tackle pose variance and image quality challenges in hairstyle transfer. Hairstyle transfer is a complex task that modifies a face image’s hairstyle while preserving key features. First, we introduce the K-hairstyle dataset, a large-scale collection of 500,000 high-resolution images annotated by expert hairstylists, serving as a foundation for pose-invariant hairstyle transfer. We then propose the HairFIT framework, which uses flow-based motion transfer for hairstyle alignment and synthesis. Image quality is further enhanced through latent optimization with a pre-trained StyleGAN2 model. Finally, we develop HairFusion, a diffusion-based model for real-world applications that employs hair align cross-attention for precise hairstyle alignment and adaptive hair blending to maintain the integrity of the original image. Our methods not only enhance the effectiveness of hairstyle transfer but also open new avenues for research in generative models and image editing technologies.