VividHairEdit: Disentangled Latent Control for High‐Fidelity Hairstyle Transfer via StyleGAN2 Inversion

Abstract VividHairEdit advances the field of realistic hairstyle manipulation through an innovative GAN‐inversion approach. Although recent deep‐generative models have made significant strides in hair editing, challenges persist in accurately reflecting user preferences and representing complex hair features. Our system addresses these issues by introducing two key improvements: structural integration with refined mask completion and improved latent code optimisation for vivid hairstyle representation. The structural integration with the refined mask completion technique significantly improves the semantic alignment between source and reference images, enabling more natural hair transfer with reduced blending artefacts. Our improved latent code optimisation, incorporating masked perceptual and style components, effectively captures and represents detailed hair features, including vivid colours and intricate strand textures. Additionally, we introduce user‐controllable components that seamlessly integrate with our enhanced hair transfer process, allowing for intuitive and precise editing. Our system not only excels in hair transfer but also provides a user‐friendly sketch interface for further refinement, leveraging the improved transfer capabilities. Quantitative and qualitative evaluations, including user preference studies, demonstrate that VividHairEdit outperforms current state‐of‐the‐art techniques in both hairstyle generation quality and usability, marking a significant advancement in GAN‐based hair manipulation.