Spatially-invariant Style-codes Controlled Makeup Transfer (original) (raw)

Transferring makeup from misaligned reference images poses significant challenges that have previously relied on pixel-wise correspondences, leading to inaccuracies and high computational costs. This work proposes a Style-based Controllable GAN (SCGAN) model that operates via a two-step extraction-assignment process. By employing a Part-specific Style Encoder, the model encodes the makeup style into a spatially invariant style-code, enabling flexible makeup editing and robust results even amid spatial misalignment. The framework supports multiple operations including makeup removal and shade-controllable transfers, demonstrating enhanced flexibility compared to existing methods.