PSGAN++: Robust Detail-Preserving Makeup Transfer and Removal

Area of Science:

• Computer Vision
• Artificial Intelligence
• Image Processing

Background:

• Existing makeup transfer methods struggle with large pose/expression differences and fine details.
• Controlling makeup degree or transferring specific parts remains challenging for current approaches.

Purpose of the Study:

• To develop a robust method for both makeup transfer and removal that overcomes limitations of existing techniques.
• To enable detail-preserving makeup transfer and effective makeup removal, even in unconstrained scenarios.

Main Methods:

• Proposed PSGAN++ (Pose and expression robust Spatial-aware GAN) utilizing a Makeup Distill Network (MDNet) and Attentive Makeup Morphing (AMM).
• Employed an Identity Distill Network (IDNet) for makeup removal and a Style Transfer Network (STNet) for feature map editing.
• Introduced new datasets: Makeup Transfer In the Wild (MT-Wild) and Makeup Transfer High-Resolution (MT-HR).

Main Results:

• PSGAN++ achieves state-of-the-art performance in detail-preserving makeup transfer and removal.
• The method successfully handles large pose and expression variations, maintaining fine makeup details.
• Demonstrated capability for partial and degree-controllable makeup transfer.

Conclusions:

• PSGAN++ offers a significant advancement in makeup transfer and removal, addressing key limitations of prior work.
• The proposed model is robust to variations in pose and expression, suitable for real-world applications.
• The release of PSGAN++ code and datasets will facilitate further research in virtual makeup technologies.