sangenan / MakeupEmbeddingGAN-tf-Implement

MakeupEmbeddingGAN-tf-Implement

• A tensorflow Implement for MakeupTransfer & MakeupEmbedding based on cycleGAN and ZM-Net

Model Design

• Using AdaIN in parallel Pnet(StylePredictingNet) and Tnet(StyleTransferNet) Architecture
• Learning How-to-Transfer-a-Given-Style in Tnet & Learning What-is-a-Given-Style in Pnet

Model Architecture

Generator Architecture

Discriminator Architecture

• Same 70*70 patch discriminator as cycleGAN

Loss Function

Following beautyGAN's setup which includes:

• adversarial loss
• cycle loss
• makeup loss (histogram match loss on face&lip&eye shadow)
• perceptual loss (calculate on relu4_1 of pretrained VGG16 model)

Results

MakeupTransfer Results

Embedding Learned from Pnet Test

Style Fusion