Hair-GAN: Aligning Portrait Images with Desired Hairstyles
Vu Le Bang Tam, Nguyen Minh Tuan, Chu Quang Linh,
Abstract Past studies point that although recent methods have notably enhanced the intricacies of hair depiction, they frequently yield sub-optimal outputs when the pose of the source image diverges substantially from that of the reference hair image.To address this primary challenge, our focus is to devise an high-performing method for altering hairstyles given in source image and reference image for inputs while ensuring high quality results. We propose a novel approach uti-lizing StyleGAN3 to address this issue. The methodology involves generating a latent vector representation of input source image using the StyleGAN3 encoder, optimizing the latent space, employing InterfaceGAN for hair and pose manipulation, blending and applying StyleGAN3 decoder to generate a new version, refining hair and facial features of this version in final step. Experimental results demonstrate the effectiveness of the proposed approach in achieving desired editing outcomes.
!pip install torch==1.11.0+cu102 torchvision==0.12.0+cu102 -f https://download.pytorch.org/whl/torch_stable.html
!pip install ftfy regex tqdm matplotlib jupyter ipykernel opencv-python scikit-image kornia==0.6.7 face-alignment==1.3.5
!pip install Ninja
!pip install pyrallis
Produce the results:
!python main_edit.py --img_list 07080 07751 --target_list 07845 07154