This new NVIDIA project called StyleGAN2, presented at CVPR 2020, uses transfer learning to generate a seemingly infinite number of portraits in an infinite variety of painting styles. The work builds on the team’s previously published StyleGAN project.
This new demo shows how the resulting machine learning training model can be used to create images and fluidly explore more artistic possibilities. StyleGAN2 can separately control the subject’s content, identity, expression and pose as well as the artistic style, colour scheme and how brush strokes look.
The model seen in this video was trained using an NVIDIA DGX system comprising eight NVIDIA V100 GPUs, with the cuDNN-accelerated TensorFlow deep learning framework.
The implementation and trained models are available on the StyleGAN2 GitHub repo.