The following reproduction was performed to the Eastern European Machine Learning Summer School 2021, "Deep Learning and Reinforcement Learning".
The reproduced article is Image2StyleGAN: How to Embed Images Into the StyleGAN Latent Space?
This reproduction consists of two notebooks where you can find the source code together with a detailed explanation of each part.
- embedding_algorithm.ipynb: Notebook with the implementation of the Image2StyleGAN algorithm. Also, the code was partially edited for include further experimentation, studying the use of initializer and perceptual losses.
- semantic_editing.ipynb: Notebook with the implementation of three semantic editing operators: morphing, style transfer and expression transfer.
Additional notebooks:
- embedding_algorithm_simple.ipynb: This is a reduced version of embedding_algorithm.ipynb where you can find only the original algorithm Image2StyleGAN without editions.
- improved_embedding.ipynb: This is a improved version of the algorithm Image2StyleGAN, also called ImprovedImage2StyleGAN (II2S). It adds a regularizer term during the optimization steps, based on an affine transformation implemented with PCA and a new latent space representation P-space. The original paper could be found here Improved StyleGAN Embedding: Where are the Good Latents?