Readme

This project contains simple implementations of Image2StyleGAN and Image2StyleGAN++.

Please note that this is not official implementations and this project is used for a course project.

We first do some exploratory experiments of Image2StyleGAN: we investigate the optimization in latent space and W space.

Then we reproduce some experiments of Image2StyleGan:

• image reconstruction
• morphing
• style transfer

At last we implement a simple Image2StyleGAN++ model, which contains noise optimization and Three blocks: Masked W+ Optimization, Masked Noise Optimization, Masked Style Transfer.

Pretrained StyleGAN model can be downloaded here.

• Image2StyleGAN running command: python execute.py

• Image2StyleGAN++ running command: python execute_v2.py

Exploratory experiments

latent space

We try to reconstruct those four images.

If we use space Z and we get the results as:

This show that space Z fails to represent original pictures.

W space

We use W space to reconstruct the same four pictures, and we get the results shown as:

We can find that for human face pictures we can reconstruct properly but for other pictures it fails to reconstruct them.

Also we can do morphing in W space:

Experiment of Image2StyleGAN

Image reconstruction

We try to reconstruct the same pictures in W+ space and we get:

we can find that not only human face pictures can be reconstructed properly but also other kinds of pictures can be reconstructed.

Morphing

We do some experiments on image morphing:

We find that it basically can work but some babbles may appear.

Style transfer

Experiments of Image2StyleGAN++

Crossover

It basically can work but not very well.