A implementation of StyleGAN using Tensorflow
This implementation is on experiment and has many issues.
official paper: https://arxiv.org/abs/1812.04948
official implementation: https://github.com/NVlabs/stylegan
This code is based on my PGGAN implementation.
-
Download ffhq-dataset from here.
-
Put images1024x1024 in ffhq_dataset and thumbnails128x128 in ffhq_dataset128.
like this
...
│
├── ffhq_dataset
│ ├── 00000.png
│ ├── 00001.png
│ ├── ...
│ └── 69999.png
├── ffhq_dataset128
│ ├── 00000.png
│ ├── 00001.png
│ ├── ...
│ └── 69999.png
├── main.py
├── model.py
...
- Train StyleGAN.
python main.py
How long does it take to train using RTX 2070,
64x64 1d00h
128x128 2d00h
256x256 3d18h
512x512 6d13h(estimated)
1024x1024 unknown
- After training, inference can be performed.
to draw uncurated images,
python pred.py -m uc
to draw truncation trick images,
python pred.py -m tt
to draw style mixing images,
python pred.py -m sm
uncurated
truncation trick
style mixing
As mentioned at the beginning, this implementation has problems.
Diversity and quality are too sensitive to 1st noise.
The above figure is from the left
Use all noise
without 1st layer noise
without 1st and 2nd layers
without all noise
Because of this, style mixing and truncation trick cannot use 1st noise.
Disabling 1st noise results slight improvement in quality and loss of diversity.
below images are style mixing and trucation trick using noise of 1st layer.
At the beginning of the stage, the kernels are unlearned, so interpolation is not performed well, this leads to quality degradation?
There are some other problems and I will continue experiments.