fishGAN - StyleGAN2 Applied to Fish

Check out my results at This Fish Does Not Exist.

This repo contains files used to scrape 60k images of fish from online and use them to train the StyleGAN neural network to generate images of realistic-looking fish.

Relies on stylegan2-pytorch and lightweight-stylegan2 for the actual StyleGAN2 model architecture.

Running yourself

Collecting data

1. Create conda environment fish_env

conda create -f environment.yml

2. Run Python script get_fish_pics.py. You can modify the script to decreaes the amount of images it saves/processes. The full script took several hours to run on a SLURM cluster.

python3 get_fish_pics.py

3. The raw fish images will be stored at fish_pics/ and the resized (square, zero-padded) images will be in resized_fish_pics/.

Running models

See the section Deploying on AWS in the StyleGAN2 implementation to see how to run the models on AWS.

Results

See the best results at This Fish Does Not Exist. Below are some more!

"Lightweight" StyleGAN2 after 150k iterations (128x128)

StyleGAN2 after 131k iterations (128x128)

StyleGAN2 after 48k iterations (256x256)

These are the images shown at thisfishdoesnotexist.net.

Website setup

1. 
2. 
3. Create a virtualenv on EC2 instance at ~/fishGAN/venv and install flask. Use command: virtualenv venv.
4. . Note: instead of sudo apt-get install libapache2-mod-wsgi, use sudo apt-get install libapache2-mod-wsgi-py3.