Image generation with Tensorflow 2 and GANs

This project's goal is to generate images using existing images like an artist.
Tested with Tensorflow 2 and GANs (Generative Adversarial Networks), (1080x773px, 3 colors) images on Nvidia 2080 GPU, VScode.

How to use it

Copy all pictures you want to use in the input directory of this project (tested with pictures of the same size : 1080x773px, 3 colors).
Change parameters in main.py and GAN.py if needed (imageHeight, imageWidth, redimRatio, dpi, percentageOfImagesToKeep, imagesPerIteration, batch_size, sample_interval, epochs).

python main.py

The result

In the resized directory : generate 4 resized images.

In the output directory :

• Generate 1 image for every sample_interval epochs.
• Generate 1 image called output.png wich name never changes, so you can see it in live.
• Generate 1 GIF after all epochs.

Parameters to change

percentageOfImagesToKeep : Percentage of images to keep in the input directory. If 20 and if you have 3000 images, it 'll keep 20% of 3000 images : 600.
imagesPerIteration : 3 will generate 1 big image containing 3x3 pictures per iteration (so 9 pictures).
sample_interval : 20 means generate an output image every 20 epochs.
batch_size : Batch size in the main GAN loop. Bigger means faster but it may give bad results.
epochs : Number of epoch in the main GAN loop. I observe Better results when the number of epoch is not that big (50 to 400).

imageHeight : Change it one time, the pixel height.
imageWidth : Change it one time, the pixel width.
redimRatio : 4 means the resized image will be (imageHeight / 4, imageWidth / 4). Used to reduce memory and go faster.
dpi : the size of the final picture. Higher means better image quality but it takes longer to create the files.

Using pokemon cards

Links to go further

Inspired by : https://skymind.ai/wiki/generative-adversarial-network-gan