This is a neural network for generating images using the GAN (Generative Advisary Network) arcitechture. It is initially setup for creating more texture-like images where some cropping, rotation and filtering can be used to increase the amount training data (input images).
-
Put your real images in the
inputfolder -
Train a GAN using:
python train.py network_name [iterations] -
Generate textures using:
python generate.py network_name [num_images]
- Python 3
- Tensorflow
- Pillow
- Numpy
- Fix the stalled learning
- only small improvement beyond 20 000 iterations (mostly shuffling)
- Better optimization?
- Better cost function?
- Better network configuration?
- Try bigger sizes than 64x64
- maybe through an upscaling GAN
- Try making repeatable textures