Generative adverserial neural network for metal band logo generation
This project is a personal project to test my capabilities for implementing a GAN. I chose to try and generate metal band logos for several reasons.
- I personally favour metal music.
- I have abundunt data (~8k images). More would be better but I have found examples of impressive GANs trained on less than 1k images.
- The logo images represent a good distibution for a GAN:
- they are mostly grey-scale (or monochrome and thus easily converted to grey-scale)
- they retain meaning when resized
- the can be quite hard to read, which makes the generator's job of creating believable images much easier
I am running most of the training on my personal computer, thus I have to keep to as little computing power as possible. The implementation is mostly based on stackGAN. currently this is a work in progress
First results with 64x64 images were not too impressive, but they proved that I am on the right track.
For the second version I added a higher definition block, and added progressive training of different layers.
Images are improving, but training is becoming much more time-expensive the higher the resolution.
I am currently tring to add a character based encoder, so that the GAN will be able to create a logo for a given name.