One of the most productive sidequests of this project I have been working on is training this little convolution network I trained to generate 48x48 pixel "layouts" based on a dataset of images. This gets used in the gabor synthesis model I've been working on, but cool on its own