Words that do Handstands

The text is very faint in many places, so I applied some thresholding in GIMP to make it crisper: https://i.imgur.com/cpC3PK9.png

There's probably a way to adapt the training process to create black and white images directly, e.g. by evaluating the classifiers on the thresholded images, but passing the gradients through to the underlying continuously-valued images.