The network I was training on sigils overfitted (it can only produce pages from the original document), so I’m starting over from the beginning. Going to train a net on all the Unicode symbols to get a continuous interpolation space, then finetune on various sigils. 20 hrs so far.
It’s pixelated because I’m using a public domain bitmap font called Unscii: http://pelulamu.net/unscii/
It’s colorful because I generated the training data with Robert Munro’s script from https://towardsdatascience.com/creating-new-scripts-with-stylegan-c16473a50fd0 — using Unscii instead of Arial has a more “computer” feel, I think.
I’ll probably train it on an anti‑aliased version of the same characters down the road. Not Arial though, jezus.
48hrs and it’s developed a “ghost pixels” mode
I think this means the generator is cheating, finding the minimum amount of color that will fool the discriminator—fingerprinting the pixelated look of the glyphs. Hopefully that comes out in the wash.
If you don’t see what I mean, look really closely at the squares when they go all white.
68 hours. It’s doing proper letter shapes now, so I think it’s time to switch datasets.
Made a Noto font version of the same thing; should learn some more refined glyph shapes. Example CJK character:
I put the laptop in the closet, fans running at top speed. I whisper to it, “I’ll be back soon, kid,” as I leave for work. It studies duteously.
After 8 hours of training on Noto characters. Did you know Noto means “no tofu”? Tofu are the little rectangles that show up when you don’t have a certain character in your font. Fun.
It’s cool how it looks like worms—developing some more “organic” shapes now.
Well, I added color augmentation and it started generating evil tofu. Going to switch to StyleGAN‑2 and start again with a better dataset.