· 3 min read

Language arts notes

Fragment from the production of Language Arts, a series of generative adversarial networks trained to interpolate a constrained visual space of text.


Font: Code2000 (an old freeware font with good Unicode coverage) Architecture: Stylegan2-pytorch, with attention, AMP, and augmentation. Script: I rebuilt the script from scratch because it was the worst Python code I’ve literally ever worked with. And I’ve worked with my own code.

7 hours of training [source]

Oh, this is 7 hours of training btw.

Haha, I found how to slow it down and now it’s l i q u i d i n k.

36 hours of training - liquid ink [source]

36 hours of training. Slowed to 1/4 speed from the one at the OP.

57 hours - halfway through first training session [source]

57 hours. Halfway through the first training session, but it should go faster from here because I adjusted some parameters (back to the defaults lol).

62 hours - shapes becoming beautiful [source]

62ish. The shapes are now so pretty that I just want to watch one at a time.

100 hours - learning new letters, developing whitespace modes [source]

100 hours. It keeps learning new letters, but it’s also developing a bunch of “whitespace” modes. I think this is from the image augmentation? Maybe I will turn that off to kind of polish it here at the end of the first training run.

Reduced whitespace - soothing magic ink effect [source]

It worked! At least, it seems like it’s producing fewer whitespace characters. But I do kind of like the whitespace in the interpolations. It looks like a blank page with magic ink appearing on it. GF says this one is “Soothing.”

After removing whitespace and punctuation marks [source]

It wasn’t just the image augmentation. There were like five images in there that were literally whitespace, and quite a few that were super tiny. So I removed the whitespace and like 600 little punctuation marks; hopefully now it’ll learn some of the more complex characters.

This is definitely working: more complicated characters, less whitespace. The color blobs and noise textures are because I’ve caught it at a phase where the generator is learning something new. 10 hrs of training (with a high gradient-accumulate-every setting) on top of the old net.

Learning complex characters - 10 hours additional training [source]

View original