The main thing we learned from this is that this a terrible problem to use machine learning for. These are far better, and readable. The algos couldn't even figure out that letters are a bunch of straight lines and neat curves.
I had a similar thought. It's amazing what you can do with machine learning, but this example shows that you can't just throw machine learning at a problem and expect good results.
And it shows that, at least for now, machine learning isn't going to make humans obsolete.
It showed that machine learning is useless if the input format doesn't match the problem. The result would probably be better if the problem was approached more on a vector graphics-based level rather than throwing a bunch of pixel clusters at the program and expecting it to figure out the underlying shapes by itself.
I mean, children in school don't learn to write by copying printed letters out of a book either, instead they're shown the individual strokes and their direction step by step.
Incidentally, to update my comment there, I asked someone who had a font GAN and he mentioned he'd already found uppercase/lowercase latent directions which can automatically "more-uppercase" letters. It's not that this is a 'terrible problem' - it's borderline trivial. (I've seen highschoolers do much more impressive GAN things.) He just didn't do it even remotely right, is all.
He put in way more work than would've been necessary to get StyleGAN or CLIP to do a vastly better job, is my point, because he used totally the wrong tools and approach. (And he could put in vastly more work, and it still wouldn't work that way either; the level of effort one would or would not spend on a SIGBOVIK paper is just irrelevant when his approach is doomed from the beginning.)
When I saw this thread I went looking for the link, couldn't find it. Felt sad. Came to read comments. Saw others have posted it and now I feel content again.
If all research papers were written in this style, I'd read a lot more of them.
> I trained the network using a home-grown (why??) GPUbased package that I wrote for Red i removal with artificial
retina networks [16]—an example of “lowercase i artificial intelligence”—and have improved as I repurposed it for
other projects, such as Color- and piece-blind chess [17]. It is “tried and true” in the sense that “every time I tried
using it, I truly wanted to throw my computer out the window, and retire to a hermitage in the glade whenceforth I
shall nevermore be haunted by a model which has overnight become a sea of infs and NaNs.”