The final version of my asemic novel _Ahe Thd Yearidy Ti Isa_ (made for NaNoGenMo 2019) is now available, either as a 100MB PDF http://static.decontextualize.com/ahe-thd-yearidy-ti-isa.pdf or as an online image gallery http://static.decontextualize.com/ahe-thd-yearidy-ti-isa-gallery/
it was generated from a suite of GANs trained on bitmaps of random words, which I then sampled from and arranged to look sorta novel-esque. training code: https://github.com/aparrish/word-dcgan novel generation/layout code: https://github.com/aparrish/word-gan-book-generator
using the same models and code I also made a book of latent space interpolations. this isn't 50k words long but I like it and wanted to share. PDF version http://static.decontextualize.com/interpolations.pdf image gallery: http://static.decontextualize.com/interpolations-gallery/
making-of thread starts here, if you're interested: https://friend.camp/web/statuses/103168661447855833
@314 in my experience linguists are very difficult to trick!
@aparrish HIGH FIVE
@aparrish that look so cool
@aparrish Do you have any way of doing a latent space interpolation between "laurel" and "yanny"?
@varx hahaha not with this model, no, unfortunately
@aparrish Ah, I was afraid of that.
Having read up on GANs a tiny bit... if you wanted to find the latent space coordinates that best matched a target, could you use a different discriminator that just knows that target, and have the generator work its way towards that?
(Not asking you to, just curious about the limitations.)
@varx yep it's definitely possible and worth exploring, just not built into the gan code i hacked to make this thing work!
This is amazing. It's such a good approximation of language that if you didn't know where it came from, it'd be a real mystery.
I could believe it was a poor copy of a manuscript in a language I can't read.
@aparrish this is pretty much what the visual symptoms of a migraine look like for me
@aparrish this is beautiful i love it
@aparrish I haven't followed the entire development thread so maybe you touched on this but it's really interesting that the "noise" around the letters is very clearly (to me) a really good approximation of font antialiasing. It looks super crisp zoomed out
@darius the GAN is trained on words drawn with antialiased letters so I think it's learning the antialiasing along with the rest of the shapes. I am pleased with how it looks!
@aparrish I love how the smashing together of words creates these ligatures and letter formations that feel familiar yet are totally unknown... I almost want to trace some and make a conlang out of them!
Hometown is adapted from Mastodon, a decentralized social network with no ads, no corporate surveillance, and ethical design.