training a gan on bitmaps of words and it's making my eyes hurt (or maybe I just need to go to sleep)

latent space interpolation on a lower resolution version of this model after just 100 batches or so, using matplotlib's default colors because it looks vaguely metroid prime-y

Show thread

after a few thousand batches of training at a usable resolution on an actual GPU. recognizably "words" now—I wonder if increasing the depth of the model (or the kernel size of the convolutions?) would help it actually learn longer-distance dependencies...

Show thread

another change I made was having it train on bitmaps of random words weighted by the frequency of the words in a reference corpus (i.e. in this case spaCy's unigram probabilities). the idea was that this would help it learn higher-frequency letter combinations and generate words that mostly replicate the "look" of English in use (rather than words in a word list). the drawback is that it looks like half the latent space is trying to spell out "the"

Show thread

prototype page layout, sampling each word at random (with the fully trained model, or at least as fully as I care to train it)

Show thread

nanogenmo, gans 

retraining with a serif font instantly makes it seem more ancient manuscript-ey

Show thread

nanogenmo, gans 

definitely bit off more than I could chew when it comes to making something that I feel is conceptually sound with this. the instant temptation is to go full "alien artifact" (and include GAN-generated body horror imagery or whatever), or at least make page layouts that resemble those of typical novels. but then the project feels like it's "about" layout, or "about" books as artifacts, which aren't topics that I personally care to spend time making arguments about at the moment

Show thread
Follow

just realized that if I finish this project I'm going to become one of those people that needs to put that diagram of GAN architecture into their slides. I'm going to find myself explaining how GANs work to someone at a party, dear god

nanogenmo, gans 

had an inkling to train a separate model for words with initial capitals, so I can introduce some structure (like sentences and paragraphs). the drawback here being that it won't have the same latent space as the lower-case model so interpolations won't work across the two. (training a separate model also for words with final punctuation)

Show thread

nanogenmo, gans, weirdly makes me motion sick? 

prototype with capitals and punctuation. you might need to take some dramamine before you try to read(?) this?

Show thread

nanogenmo 

I wish more of this project was more "making weird text things" and less "reinventing typesetting from scratch," but here we are

Show thread

nanogenmo 

(right now I'm just blitting the GAN output straight to Pillow buffers in Python, bc otherwise I would have to write out all of the images to a directory and include them in a LaTeX template or something? and I think I'd rather be lost in the hell of reinventing typesetting than the hell of getting LaTeX to do what I want with tens of thousands of images)

Show thread

nanogenmo, gans 

it does full justification and indentation now! shown here zigzagging through interpolations of the latent space. (the capitalized words and end-of-sentence words are separately trained models, which is why they don't look like the surrounding words)

Show thread

nanogenmo, gans, weirdly makes me motion sick? 

@aparrish Pretty for a moment, then oooogh.

nanogenmo, gans, weirdly makes me motion sick? 

@aparrish this is very cool, and shortly after you realize that, you realize you are becoming rapidly disoriented. amazing!

OCR Output (chars: 1259) 

@varx
(Couldn't find a language with the name 'can', falling back to default.)

Fra ansycls thay sneres. Stout theac vturo of the
thd Bate hiaue there editesy ther ds ma of izoih
mar. itr1 doc and Chan pedi nia anmay shanuy
vral toove noiieh ohek atarnzy anu pec wonc x
nowil men Iiehd nnd vad mamls wei the is
thurek anm vern the mil. Yoone: withg andbiy
wal borlnty wal. oot otanc matte aralug whse
goue mi bery thekt M Had istials bran orerras
doe. the ot niartery aad pear she: Aw yarn issule
icoiny’ isdelb imum. cei she yed svene. tnocay
shetkes cn a® is tame outlets icom arche! Sczmn
soere cermaa so gall gee crue oceir ch twese aodig
tin ort Svianc choces Lrelt beng} muve’ is aaoye
obr: tavrks the: Vi. vor ur latteiy heite hou
daen tair is ale arltre bow arere’ w chtertsy the
want is fit nc ts yew Itvnney sozoe and ig tlad
clid ropn var isrthi mar itoire petrhi crghsy
anens and! Co these amlh and warllly vinms of
cnee sve thenir hie ones Yn. aecne’ het gea
hetht narely mosas met go:l nore wesc bed
thardr acare andi snes over the: Honir or «ar
aga a1 merc thon. shay arekaly be dnr aren ao
that oe! SVite: pr aretite, ceah: shes gmilely cost
m pew nay and to nud me: cninils thore’ witl

aacha crese along herres izurh’ oucsos bid.

Co these amlh and warllly vinms of cnee, indeed.

nanogenmo, gans, weirdly makes me motion sick? 

@aparrish Have you put these through an aggressive-fit OCR?

nanogenmo, gans, weirdly makes me motion sick? 

@a well, I mean, I used mastodon's built-in OCR to caption the image...

nanogenmo, gans, weirdly makes me motion sick? 

@aparrish Oh, I had totally missed the captions! Uh... good idea, then! 😳

nanogenmo, gans, weirdly makes me motion sick? 

@aparrish I like these so much. I keep looking at the early ones, in particular, with my brain waiting for the image to resolve and clear up all the jpeg artifacts.

nanogenmo, gans, weirdly makes me motion sick? 

@aparrish this reminds me uncomfortably of one time when I was reading a PDF in acrobat reader, hit page down, and it rendered the next page in a way not entirely dissimilar from this. It momentarily made me wonder if I was having a stroke or something

nanogenmo, gans, weirdly makes me motion sick? 

@sunward yeah that seems to be the effect that it has on people (on me too). this is the particular phenomenological territory that this piece is exploring apparently

nanogenmo, gans, weirdly makes me motion sick? 

@aparrish This is excellent –Incredible how much this parses as Middle English

Sign in to participate in the conversation
Friend Camp

Hometown is adapted from Mastodon, a decentralized social network with no ads, no corporate surveillance, and ethical design.