Shoutout to that dude Dave at the coffee shop who changed my whole view on AI art training data

I was always kinda skeptical about using public domain stuff for training models. Had this convo with Dave, he runs a small graphic design studio near the courthouse here in Austin. He pulled up his phone and showed me a dataset of 1800s botanical illustrations his team had been cleaning up for training. Dude said they got better results on texture generation from those old lithographs than from any modern stock photo set because the imperfections actually taught the model more about variation. That honestly hit me different. He had numbers to back it up. Has anyone else found niche historical datasets that worked way better than expected for specific tasks?

3 comments

3 Comments

robinson.hannah29d ago

And the thing is, those old botanical lithographs have that hand-drawn quality that makes the edges and textures unpredictable. I messed around with some vintage map datasets a while back for a personal project and the model picked up on paper grain and ink bleed patterns that totally changed how it handled color transitions. It was wild seeing it apply that same rough sketchiness to unrelated images without being told to. Makes me wonder what other forgotten archives are just sitting there waiting to be digitized and thrown into a training pipeline.

the_eric28d ago

That bit about "paper grain and ink bleed patterns" really got me. A buddy of mine stumbled into something similar when he scanned a bunch of old field notebooks from the 1920s at a university library. They had these faded pencil sketches of rock formations with coffee stains and folded corners. He tossed them into a style transfer thing just for fun and the AI started adding phantom smudges and dog-eared edges to everything it processed, even text. It was eerie watching digital landscapes come out looking like someone had spilled breakfast on the results.

bennett.vera29d ago

A buddy of mine, he's an archivist at a small historical society in Vermont, found a collection of 19th century seed catalogs. He scanned a few hundred pages of those old watercolor illustrations of fruits and vegetables, the kind with those pearlescent, almost wet-looking highlights. He told me he fed them into some old style transfer code he had lying around and the results came out looking like they were painted on pulp paper with actual berry juice. It freaked him out a little, honestly, seeing digital images adopt that specific physicality of an old document.