patricia32•16d agoProlific Poster

I was training AI on my own sketches and wondered why the outputs looked flat, then a friend pointed out I was only feeding it 12 pieces of reference art

After 3 months of wondering why my AI stuff had no soul I realized I was starving the model by giving it such a tiny batch of my own work instead of letting it learn from a big varied set of styles, anyone else accidentally limit their datasets like that?

3 comments

3 Comments

spencer_gonzalez116d ago

Did you try mixing in some public domain art from different eras too? I had the same problem with a style transfer project and fixing it was as simple as grabbing 50 random charcoal sketches from an old museum archive and 30 ink wash paintings from another source. The model started making stuff that actually looked like it had some life behind it instead of just copying my own bad habits.

stellat4616d ago

Throw some random old sketches in there and see what happens.

phoenix_martin4016d agoTop Commenter

Ngl mixing in random old sketches actually sounds like a solid fix? Like throwing in some messy charcoal studies from the 1800s or faded ink wash landscapes could break the model out of your own predictable habits. Ive seen people dump in random retro anime cells or vintage newspaper comics and suddenly the outputs get way more interesting. Its like feeding it a bunch of weird ingredients so it stops making boring soup.

-1