PSA: That AI art demo at SXSW last week made me rethink how I train models

Was at the Austin Convention Center for a workshop on generative image models. The presenter showed a side-by-side where one model was trained on 10,000 generic cat photos and another on just 200 photos of his own cat from different angles. The 200-photo model actually captured more personality and detail. Made me realize we are all overloading our datasets with quantity over quality. Anyone else scaled back their training data size and got better results?

2 comments

2 Comments

skyler_fox7223d ago

200 cats is kind of a small sample size though. I tried it with 50 photos of my dog and the model just kept generating the same blurry shape over and over. Maybe it works if the subject is really consistent but for most stuff you still need a bigger base.

taraj1123d ago

Yeah I tried training on 30 pics of my cat and it just gave me a weird smudge every time.