robinson.jake•1mo agoMost Upvoted

Found a hack for training AI image models on just 15 reference photos

I was trying to make a custom AI portrait model of my dog for months but kept getting these weird results because I only had like 15 good photos of him. Every tutorial I found said you need at least 50 to 100 images for decent training. I stumbled on this method where you use a technique called "textual inversion" combined with a pre-trained model instead of starting from scratch. It took me about 2 hours to set up on my home PC and the first batch of generated images actually looked like my dog, not some mutant creature. The secret was picking photos with different backgrounds and lighting instead of just similar poses. Has anyone else gotten good results using way less training data than the guides recommend?

2 comments

2 Comments

cole_patel411mo ago

This kinda tracks with how most "rules" work in my experience. Everyone says you need a massive dataset or a whole studio setup, but it's usually about quality over quantity. I've seen the same thing with learning guitar - people obsess over having the right gear or practicing 5 hours a day, but the real trick is just picking a few good habits and sticking with them. The different backgrounds and lighting thing makes sense too, that's basically forcing the model to learn the actual subject instead of memorizing one specific couch.

the_ben1mo ago

Honestly I used to buy into the 'more data is better' thing too but this flipped my thinking completely.