Hot take: I fed 30 old comic scans and 30 phone photos of my own drawings into two AI tools, and the phone photos won by a mile

Shane Black · 2026-10-09T01:30:03.25356+00:00

I ran this test in my garage over two weekends because I kept hearing people say more training data always means better output. The tool that trained on 30 of my actual pencil sketches and ink pages gave me results that kept my weird hand wobble, while the one working off a pile of old scanned comics gave me the same stiff generic hero pose every single time. The difference came down to one thing: the machine learning on my own work picked up my bad habits, and those bad habits are what makes a drawing look like mine. It's like handing a cook your grandma's 3x5 card versus a random cookbook. So which matters more for real art, hundreds of hours of training data or 30 samples of one person's hand, and has anyone else tried the small batch route?

Opinions Matter

Loading…

Hot take: I fed 30 old comic scans and 30 phone photos of my own drawings into two AI tools, and the phone photos won by a mile | AI Art Debate - Opinions Matter