Just realized my AI image generator was recycling faces from its training set

I was making some concept art for a board game last week and noticed two different characters had the exact same nose and eye shape. Ran a reverse image search on one and found a real person's photo from a stock site that matched almost perfectly. Kind of makes me wonder how much of this stuff is actually generative versus just remixing data in clever ways. Has anyone else dug into where their AI art is pulling from?

2 comments

2 Comments

parker_patel8416d ago

Nah I gotta push back on "could just be similar angles and lighting" making it look the same. I actually looked into this pretty deep after I found a generated face that was almost pixel-for-pixel identical to a stock photo of a woman from a 2017 photoshoot. Different lighting, different angle, but the exact same asymmetrical mole placement and scar on the eyebrow. That's not a coincidence. There's been studies and lawsuits showing these models literally memorize and regurgitate specific training images when they're overfitted. The fact that you can reverse image search and get a direct match to a copyrighted photo means the model didn't "learn" to create new faces, it just stored and recombined pieces of existing ones. It's not just pattern matching, it's straight up copying with extra steps.

sanchez.sean16d ago

Is it really that serious though? I mean yeah its kinda interesting that the model is pulling from training data but thats literally how all AI works right now. They're just fancy pattern matching machines trained on whatever they were fed. I get why people get weirded out by it but at the end of the day you're using a tool that was trained on billions of images so some overlap is basically guaranteed lol. Plus reverse image searching faces and finding a stock photo match doesn't necessarily mean it was copying, could just be similar angles and lighting making it look the same.