Hit 10,000 AI-generated images and finally saw the pattern

I've been messing with image gen models since last spring for fun, mostly making fake album covers. Last week I hit 10,000 generations on my local setup and something clicked. The weird hands thing? It's not random. Roughly 1 in 37 images has a hand issue, but only when the prompt mentions hands. If hands aren't in the prompt, it's like 1 in 200. I counted over 3 days using a simple script. That made me realize the model is learning what to expect, not what to draw. Has anyone else noticed these weird breakdowns by prompt category, or am I just way too deep in this rabbit hole?

2 comments

2 Comments

kim_davis4d ago

Honestly, I see it kinda different. You ran a script counting hands, but that's one slice of a huge pie. Tbh, I've done similar tests with faces and lighting, and the pattern I get is more about training data bias than some deep "expectation" thing. Like, when I prompt "hands holding a cup," the model doesn't know what hands are, it just matches pixels to a huge pile of photos where that combo shows up. Your 1 in 37 vs 1 in 200 stat could just be because hands are way more often in the training set when they're mentioned, so the model takes more shots and misses more. Ngl, I think you're seeing a pattern because you're looking for one. Try the same count with "cats" or "cars" next time, you'll probably get a similar weird ratio.

fionat554d ago

Wait, you actually counted 10,000 images by hand with a script and tracked every single one over three days? That's some serious dedication, I gotta give you that. But hold on, if the hand issue jumps that much just from mentioning them in the prompt, that feels like the model is doing something way more sneaky than just matching pixels from a photo pile. Your cat and car test sounds like the perfect next step, and honestly I'm kinda hoping you run it so we can see if this weird ratio holds up.