Found a weird trick for training AI image generators to stop making extra fingers

I was messing around with Stable Diffusion for my moving company's website graphics and kept getting hands with 6 or 7 fingers. Totally useless for a professional ad. I tried adding 'correct anatomy' to negative prompts but it did nothing. Then I found a community dataset that focused on hand photos with captions describing finger counts, and after training on about 50 of those images the output cleaned up completely. Has anyone else found a specific dataset that fixed a weird artifact like this?

2 comments

2 Comments

jordan18412d ago

Nah @jason73 I gotta disagree with you there. The captions matter but a dataset with specific hand photos works way better than just rewording your prompts. Those hand-focused sets force the model to actually learn finger patterns instead of just guessing. I tried the caption trick first and it still gave me six fingers half the time.

jason7312d ago

Did you try training on something wider than just hands? Cause the real fix is usually in your captions not the images. A lot of these datasets are garbage because people just tag 'hand' and don't specify anything else. Like your moving company pics, if you caption each image with 'man holding box with 5 fingers' instead of just 'hand' the model actually learns the right count. I've seen the same trick fix weird eye spacing and messed up teeth too. The key is being super specific with what you want the model to unlearn.