Can this be using to eliminate badcases?
just read the paper, great job! I was wondering if this could be a way to auto-detect (or auto-correct) badcases such as anatomy errors, limb/hand problem. Don't know if that's something on the radar? would be wonderful to try it out
Hello @Jaylom95 , thank you for your interest in our paper!
I think it would be an interesting approach to use MaPO for resolving generation errors!
The key strength of MaPO is its versatility and data efficiency, where you can curate any offline preference dataset according to your own standard of preference, and indeed, "limb/hand error" can be defined as preference.
We can think of formulating the binary preference data with "chosen" images as the ones with proper limb/hands and "rejected" images as the ones with limb/hand errors.
As the base SDXL frequently suffers from limb/hand errors, making a good "chosen" image set would require superior models, like DALLE-3.
Again, I think it would be an interesting approach!
yes, that's exactly what i mean. does this image set have to be newly generated? or the already existing ones can be used? cause i'm sure there are tons of good sdxl(even sd1.5) good cases out there with good limb/hand.
The chosen and rejected images should share a similar context since we give SDXL the same prompt for both images.
If you could pair the images with a single prompt, you can use the existing good and bad cases!