Generating "Spot the Difference" Puzzles with AI
kamens.comNice article! I’ve experimented a bit with autogenerating “Where’s Waldo?”-style images. Even models that can output higher resolutions (Seedream can do 4K) tend to generate faces that look like they’ve been shoved into a fireplace, like Sandor Clegane.
This is where something like ADetailer (YOLO + Img2Img) really feels necessary to clean up all the finer details but it would probably take a lot of manual tweaking.
I agree. There’s another guy (in the quote tweet) here who’s pushed on Where’s Waldo stuff, but like you I think it’s currently stuck at the “deformed bodies/faces” issue: https://x.com/kamens/status/2001396716654727607
I also suspect it may be solvable by switching to something other than humans - we probably won’t be as weirded out by malformed cars or plants or whatever.