Maybe worth trying to train a better style for this. This is probably something where you could put a little effort in up-front (ie: using a model that's for segmentation to get outlines, using some classic image-processing for boundary detection) and then have AI touch it up a little more lightly and a less of the "default" style.
Also, do you have AI images for the "real world" samples on the left? They have a certain "I don't exactly know what, but it's creeping me out" vibe.