my wardrobe app got its visual layer and then i noticed the closet was full of lies. a tile labeled black sunglasses showed a watch strap. blue gloves showed a face. and the ai product thumbs were worse. a perfectly good headphones crop regenerated into a gray t-shirt, and a watch crop became blue plaid fabric. every single thumbnail had to be checked against its label by eye, which is not something a person should have to do for half a thousand items.
built an audit loop that does what that human eyeball was about to do. a vision model verified every main tile, every member crop and every thumbnail against what it was supposed to be, deleted wrong and empty crops, promoted the best lookalike when the main was wrong, and flagged thumbs for regeneration. the regeneration prompts named the exact item type and then re-verified the output before accepting it, with a guard that a description had to actually contain the right word.
that loop converged in passes. delete the junk, regen the thumb, verify, repeat. accessories went from a mess to every tile matching. bottoms and tops took more laps because the image model kept inventing sweaters and recolors for plain jeans, and each one got verified again before it was trusted. eventually it bottomed out on a handful of correct items that just have thin framing because the source photo offers nothing better.
the whole pass cost almost nothing and the real lesson is that any pipeline making images for you needs a verification stage, or the hallucination just compounds into the next layer. a model that checks the previous model is the entire trick, and it works until the checker starts rubber stamping slivers, at which point you make the checker stricter and run it again.