The failure
Variant A is a good picture. Nice coat, open hand, the bag reads LL. Nobody on the team would have stopped it.
The QA layer did. The pose detector could not confirm his right knee, and his right upper arm came out at 1.61 times the length of his forearm. Our band for a human arm is 0.7 to 1.6. Small numbers, but this is exactly how a broken arm or a third knee slips into a campaign: it looks fine at a glance.
Why models do this
Image models do not know what a skeleton is. They learned what bodies look like, pixel by pixel, not how they are built. Most of the time that is enough. Then a coat hides a knee, a bag covers a hand, and the model fills the gap with something that is almost right.
The same goes for fingers, faces and logos: text and patterns are drawn as shapes, not spelled. A monogram can drift into something that is close, but not your brand.
How we catch it
Every image runs through the same checks before a person sees it: a pose skeleton, a hand check (is the hand there and open, as the prompt asked), limb ratios, a face check and a logo check on the product. Each check passes, fails or asks a human to look.
The gate is strict on purpose. We would rather block a good image than ship a broken arm. A human can always overrule a block. Nobody can un-ship a campaign.
What it means for you
If you generate images at scale, your bottleneck is not making them, it is trusting them. A QA layer turns "someone looks at everything" into "someone looks at the few that need it".
That is what we build into every system. How we work →