What I Learned About AI Product Features From Rejecting My Own Output

I shipped an AI product image enhancer. Then I spent a year building QA workflows that tell users when NOT to use it.

Here's why, and what it taught me about product design.

The feature I almost shipped

The original plan was straightforward: add an "Enhance Image" button. Upload a product photo, get a sharper version. Ship it, collect the feature point.

But during internal testing, I kept hitting the same wall. The enhanced images looked better — sometimes dramatically better — but they were also wrong. A bottle cap gained a subtle rim. A backpack strap shifted shape. A skincare label had text that was close to correct... but not.

"Better but wrong" is worse than "soft but accurate" for ecommerce. A soft image might lose a sale. A wrong image guarantees a return.

What I built instead

Instead of an Enhancement button, I built a SKU-locked pipeline:

1. Product facts get locked before any AI touches the image

2. The prompt lists facts to preserve before asking for quality

3. A 7-check QA gate runs against every output

4. Anything that fails gets rejected, not polished

The output is sometimes less impressive visually. But it's the same product from image to image. That consistency is what buyers actually trust.

The lesson I keep relearning

We track what sellers ask for versus what they actually implement. "More beautiful images" is the most requested feature. But "consistent accurate listings" is what correlates with lower return rates.

The hardest technical challenge isn't making images prettier — it's making them provably faithful to a source. Everything else is decoration.

— — —

Full enhancement workflow with prompt templates: https://loomadesign.ai/en/blog/ai-product-image-enhancer-sku-detail-qa