What 111 Flagged AI Product Photos Reveal About Image Models

Summary

ProductAI used a vision model to compare sellers’ uploaded product photos with generated renders and ask whether the product was still the same. Across 495 checks from 13 accounts over 18 days, it flagged 111 renders, or about 22%; the team manually reviewed every note. The largest category was text or other label changes, accounting for 39 flags: numbers, words, and entire labels could change or disappear while remaining plausible at a glance. Other flags involved color or finish drift, altered shapes, missing components, and renders containing the wrong or no product. About 14% of the flags were judged false alarms, often because the checker did not know that a seller had intentionally opened, held, tilted, styled, or changed an item, or had supplied a mood reference rather than a strict product reference. In the model comparison, GPT Image 2 had the lowest flagged rate in this sample at 11%, compared with 24% for Nano Banana Pro, 25% for Seedream 5 Pro, and 24% for Nano Banana, but the result is only directional because sellers and products differed across models and the sample covered just 13 accounts. ProductAI is testing a checker that reads the prompt and describes the detected change, while the studio now gives contextual tips about readable labels, complete product views, exact colors and finishes, and using one product per generation. The company cautions that the data is uneven, concentrated in personal care and jewelry, and based on a conservative vision-model judge that can still miss errors.