Test AI-generated product images for text errors with automated checks, not visual review

The author argues that visually eyeballing AI-generated product artwork misses errors like wrong prices, which customers may spot first, and proposes treating renders like build artifacts. They ran identical prompts through GPT Image 2 and Nano Banana Pro on three e-commerce scenes, with 14 required text strings, and both models rendered all of them correctly. The main differences were speed, with Nano Banana Pro about four times faster, and file formats, with GPT Image 2 producing larger PNGs and Nano Banana Pro producing smaller JPEGs. The author recommends OCR-based assertions against declared required and forbidden strings, which can also detect regressions when a provider silently changes a model. The piece also covers retry classification, request timeouts, and editing features of each model.