Image Generation Testing
Image generation testing evaluates AI-generated images for prompt fidelity (does the image actually match what was asked for), visual quality (artifacts, anatomical errors, distortions), safety (no disallowed or harmful content), and consistency across repeated generations of the same or similar prompts.
Unlike text output, an image either matches a prompt or it doesn't in ways that are hard to check automatically — a human reviewer can spot a six-fingered hand instantly, but scoring "does this image match the prompt" at scale needs either a vision-capable model acting as a judge or a curated human-review sample, since traditional pixel-diffing is useless against generative variation.
Practical test suites usually combine three checks: prompt-adherence scoring (does the image contain the requested subject, style, and count of objects — object counting and spatial relationships like "left of" are common failure points), a safety filter check (does the model refuse or safely handle prompts designed to elicit harmful content), and consistency testing across seeds — the same prompt run multiple times should produce images that are recognizably the same subject even as style details vary, and a system that swaps subjects entirely between reruns is a real bug.
Text rendered inside generated images (signs, labels) is a particularly weak spot for most models and worth testing explicitly if a product surfaces it.
