A Reproducible Test Card for AI Image Edits

Owen Carter — Fri, 25 Sep 2026 08:55:02 +0000

A convincing AI-edited image does not tell you whether the editor followed the brief. Did it change the requested outfit but quietly alter a face or crop? For a repeatable review, save the input, the exact request, the actual output, and what you can and cannot verify. This is a blank test-card template, not a reported model test.

1. Define the task before generating

Write two lists. Must remain might include subject count, facial features, product shape, location, and crop. May change might include clothing, lighting, background, and finish. These are task-dependent: a portrait restyle may allow a new setting, while a product cutout should preserve its silhouette.

Give the task a stable case ID. Give each attempt its own run ID. Do not overwrite an older output when a later one looks better.

2. Store a record you can actually check

The following JSON is a schema example with placeholders, not an actual run. Fill the values only after you generate and inspect an output.

{
  "case_id": "[stable task ID]",
  "run_id": "[unique attempt ID]",
  "tool": "[tool actually used]",
  "model": "[model shown by that tool, if any]",
  "date": "[actual run date]",
  "input_file": "[source image path or private reference]",
  "output_file": "[returned image path]",
  "exact_prompt": "[unaltered prompt]",
  "must_remain": ["[requirement A]", "[requirement B]"],
  "may_change": ["[requested change A]"],
  "observations": []
}

Keep the source and output together where you have permission to store them. If the images contain people or private work, a public post may need redacted or permission-cleared examples. File paths in a record are pointers to evidence, not proof by themselves.

3. Review in a fixed order

First check the number and placement of subjects. Then inspect faces, hands, hair, thin edges, labels, reflections, and small text at full size. Finally review background, crop, lighting, and style. A thumbnail is useful for overall impression but can hide the exact error you need to find.

For each requirement, record one of four plain-language statuses: kept, changed as requested, changed unexpectedly, or unclear. “Keep both faces” is a request, not an observed result. Mark a face unclear when the source or output is too small to compare confidently. Do not turn a visual glance into a biometric identity claim.

4. Make the next attempt diagnostic

If the background disappeared, make the location-preservation requirement explicit and remove a conflicting backdrop instruction. If a face drifted, try a clearer source and fewer simultaneous appearance changes. If a product edge looks wrong, inspect it against both light and dark backgrounds. These are hypotheses to test, not guaranteed fixes. Save each new output under a new run ID.

A note such as “wardrobe changed as requested; crop removed a hand” is more useful than “version two is better.” It tells the next reviewer exactly what to inspect.

Scope

This test card helps with prompt iteration, design review, and regression discussions. It does not rank models, certify safety, or establish a success rate. If you only need color correction or grain, a conventional editor may be simpler to audit.