The Reference-Image Editing Test Loop
Treat an image edit like a controlled experiment: lock what must remain, change one thing, and compare the result against visible criteria.
Separate Fixed Details From Requested Changes
Preserve
Identity, product geometry, logo placement, pose, camera angle, crop, or any feature that must not drift.
Change
Name one primary edit: background, garment color, lighting, object removal, material, or visual style.
Blend
Describe contact shadows, reflections, perspective, edge treatment, and color spill needed for a believable result.
Reject
List failure conditions you can see: warped details, altered face, changed proportions, new text, or mismatched light.
Edit the attached image. Preserve [fixed features] exactly. Change only [single requested edit]. Integrate the change using [perspective, lighting, shadow, reflection, and edge instructions]. Keep [crop and composition]. Do not alter [protected details]. Return one clean version with no added text or marks. The edit fails if [visible rejection criteria].
- Save the original and label every run with the model, date, and prompt version.
- Make one controlled change and generate several candidates.
- Score preservation and edit quality separately; a dramatic image can still be a poor edit.
- Revise the weakest instruction block, then rerun against the same reference.
Original Source and License
Inspired by the side-by-side examples in PicoTrex's Awesome Nano Banana Images repository, released under the Apache 2.0 license. This page presents an original testing workflow and does not reproduce repository images.