If your service currently offers options labeled GPT Image 2.5 and Nano Banana 2.1, compare them with a small test of your own deliverable. Write the acceptance rules before seeing either result. A model that makes an attractive landscape may still fail your product label, character identity, or required edit. This guide supplies a comparison method and blank reader materials; it reports no completed generation benchmark or winning model.
Decide what would make a result unusable
Begin with one real brief, such as changing a portrait's coat while preserving the person, or making a product campaign image with an exact approved label. Name the deliverable's required dimensions and the regions that must remain faithful. Mark hard failures before assigning aesthetic scores. A beautiful image with the wrong person or incorrect essential text cannot be rescued by a high average score.
Distinguish a service's menu label from the exact identifier and capabilities it publicly exposes. Imagild's image-model catalog describes available options at the time it is read; it is not an independent performance evaluation or proof that a later task executed on a particular route. Confirm availability in the interface before planning a paid comparison.
Keep the comparison fair enough to be useful
Use the same authorized reference files, intended change, framing, and delivery requirement. If the models expose different settings, record the difference rather than pretending every control is equivalent. Decide beforehand whether you are comparing equal numbers of attempts or equal total spending. Those answer different questions, so do not switch the basis after one option needs more retries.
Keep all attempts in the comparison record, including failures and discarded results. Give files neutral labels such as A and B before asking someone else to review them. Blinding cannot remove every clue about a model, but it can reduce the influence of a familiar name on an initial judgment.
Use the blank worksheet
Download the blank image-model score sheet and its instructions. The sheet has no filled scores or conclusions. It is a reader worksheet, not an Imagild benchmark, certification, or claim of model superiority.
Record identity, requested change, protected regions, text accuracy where relevant, usable output size, and the effort needed to reach an acceptable result. Use your project's priorities. If a criterion does not apply, mark it as not applicable rather than inventing a score. Keep reviewer comments close to the affected image region so another person can understand the rejection.
Make a bounded decision
Choose the option that meets this brief within the budget and hard-failure rules you set. Report the scope: references, date, settings, attempt allowance, and delivery target. A conclusion such as “B was suitable for this authorized portrait task” is more defensible than “B is always better.” Recheck when your project, model version, or available service route changes. Starting with a modest, preplanned sample gives you useful evidence without making a large purchase on the strength of a menu name.