How do you compare remix candidates when each changes several things?

Choose the decision before viewing candidates

Write a single selection question, such as whether a low product position leaves enough room for the approved headline. Keep source product, destination and factual copy fixed. Record the requested variation and recognize that generative outputs may introduce other differences. CartVisual does not promise deterministic matching or identical scenes between requests.

For an illustrative bottle campaign, compare two composition requests against the same source. Do not call them a controlled marketing experiment merely because there are two files. Creative selection and measured audience testing are different activities.

Define an external layout exercise before requesting candidates: a 1080 × 1350 canvas with one fixed approved headline area from x=80 to 1000 and y=100 to 320. Keep that same area, exact headline, product source version and destination for every candidate. The coordinates are fictional planning values, not a platform safe-area rule or a measured readability threshold. A candidate whose bottle overlaps the area fails this particular composition brief even if its lighting is preferred. Record a second question separately if you also want to compare lighting.

Record requested and incidental changes

This is an invented review sheet, not observations of generated files. Headline space is the intended layout result; extra props and altered lettering are separate differences. Use visible observations rather than numerical quality scores invented after seeing the images.

Compare at equal presentation scale

Place candidates on the same background at the same intended display size in an external review document. Hide decorative mockups that favor one option. Check product accuracy before composition, because an incorrect product cannot be the winning factual ad merely by leaving more empty space.

Generate another candidate only when the comparison leaves a useful unresolved decision. If the existing image needs a small precise copy correction, use an appropriate external layout workflow instead of changing every scene element again.

Overlay the same headline area and approved copy in your external review document; inspect the whole canvas and the intended display size. Do not independently enlarge each product to fill its frame, because that would hide the product-position difference being compared. For A, remove or reject the unsupported prop before selection; for B, preserve both layout and label failures. For C, compare the label and product with the actual source before writing “confirmed.” If all candidates fail, record “none selected” and request a specific new direction; you do not have to choose the least incorrect factual image.

Pebblely’s tutorial shows that description and reference both shape composition; it does not establish identical outputs from repeated settings. The current CartVisual remix request likewise carries the selected reference and product photos, not a controlled audience-test result. Use a traceable review contact sheet for source/version records without exposing private account data. No files were generated, overlaid or audience-tested for this article.

Keep the reason for selection

Only after C passes the source and copy checks, an illustrative selection note might say: C selected for headline space; source label confirmed; extra prop absent; not an audience-tested result. That note is more informative than selected best. Keep rejected observations too, so the next campaign does not repeat the same ambiguity.

Use creative variant comparison for audience-test planning and feedback resolution when reviewers disagree. This article’s output is a design comparison sheet, not a conversion claim.