Two candle clips sit side by side. The first is square and five seconds long. The second runs for ten seconds, uses a vertical format, and follows a different camera instruction. It may look smoother, but the comparison tells us little because the task changed too.

When comparing models in ZMS AI, keep the creative brief steady and change only the model. This makes the differences easier to see and explain.

Lock the basics first

Write one brief before generating anything. For the candle clip, keep these details fixed:

Product image

Candle and flame movement

Setting and lighting

Camera direction

Duration, aspect ratio, and resolution where available

Copy the brief without rewriting it for each model. Even a small wording change can affect the result.

If two models do not offer matching settings, record the difference. The comparison can still be useful, but it is not exact.

Change only the model

Do not add another reference because the first result looks weak. Do not improve the prompt before generating the second version.

Give both models one product image and keep this instruction unchanged:

The camera moves slowly toward the candle while the flame reacts to a light draft.

New references and revised prompts may improve the final clip, but they belong in a later test. For this round, the goal is to see how each model handles the same request.

Use a short review list

“This one looks better” is difficult to explain. A short checklist makes the decision more useful:

Did the clip follow the requested action?

Did the candle keep its shape and label?

Was the camera movement close to the brief?

Did the lighting remain stable?

Were there visible glitches?

How much editing would the clip need?

You do not need to score everything out of ten. A short note beside each item is enough.

Run the test more than once

One result should not decide the whole comparison. Generated outputs can vary between runs.

If time and budget allow, create more than one clip with the unchanged brief. Label each file with the model, settings, and date. Model access and controls may change, so these notes help explain what was compared.

Do not call the result a controlled benchmark unless the test was designed and documented that way. “Better fit for this brief” is more accurate than “the better model.”

Use only reference material you own or have permission to use. Do not upload private or unreleased client material unless the owner has approved its use and the platform terms allow it. Do not include recognizable people without their consent.

Check generated product footage before publishing. It should not be presented as proof that a real product performs in the way shown.

A useful comparison does not need a complicated spreadsheet. Keep the brief steady, change one variable, and record what happened.