Ad variants from one product photo for six audiences

To make ad variants from one product photo, keep the photo as the first reference and change one thing per audience: the setting, the light and the props around the product, never the product itself. Six audience edits on Ideogram 4.5 cost $0.45 at medium quality on Sume ($0.075 each), or $0.225 at low. Morphic lists colorways, relighting and interior swaps among the model's uses, which is the same set of changes.

One brief, six edits

That is six scenes; drop the one that does not fit your product. Send the same photo as the single input_references entry each time and leave aspect_ratio out so every variant keeps the photo's shape.

• Fixed: "Keep the product, its label and its proportions unchanged."
• Commuters: "Place it on a train seat table, morning light."
• Students: "Place it on a dorm desk with a laptop and notebooks."
• Parents: "Place it on a kitchen counter with a fruit bowl."
• Gym: "Place it on a bench in a gym locker room."
• Gifting: "Place it with ribbon and wrapping paper on a table."
• Travel: "Place it on a hotel bed beside an open suitcase."

What it costs against alternatives

Ideogram 4.5 is priced per image on Sume at the provider list times 1.25. The figure for any other model is on its endpoint record: GET /v1/images/models/{id}/endpoints returns a pricing array. Read it before comparing.

Keep the product honest

Variants are a tempting place for the product to drift: a label that gains a word, a cap that changes shape, a bottle that gets taller. Check the product crop of each variant against the original at 100%. Ad rules differ by market, and a variant that alters the product's look can mislead a buyer; reject any that do.

Decide the audiences before you generate, from data you hold: who bought, who clicked, who asked a question. Six audiences invented in a meeting produce six pictures that nobody tests. Name each variant file after its audience and its prompt, so a result in the ad dashboard can be traced back to the exact wording that made the image.

Run a low-quality pass first. At $0.225 for all six, you can see which scenes the model handles and which come out muddled, then spend the medium or high tier only on the ones you will run. A scene with several props is more likely to confuse the product's scale than a plain one.

Say what you want from the variable part in terms the model can see. "Cosy" gives you anything; "warm lamp light, wooden table, a mug" gives you something you can compare across six outputs. Keep the same light direction as the source photo unless relighting is the point; the light on the product and the light on the scene should agree.

If you want several outputs from one request, some models accept n; the per-model limits are listed in n variants in one call. For resizing each variant to each ad slot, see edit once and crop for every placement.