To create product-image ad variants with Gemini, submit an approved product photograph with a narrow editing instruction, such as changing the setting while preserving the product. Keep the reference, model and prompt version together. Compare every output with the original before adding the offer and CTA in your design tool.
This guide prepares reference-based edits through the Gemini API using the currently documented Nano Banana model. Download the variant review sheet and reference-edit prompts. The examples are proposed inputs and review criteria. We did not run paid image generation for this article.
Choose a reference with verifiable details
Start from an approved product photograph. Preserve a full-resolution original and note the product or SKU it represents. Before prompting, list the features that must remain unchanged: silhouette, cap, material, color, logo and label text.
For our fictional bottle, the proposed creative difference is the setting: a neutral desk or a kitchen counter. The product remains the same. Do not compare two outputs with different bottle colors and describe the experiment as a background test.
Use an image you have the right to upload and edit. The product-image quality checklist helps turn visual fidelity into a practical review, rather than leaving approval at “looks good.”
Match the model and API example you are using
Google's image-generation guide documents text-and-image editing. At the review date, its example uses gemini-nano-banana-2.1 with the Interactions API. Do not combine an old model identifier, a different endpoint and copied request fields from another tutorial.
Check access for your project and follow the API-key setup guide. Keep the key in your server or local environment, outside the prompt, browser bundle and shared worksheet. Confirm the applicable billing before making generation requests.
Send the approved reference image with the editing instruction using the current provider example. Record the exact model identifier, image filename and prompt version. The model reference is the place to verify current capabilities before choosing output settings.
Ask for a narrow edit with explicit invariants
Our desk variant changes only the surrounding environment. The prompt below is an original brief you can adapt after replacing the fictional product details with your approved reference:
Use the supplied bottle photograph as the product reference.
Place the same bottle on a neutral work desk in soft daylight.
Preserve its silhouette, cap, color, material, logo and label.
Keep the complete product visible with open space to its right.
Do not add text, price badges, people or new product features.
The setting may change. The product identity must remain fixed.For the second version, change the setting sentence to a neutral kitchen counter. Keep the rest of the brief unchanged. Save both instructions, even if the tool offers conversational editing, so the creative team can understand the intended comparison.
Model instructions are not a lock on pixels. If strict packaging fidelity is essential, review whether ordinary compositing of the original product cutout is a better production method for that asset. Generation should earn its place by producing a usable result.
Compare each output with the original reference
Place the output and source beside each other at the same viewing size. Check the silhouette first, then cap geometry, printed details, reflections and contact with the surface. Zoom in on the label instead of approving it from a thumbnail.
| Check | Example reason to reject |
|---|---|
| Identity | The cap or bottle proportions changed |
| Printed details | The logo or label contains invented characters |
| Physical scene | The bottle floats or intersects the desk |
| Advertising message | Added objects imply an unsupported feature |
| Composition | The required crop removes the product or text space |
Record an observable reason for each rejection in the sheet. A new generation should address a named problem, not simply continue until the team stops looking closely. Keep source, prompt and output versions linked using the prompt-versioning guide.
Turn approved images into a controlled creative batch
Add the offer and CTA in the design tool after image approval. That keeps factual copy editable and avoids asking the image model to render the final campaign text. Export the placement-specific crops and inspect each at phone size.
For the fictional desk-versus-kitchen comparison, keep the ad text, offer and destination consistent. Record which approved file belongs to each creative ID. If a crop forces the product to move or changes the message hierarchy, document that additional difference before interpreting results.
An approved still can also become the first frame of a Veo product shot. Repeat the fidelity check after animation; approval of an image does not automatically approve every frame generated from it.
Sources checked October 15, 2026. This is a source-reviewed editing workflow with original prompt examples, not a benchmark or a claim that the illustrated variants were generated successfully.