How to Combine a Product Image and a Background with AI

Actual corrected test outputs

These are the two retained corrected-run outputs, not mockups. The backgrounds were drawn test fixtures.

Original product reference · Warm background reference · Cool background reference

What this tutorial demonstrates

This guide describes a small two-reference experiment using one synthetic LUMA bottle image and simple backgrounds drawn for the test. It demonstrates a way to request a product composition, not a guarantee for arbitrary photographs. No people were used, and the results do not establish face preservation or person insertion. The useful lesson is to separate the subject reference from the base scene and then inspect both kinds of detail. An attractive bottle alone is not enough if the background has changed in a way that defeats the task.

1. Prepare the subject and the base scene

Use the first reference to identify the product and the second reference to define its intended setting. Choose a clear product image and a scene with enough space for it. In the studio, the first upload is the subject reference and the second upload is the base scene. The two images remain separate provider inputs; they are not joined into a contact sheet before generation. Check their order before submitting. Use JPG, PNG or WebP files within the stated five-megabyte limit for each reference.

2. Describe one scene, without unrelated conditions

Our initial prompt included an instruction about a plant in a different test scene. Outputs for backgrounds without plants then introduced plants and extra objects. That is a flaw in the test prompt, so the initial batch cannot fairly measure background preservation. The corrected instruction refers only to the current pair of images. Avoid clauses such as “for the other scene” or a list of alternative compositions. The model receives one task, so give it one explicit target rather than asking it to choose among your experiments.

A prompt structure you can adapt

Take only the product from reference 1 and place it on the surface in reference 2. Keep the product silhouette, main color and label recognizable. Preserve the base scene layout and pedestal shape, adding a natural contact shadow. Do not copy the old headline, old background or old pedestal from reference 1. Do not add extra products, words or objects. Adapt the placement description to something that actually exists in your second image. These are generation instructions, not guarantees of exact preservation.

3. Review the cost and submit once

Nano Banana 2 is the model selected for the two-reference workflow. The studio uses its existing six-credit cost per generation. Review your balance, prompt and aspect ratio before submitting, and wait for a result or a visible terminal failure state before deciding what to do next. Preparing the page or choosing files does not generate an image. A new attempt consumes credits again even if the prompt changes only slightly. Avoid repeated clicks while a request is processing; the existing task system tracks the submitted request.

4. Inspect the result against both references

Read the product label and check the cap, bottle shape, color and number of products. Then inspect the scene: pedestal shape, position, extra plants, other objects and the contact shadow. In our corrected warm-background example, the single bottle and round pedestal were acceptable for the controlled composition. In the corrected cool-background example, the output still brought in water drops from the original product scene and changed the pedestal. The second result looked polished but did not satisfy strict background preservation.

What the test costs tell us

Eight API calls were retained across the initial and corrected batches. The provider reported the same charge of $0.06951 for each 1024-square output with two references, totaling $0.55608. These figures describe this experiment, not a permanent supplier price or the amount charged to a customer. They are provider response values rather than a payment-statement reconciliation. The first six calls cost $0.41706 and the two corrected calls cost $0.13902. There were no video generations or automatic paid retries in these batches.

What this test does not prove

One product and drawn backgrounds are not representative of every commercial photograph. We did not establish reliability for people, hair, glass, complex packaging, occlusion or dense text. Even a visually acceptable result is not evidence that the original pixels stayed unchanged. Do not describe the small sample as a general success rate, especially because the first prompt contained a cross-scene instruction. Use your own acceptance criteria and source material before making a claim about a specific product or campaign.

Finish with a deliberate choice

If the result meets your actual requirements, save it and keep the original references. If an important detail is wrong, decide whether a narrower instruction is likely to help or whether a conventional editor is more appropriate. For merely showing the product beside another image, use free Stitch or Collage instead. For exact new lettering on a finished image, use Add Text. The related model guide explains the current studio limits. This tutorial does not claim that multi-image functionality is already deployed wherever an older production interface still shows only one upload.

Continue with the right tool

AI image combiner · Stitch images · Photo collage · Nano Banana 2 guide · Related composition guide