Nano Banana 2.1 Product Photo Migration Test

Google now recommends Nano Banana 2.1 for new image-generation projects. For product sellers, that is a reason to test the model, not a reason to switch every catalogue workflow overnight.

This is a migration protocol rather than a published head-to-head benchmark. Its purpose is to help a seller compare Nano Banana 2.1 with the version already used in production, using the products and failure cases that actually matter to the business.

What changed, and what did not

Google describes Nano Banana 2.1, whose model ID is gemini-nano-banana-2.1, as an October 2026 update to Nano Banana 2. The documented improvements include visual quality, prompt adherence, multi-turn character consistency and text rendering. The model supports 1K, 2K and 4K output, with 1K as the default. Google also says that tiling artefacts have been addressed for very wide and very tall aspect ratios at 2K and 4K.

Those changes could matter to sellers producing packaging images, listing graphics, banners and repeated lifestyle edits. However, the reference capacity has not expanded. The documentation still allows up to 14 reference images, including up to 10 objects for high-fidelity inclusion and up to four people or characters for consistency. A version upgrade does not remove the need to control how many products appear in a scene.

The sensible question is therefore not whether 2.1 is newer. It is whether it creates more approved images per hour without introducing new product errors.

Freeze the comparison before generating anything

Choose a small but difficult set of real catalogue products. Include at least one reflective item, one product with a label or engraving, one colour variant and one multi-product composition. Jewellery sellers should include a chain, clasp, stone setting or fine engraving because these details expose geometry drift quickly.

Run the old and new models with the same source images, prompt, aspect ratio, output resolution, thinking level and number of outputs. Do not improve the prompt for one model during the baseline. If the interface does not expose a repeatable seed, create several outputs for each case rather than judging a single lucky result.

Save the model ID, date and settings beside every output. Otherwise a visually attractive contact sheet can become impossible to reproduce a month later.

Case 1: background replacement without product drift

Start with the simplest commercial edit. Place the same product into a clean lifestyle background while instructing the model to preserve its shape, proportions, colour, surface finish, logo and small construction details.

Check the product before the scene. Compare the silhouette, openings, clasp, prongs, seams and label placement with the source. Then inspect contact shadows and reflections. A beautiful scene fails if the pendant bail changes shape or a bottle gains a different cap.

Case 2: packaging, logos and engravings

Use a product whose visible text customers could verify. Ask for a modest scene change, not a new label design. Compare every character, line break and relative position with the original.

Google specifically reports improved text rendering in 2.1, so this is a high-value regression case. Still, readable text is not automatically accurate text. Count misspellings, substituted characters, warped logos and invented claims as product errors. For regulated or safety-related packaging, keep the verified original artwork as a separate compositing layer rather than trusting generated pixels.

Case 3: a controlled colour or material variant

Request one clearly defined change, such as yellow gold to rose gold, black to navy fabric, or a matte finish to a polished finish. Tell the model that all geometry, stones, fasteners, stitching, labels and camera position must remain unchanged.

Compare the outline and internal landmarks with the source. On jewellery, inspect the number and placement of stones, prong shape, chain links and engraving. On packaged goods, check cap geometry and label boundaries. Reject an output if the model has redesigned the product while changing its colour.

Case 4: a multi-product bundle

Build a scene from three known SKUs before attempting a crowded catalogue composition. Provide one clean reference for each product and name them consistently in the prompt. Ask for each item to appear exactly once.

Score presence, duplication, identity swaps and relative scale separately. The reference-image ceiling tells you what the model accepts, not what it will reproduce perfectly. Three accurate products are more commercially useful than ten plausible approximations.

Case 5: a three-step edit chain

Multi-turn editing is valuable when a seller wants to refine a scene without rebuilding it. It is also where small changes can accumulate. Start with a background replacement, then adjust lighting, then alter one prop while asking the model to leave the product untouched.

Compare the product after every turn with the original reference, not only with the previous generated image. Note when drift first appears. If the third edit damages the clasp or label, test whether restarting from the verified source and issuing a combined instruction produces a cleaner result.

Score factual accuracy before aesthetics

Give every output a product-accuracy pass or fail before rating composition. Record the reason for each failure in a short controlled list, such as geometry, text, colour, missing detail, duplicate item, wrong scale or unwanted redesign.

Then record clean-output rate, median generation time and actual cost per approved image. Cost per generation can be misleading if one workflow needs many retries. Use the current Google pricing page when calculating the test because prices and billing details can change.

If wide banners matter to your shop, add a 4:1 or 8:1 case because 2.1 has a documented tiling fix for extreme ratios. If you export at 4K, inspect whether fine product detail is genuinely better rather than merely larger.

A practical migration rule

Move a workflow to Nano Banana 2.1 only when it matches or improves the current model's product-accuracy rate on the hardest relevant SKU classes. A small quality gain may still be worthwhile if latency and cost remain acceptable, but do not average away a serious failure on labels, engravings or product geometry.

Keep the previous workflow available during the first production batch, version your prompts and audit a sample of approved images at full resolution. Official release notes describe broad model improvements. They cannot guarantee accuracy for a particular ring, bottle or garment.

The useful outcome of this test is not a universal winner. It is a clear decision about which jobs can migrate now, which need adjusted prompts or compositing, and which should stay on the existing process.

Sources

Google AI for Developers, Nano Banana 2.1 model page, reviewed 11 October 2026: https://ai.google.dev/gemini-api/docs/models/gemini-nano-banana-2.1

Google AI for Developers, image generation guide, reviewed 11 October 2026: https://ai.google.dev/gemini-api/docs/image-generation

Google AI for Developers, Gemini API pricing, reviewed 11 October 2026: https://ai.google.dev/gemini-api/docs/pricing