Product Images Just Got Cheap. Keeping Them Right Didn’t

On 6 October, Google released Nano Banana 2.1, the latest version of the image generation and editing model behind the Gemini app. It’s rolling out across the Gemini app, AI Mode in Google Search, Google Ads and Google’s creative and developer tools, including Flow and AI Studio. Google says the update improves visual quality, text rendering and consistency across repeated edits, and it can now combine up to 14 reference images while keeping up to ten objects consistent in the result.

The cost and time of making a product image have largely stopped being the constraint. Until recently, a photoshoot set the limits on what a store could show: how many lifestyle shots, how many angles, whether every colourway got its own image or just the hero product. A merchant can now produce lifestyle and context images for an entire range in an afternoon. That’s genuinely useful, particularly for smaller teams. But it doesn’t remove the work. It moves it somewhere less visible.

An image model doesn’t know your product. It knows what you’ve shown it and what it has learned products like yours usually look like. Even good models can shift a colour slightly, smooth out a texture, change a proportion or add a detail that isn’t there. On a product page, that matters more than it sounds. Until the parcel arrives, the image is the product as far as the shopper is concerned.

The practical answer isn’t to avoid generated imagery. It’s to be deliberate about where it’s used. Many retailers will find a sensible split: real photography for primary and colour-critical images, and generated images for lifestyle scenes, seasonal campaigns and context shots, with each one checked against the physical product before it goes live. That approval step has to exist somewhere in the workflow, and it needs to be someone’s job rather than an assumption.

Resizing tools, compression, content delivery networks and catalogue exports often strip embedded metadata by default, because for years nobody needed it. An image that was compliant when it was generated can arrive in a shopping feed without the tag that makes it compliant, and nobody along the way made a decision to remove it.

When images were expensive, there weren’t many of them. When they’re cheap, the number of assets grows faster than anyone can manage by hand. Each product might now have a dozen images across several variants, and each of those needs to appear correctly on the website, on marketplaces, in Google Shopping feeds and anywhere else the product is sold. A lifestyle image of the navy jacket showing up on the charcoal variant’s page is a small error that erodes trust quickly, and it’s far more likely when images are uploaded channel by channel rather than managed in one place.

This is where image generation stops being a creative question and becomes a systems one. The businesses that will get the most out of cheaper imagery are the ones with a single source of truth for product data — whether that’s a product information management system, a well-structured storefront catalogue or an ERP-connected platform — that holds approved images alongside the SKU, variant and attribute data, and pushes them consistently to every channel. Without that, faster image production just means faster accumulation of inconsistencies.

B2B stores often have the opposite problem to fashion retailers: thousands of SKUs and very little photography. Generated context images can help fill those gaps, and better text rendering makes things like specification graphics more practical. Anything that shows a specification should be built from the product data, not trusted to the model.

Google’s new model can ground its images using Google Search and Google Image Search, and AI shopping assistants increasingly compare products using images as well as text. That makes consistency across channels more important, not less. When your website, your marketplace listings and your shopping feed all show the same accurate images against the same accurate product data, every system reading them — human or otherwise — gets the same story.