Generating product images
The gap between a catalogue photographed on a table and one photographed in a studio is worth real money on a Shopping surface, where your image competes directly with five others in the same row. Generation closes that gap without the studio.
Two different jobs
They get discussed together and behave nothing alike. A packshot is a corrective job: the same product, on clean white, with the lighting fixed, the background removed and the detail sharpened. The output has to be the product, exactly, because it is the main image and Google requires that it represent what you sell.
A lifestyle shot is a generative job: the product placed in a scene it was never photographed in. It is what makes a listing feel like a brand rather than a warehouse, and it belongs in additional images, on the product page and in social ads.
What the main image has to survive
- It must show the product being sold, not a similar one, and not a version with details the buyer will not receive.
- No promotional text, badges, borders or watermarks burned into it.
- Consistent with the landing page. An image that disagrees with the product page is a mismatch disapproval waiting to happen.
- Large enough to render well. The floors are 100 by 100 pixels, and 250 by 250 for apparel; the surfaces deserve considerably more.
Corrective editing that does not substantially alter the product, such as lighting, background removal and sharpening, is treated differently from generating a new scene, both by platform policy and by the EU AI Act. The compliance chapter covers where that line sits.
Where it holds up and where it does not
Generation is reliable on the things that are properties of the scene: light, shadow, surface, backdrop, context, model and setting. It is least reliable on the things that are properties of the product: text on packaging, logos, fine texture, stitching, the exact shape of a component. Those are precisely the details a buyer inspects before spending.
The practical rule is to let generation own the scene and let photography own the product. A single decent source photograph of the item, relit and placed, beats a fully synthesised product, and it is also the version that survives a returns dispute.
Doing it at catalogue scale
- Decide the look once, covering lighting, background, framing and crop, and apply it as a template. Consistency across a row of your products is what reads as a brand.
- Batch by category. Shoes, apparel and small hard goods want different framing, and a single template across all three looks like a mistake in each.
- Review before publishing, in a grid rather than one at a time. Errors are obvious next to their neighbours and invisible on their own.
- Keep the source photograph. It is your evidence of what the product is, and your input if you re-render later.