You sell something real: candles, jam, ceramics, a plate that leaves your kitchen every lunch. The phone photo of it looks like a phone photo, on a counter, under a ceiling light, and a photographer isn't in the budget this month.
Wild Fig Jam
Label: WILD FIG JAM
Lid: dark green
On the counter, ceiling light
What the buyer will unwrap
Wild Fig Jam on a breakfast table
Label: WILD FIG JAMM
Lid: dark green
Morning light, marble, a spoon
A promise the parcel breaks
An image model can put that product on a marble shelf or a breakfast table in a minute. The hard part is making it look better without making it look different. A buyer orders what the picture shows, so the picture has to show what's in the box.
Shoot the product in good light
Stand near a window in daylight and turn the flash off. Put the product on a plain surface with the label facing you and fill the frame with it. Take three shots: straight on, from a three-quarter angle, and a close-up of the label.
Photoroom says a phone, or the product shots you already have, are enough to start from. The catch is that every later step is held against this picture. Keep the originals in their own folder and never edit them: they're what you check against at the end.
Checkpoint
Three sharp phone shots of each product, with the label readable.
Cut out the background in Photoroom
Upload the straight-on shot to Photoroom's background remover. It finds the main subject and removes everything behind it. Keep the result transparent, put it on white, or pick a colour, then export a PNG. Photoroom's free start covers the cut-out; high-resolution exports need a paid upgrade, and this is the image buyers will zoom into.
This cut-out is your clean listing shot, and it's still a photo of the actual product. Make one for every product before you generate anything, so a range of jars or mugs all sit on the same background.
Checkpoint
A clean cut-out of each product on white or transparent.
Stage it in a scene
the step that pays off
Open Gemini's image tool, upload the cut-out and paste the prompt below. On the free plan Gemini runs this on Nano Banana 2; paid Google AI plans run Nano Banana 2.1. In ChatGPT, upload the cut-out and paste the same prompt. Use one or the other, and run it two or three times so you have scenes to choose from.
Fill in the label text exactly as it's printed, and the real size in centimetres. The scene is the easy part. The product is the part that has to survive, which is why the prompt spends its rules on the product and one line on the scene.
Put the product in a scene without changing it
Pro tip
Describe the scene the way a buyer would use the product: a candle on a bedside table at night, a jar of jam beside a breakfast plate.
Checkpoint
Two or three scene images of the product to check.
Check every image against the real product
Open a fresh chat in ChatGPT or Gemini, attach the original phone photo and the scene, and run the check below. Then look yourself: open the scene at full size, zoom in on the label, and hold the real product next to the screen.
The usual failure is small. A letter in the label changes, a logo softens, a lid turns a shade darker, a jar looks a size bigger next to a tiny spoon. A photo like that is a promise the parcel breaks, and the buyer finds out at the door.
To fix one detail in ChatGPT, select that area and describe the change. OpenAI notes that selections are not always precise and edits can spread past the area you picked, so run the check again after every fix.
Check the new image against the real product
Checkpoint
A pass on every line, and your own look at the label at full size.
Put the real shot first
Make the cut-out from step two the main image on every listing. The scenes go on your own site, your social posts and the menu board.
On Etsy, a generated scene stays off the listing. Etsy's rules ask for original photos of the actual product buyers will receive, not renderings or stock photos. The exceptions are narrow: computer-made mockups in extra images to show a personalised item's options, and stock mockups for a design printed onto a base item by a production partner. Read your other marketplaces' image rules before a scene goes up there.
Leave the watermark on, and never call a scene a photo. Gemini says its images carry an invisible SynthID watermark and a visible one. OpenAI says ChatGPT images carry Content Credentials and a SynthID watermark.
Checkpoint
Each listing led by the real photo, and checked scenes kept for your own channels.
The takeaway
You end up with a clean cut-out of every product, a few scenes that pass a line-by-line check, and the real photo leading every listing.
The limit: no prompt guarantees the label, the colour or the size survives a generated scene, and a model's check misses things too. The check is what stands between the scene and the buyer, and it's yours. If a scene can't pass it, the cut-out on white is a perfectly good product photo on its own.
Variations
Use Photoroom's own AI backgrounds. Photoroom can generate a scene around the cut-out from a text prompt or a reference image. It saves a trip to a second tool. Run the same check on what comes back.
One scene for a whole range. Keep the scene description fixed and run it for every flavour or colour, so the range looks like one shop. Check each image on its own: the colour variant is the detail a buyer orders by.
Menu photos. Shoot the dish exactly as it goes out, cut it out, and put it on the menu board's colour. Don't let a scene add a garnish the kitchen doesn't serve.
Don't generate the product from a description. With no photo to hold it to, the model draws a product you don't sell, and there's nothing to check it against.