Edit your real product photo instead of generating a new one, and do it in order: clean the background first, then add the scene, then check the details the AI tends to wreck. That sequence is the whole workflow. This post walks through it with ChatGPT and Gemini image tools for a small store, with prompts and a checklist for the things AI gets wrong.
Status of the prompts: untested templates. I didn't run image generations for this post, so there are no result images here, and nothing below claims a specific output quality. Treat the prompts as starting points and run them on one product before a whole catalogue.
If you want a big list of ad-style image prompts instead, there is a library of product-ad image prompts. This post is about the process of getting from a phone photo to an image you can sell with.
What should you shoot before you open an AI tool?
AI editing amplifies what you give it. A blurry, tilted, badly lit photo gives you a polished blurry, tilted photo. Spend ten minutes on the input.
- Shoot near a window, product on a plain surface, no direct sun.
- Take it straight on and also from a 45 degree angle. Fill the frame, leave a margin.
- Wipe the product. Dust and fingerprints get "enhanced" into textures.
- Shoot every colourway the same way, same distance and angle. This pays off in the consistency step.
- Keep the original files. Every edit below should be a copy.
Step 1: clean the background first
Do the boring job before the creative one. A clean cutout on a neutral background is a stable base: if later steps go wrong, you restart from here, not from your phone photo.
OpenAI's help page for images in ChatGPT says you can upload a photo and describe changes, including making the background transparent, and that you can select an area and describe the change or just describe it in the chat. It also says the highlight is not always precise and edits may spill past the selected area (help.openai.com/en/articles/11084440, checked 2026-10-09 via search excerpt, as the page itself blocked automated fetches). That last note matters: it's the reason to keep prompts narrow.
Edit this photo. Keep the product exactly as it is: same shape, proportions, colour, label text and logo. Change only the background to plain white with a soft natural shadow under the product. Do not add props, text or reflections. Do not retouch the product itself.
Variations: ask for a light grey background if the product is white, or for a transparent background if you plan to composite later.
Check the result at 100% zoom, not as a thumbnail. Look at edges (hair-like fuzz on fabrics, glass and chain links cut away), at the shadow (should touch the product) and at the text.
If the cutout is bad, a dedicated background remover in your editor is often more reliable for the product edge, and you can still use the AI for the scene. That mixed approach is common and nothing to be embarrassed about.
Step 2: add a scene without changing the product
Once you have a clean base, ask for a scene. Describe the setting, the light and the camera, and repeat the "don't touch the product" rule every time.
Using this product photo, place the product on a [SURFACE, e.g. light oak table] in a [SETTING, e.g. bright kitchen corner]. Soft window light from the left, shallow depth of field, eye-level camera. Keep the product's shape, size, colour, label and logo unchanged and in sharp focus. No people, no extra text, no brand logos in the background. Square 1:1 image.
Tips that save retries:
- One scene change per prompt. Surface, props and lighting together invites drift. Do the surface, then add one prop in a follow-up.
- Name scale. "A palm-sized jar next to a coffee cup" stops giant-jar results.
- Say where the light comes from. It removes the plastic look more than adjectives like "photorealistic".
- Use selection and follow-ups. If the tool lets you select a region, say "change only the area behind the product".
- Aspect ratio. ChatGPT lets you pick the ratio in the picker or in your prompt (per the same help page). Ask for the ratio your ad or theme needs: 1:1 for a catalogue, 4:5 for feeds, 9:16 for stories.
Google's Gemini help page says you can upload an image and ask Gemini to edit it with a text prompt, and combine several images into a new one (page checked 2026-10-09). It also says image editing is limited by the languages and countries your Gemini app supports, requires you to be 18 or older on a personal account, and that lighter models and daily image quotas can restrict multi-image references or re-edits. The same prompts work as a starting point; check your own quota in the app. For wider Gemini habits, see the Gemini Gems guide.
Step 3: keep a whole product line consistent
A store with five colourways, each in a different "AI look", reads as untrustworthy. Consistency is a prompt-writing job:
- Write one scene block and reuse it word for word. Only swap the product photo.
- Use the same aspect ratio and the same base distance in your original shots.
- Keep a style note in a doc: "oak table, left window light, 1:1, 3 o'clock shadow".
- Generate all variants in one session when you can, because earlier turns in the same chat steer later ones.
- Review a contact sheet: put all the images in a grid and look for outliers in colour temperature, shadow direction and scale.
Scene block (reuse exactly):
"on a light oak table, bright kitchen corner, soft window light from the left, shallow depth of field, eye-level camera, 1:1, no people, no extra text"
Task: place the attached [COLOUR] [PRODUCT] in that scene. Keep the product unchanged.
Colour is the one that hurts most. If your cutout is a deep green and the scene render turns it teal, customers receive a different product and you get a return. Compare against the original on a calibrated or at least consistent screen, and lock the product pixels when it matters (see the compositing option below).
What AI gets wrong with product images
Run this checklist on every image before it goes on a product page.
| Problem | How to spot it | Fix |
|---|---|---|
| Logos and label text redrawn | Zoom to 200% and compare to the original, letter by letter | Composite the original product cutout over the AI background |
| Small text and numbers garbled | Ingredient lists, size charts, serial numbers | Never rely on AI-rendered text; add text in a design tool |
| Proportions shift | Product looks taller, thinner or larger than real | Overlay with the original at 50% opacity to compare |
| Added or removed parts | Extra buckle, missing zip pull, different number of stitches | Count features against the product |
| Colour drift | Swatch-compare against the original | Correct in an editor; keep the original as truth |
| Impossible physics | Floating products, wrong shadows, reflections that do not match | Ask for contact shadows; regenerate |
| Hands and people | Extra fingers, odd grips | Avoid hands, or use a real photo of hands |
| Material changes | Matte becomes glossy, leather becomes plastic | Name the material and say "do not alter finish" |
| Misleading scale | Tiny item looks big | Add a known-size object or a real size photo |
The deeper rule is the one in the FAQ above: the image has to show what the customer will actually receive. A prettier lie is still a return, a bad review or a platform complaint.
The compositing fallback for exact products
When the logo or label is critical (cosmetics, packaged foods, anything with regulatory text), don't ask the model to preserve it. Take the model's scene, then place your original, untouched cutout on top in an ordinary editor with a layer for the shadow. It's an extra ten minutes and it removes the biggest risk. An AI can create the background, lighting direction reference and props, while your pixels stay the product.
Marketplace image rules: what I could and couldn't verify
Marketplace rules are strict and vary by platform. I tried to read Amazon India's seller image guidelines and could not: the primary pages need a seller login or returned errors. A search summary of third-party guides describes a common pattern for main images (plain white background, product only, no logos or watermarks, enough pixels for zoom), but these are secondary sources and I'm not treating them as verified. Check these in your own seller account before uploading:
- Main image background and framing rules.
- Minimum size and accepted formats.
- Whether text, badges or watermarks are allowed on each image slot.
- Any policy on edited or generated images.
For Shopify, the help page on uploading files lists these limits for images: up to 20 MB, up to 25 megapixels, and JPEG, PNG, WebP, HEIC or GIF (page checked 2026-10-09). It does not give a recommended display size, so preview the image in your actual theme on mobile.
A repeatable routine for a small catalogue
- Shoot every product the same way, keep originals.
- Run the background-clean prompt. Check at 100%.
- Pick one scene block and apply it to all products in that line.
- Run the failure checklist. Composite where logos matter.
- Export at the size your theme or marketplace asks for, with alt text that describes the real product.
- Put a plain, well-lit original photo among the images so customers see the real thing as well.
If you're building the rest of the listing, the companion pieces are prompts for Shopify product descriptions and AI ad creative prompts for Meta and Google. For image-model basics, how Claude's vision analyses images is useful for auditing your own photos, for example asking a model to list every visible feature so you can compare the list to the real product.
I haven't measured whether AI-edited images raise conversion for any store, and I won't suggest they do. Test it the boring way: run old and new images on a few products for a few weeks and compare, on your own numbers.
