AI-Generated Product Images: Generate or Edit?

AI-generated product images for ecommerce are photos made or changed by a generative model instead of a camera. The term covers a background swap on a supplier photo and a fully made product shot.

Editing and full generation carry different risks, so choose the operation before choosing a tool.

Supplier photos were a weak point in both stores I ran. One sold print-on-demand apparel, while the other sold niche products.

Replacing the product in the picture seemed easier than fixing the source photo. That shortcut risked promising details the parcel couldn't deliver. The better goal is a polished set that still traces back to the item you ship.

Key takeaways

Should you edit or generate product images?

Editing changes a real product photo, while generating creates the depicted product from a description. Shopify Magic shows the practical split by letting sellers remove or replace a background around an uploaded subject. A text-to-image request asks the model to invent the subject too.

An editor can keep the item's shape, material, stitching, and hardware while changing its surroundings. A generator makes pixels that fit the prompt. Its clasp or fabric can look real while differing from your stock. That source check also applies when AI drafts other store work.

I treat a generated product as a concept image and an edited product as a listing candidate. For the candidate, compare the source and output at full size with a product fidelity test. Check the item's outline, color, material, hardware, logos, and countable parts. A generated scene can look better, yet appearance can't prove product accuracy.

In AI dropshipping, the model can prepare the image asset. You still trace each product claim to its source.

The four image slots, and which ones you can generate

Each product image has a different job, so decide whether to edit or generate it and test the result. The slot is the job an image performs on the product page. Apply the same checks when comparing AI tools for dropshipping,

1. Primary image: Establish the exact item for the sales channel.
2. Alternate angle: Reveal a view the first photo can't show.
3. Scale and detail: Prove size, material, and construction.
4. Lifestyle and context: Show the real item in an invented setting.

The decision matrix keeps those jobs separate:

‍

The four decisions below explain what earns each verdict.

1. Primary image

Edit a real photograph for the primary image because it has to establish the exact item you sell. Google Merchant Center requires the image to display the actual product. It must show the whole item with minimal staging, without placeholders, substitutes, or promotional overlays.

Keep the product pixels intact while you remove the background, correct lighting, or crop the frame. Compare the edit with the full-size source. Confirm that the color, material, shape, and variant still match.

Merchant Center can disapprove a product with an obstructed image until you replace it. A polished invention can therefore cost you the listing.

A weak supplier image limits what editing can recover. If the product is hidden, tiny, or blurry, get a better supplier asset or photograph a sample. I stop there because this slot gives the channel its product evidence.

2. Alternate angle

Use a supplier-provided angle or photograph a sample because a model can't reveal a view it has never seen. A generated back, underside, or interior can look consistent with the front while inventing ports, seams, fasteners, or openings.

This slot matters for products with hidden ports, closures, interiors, or backs. A front view can be enough for a flat print or another item with no hidden feature. Compare every added angle with the supplier's assets or the sample in your hand.

You usually discover a false angle after the customer receives the item. Cloudinary surveyed 2,693 consumers across Australia, Germany, the UK, and the US. 30% reported a return because the product looked different on the website. The survey isn't dropshipping-specific, but it shows why a convincing false angle is worse than a missing one.

3. Scale and detail

Show scale and material with a real reference because a generated hand, model, or room can change the product's proportions. The reference can be a photographed ruler, a known object, or a measured detail shot. Put written dimensions in page content or a non-primary image.

On my apparel store, a clean photo still couldn't settle fit. I treated the image as visual proof and kept the measured sizes in page content. Our size conversion guide handles the written size mapping.

Google Merchant Center treats watermarks, logos, and promotional elements as primary-image obstructions. Put dimension overlays in page content or supporting images, and keep the feed's main image clear.

Because perspective can distort a real reference, compare its apparent size with the listed dimensions before publishing.

4. Lifestyle and context

Generate the setting when the product is a real cut-out from your source photo. This slot shows where the item fits or how someone might use it. Google Merchant Center treats these scenes as additional images rather than the main product image.

This is where I would use generation. Composite the cut-out product into the scene. Then inspect every edge where the model may have redrawn part of it. Match the palette and setting to the choices in our guide to branding a product.

A generated person presented as a customer creates a new truth claim. A model wearing clothing must also preserve the garment's real cut, color, and print.

Run the Generate-or-Edit Line on one slot

The Generate-or-Edit Line uses three checks to decide whether to edit or generate one image. First name the claim, then find its source photo, and finally check the destination,

1. Name: Write down what the image has to prove.
2. Find: Locate the real photo that supports that claim.
3. Match: Choose the operation, then verify the destination accepts it.

Run the checks in order before opening an image tool.

Step one: Name what the slot has to prove

State the buyer or channel question the image must answer in one sentence. A primary image proves identity. An alternate angle proves construction, while a scale shot proves dimensions and a lifestyle scene proves context.

Write that sentence beside the image. Split claims when one image would become crowded or make either fact hard to verify. A clear real photo may prove identity and color together.

Step two: Find the real photo behind it

Match the claim to a real supplier photo or a photograph of your sample. Record which image proves the product's shape, color, material, and visible features. A lifestyle background needs no real reference when the product placed inside it has one.

The sequence stops when the source set can't support the claim. Order a sample or reconsider the product. Apply the same evidence standard when choosing a supplier.

Step three: Match the operation to the channel

Choose the operation from the claim, then verify that the destination accepts the finished asset. Edit when the slot proves a product fact. Generate when it supplies context around a verified product.

Open the destination's current image-policy page and preview the exact feed asset. When the policy is silent, use the stricter primary-image rule for product facts. The check passes when these conditions hold,

1. The product still matches the source.
2. The destination accepts the image.
3. You can name what the image proves.

What you have to disclose once an image is generated

Disclose an AI image when the market's law or the destination's policy covers its claim. Run the check in this order,

1. Name the buyer's market and country where customers will see the image.
2. Ask whether the image falsely looks authentic to those buyers.
3. Check the destination's AI-label rule before uploading the final file.

Provenance metadata such as the C2PA and IPTC provenance standards travels inside the file. Meta can read those technical signals when labeling images. Missing metadata doesn't prove an image is human-made.

Google Shopping free listings and Shopping ads need images that represent the item offered. Google's policy explains the image rule.

Etsy is a marketplace requiring photos of the actual item, so use a generated setting only when the product itself remains true to the listing. Shopify's media guidance is useful for an edit, but it doesn't establish that a generated product claim is accurate.

Article 50 requires EU AI-system deployers to disclose image content that constitutes a deep fake. The content must resemble an existing person, object, place, entity, or event and falsely appear authentic.

Article 50 explains the same transparency duty. It has applied since 2 August 2026.

The Commission's guidance says metadata alone isn't enough.

Did the new images actually sell anything?

Judge a new image set by product-page conversion and return rate, using a bounded comparison on products you already sell. Image production speed measures the workflow. Conversion rate and returns measure what shoppers did after seeing the result.

Set up the comparison this way,

1. Change the images for one established product family.
2. Keep a comparable family on its current supplier photos.
3. Hold price, traffic source, offer, and ad spend steady.
4. Wait for both groups to collect orders and returns.

Use this comparison as directional store evidence. Statistical proof needs a designed experiment with enough orders for a reliable result.

Different traffic or a discount makes the comparison useless. Replacing the whole set also hides which slot caused the change. If conversion rises while returns worsen, inspect alternate-angle and scale claims before keeping the new images.

I recommend publishing the new set only when conversion holds or improves without a worse return rate. A faster generator earns no catalog rollout by itself. A failed comparison should send you back to the source photo instead of another prompt.

A product library can supply the comparable listings that make a generated-image test measurable.

FAQ

Will Google disapprove AI-generated product images?

Google checks whether the asset shows the actual product without promotional text or watermarks. An AI-made image can pass that test, while an image that obscures or misrepresents the item can trigger disapproval.

How many products should you re-shoot at once?

Re-shoot one established product family first. Expand only after the unchanged comparison family shows that conversion and return-rate changes weren't store-wide.

Do you need a sample to make good product images?

You need a sample when supplier assets can't prove the angle, scale, material, or detail you want to show. Keep the existing image or leave that slot empty until a real photograph exists.

Can you put an AI model in clothing photos?

You can generate a model around a real garment image, but the output must preserve its fit, color, and construction. Save the model release or tool license, check the destination's AI-label policy, and apply any required disclosure.