# AI product photography and image editing for online shops

> What AI image editing does reliably for a shop with hundreds of products, where it fails, and a batch workflow that keeps every picture true to the item.

- Published: 2026-05-01; last updated 2026-11-14
- Author: the WeStack AI team
- Subject: Generative AI
- Address: https://westack.ai/blog/ai-product-photography-and-image-editing

> AI editing is now dependable for the repetitive part of product photography: cutting products out, replacing backgrounds, adding consistent shadows and building lifestyle scenes around a real photograph. It is not dependable at redrawing the product itself, so the working method is to keep the product's original pixels and generate only what surrounds them. Put a person's approval between the batch job and the shop, because the picture is a claim about what the customer will receive.

## What AI editing does well today

Product photography for a shop is two different jobs. One is making a true picture of each item. The other is producing every version of that picture the shop needs: a white-background main image, three room settings, a square crop for social, a wide banner, a zoomable detail. The first job still belongs to a camera. The second is where image models now earn their place.

The second job has changed because current models edit by instruction. You hand over a photograph and say what should change. OpenAI's GPT Image models, Google's Gemini image models and Black Forest Labs' FLUX models all work this way, and all accept reference images so that a product can be carried into a new scene. Shop platforms have built the same thing into the admin: Shopify's [media editor](https://help.shopify.com/en/manual/shopify-admin/productivity-tools/shopify-magic/media-generation) removes or recolours a background, generates a new scene from a short description and extends the canvas for a different aspect ratio.

Here is how we rate the common jobs for a catalogue of a few hundred products.

| Job | How dependable | What to check |
| --- | --- | --- |
| Background removal | High for solid, opaque products | Edges: handles, gaps, fringes, thin cables |
| Plain or coloured background | High | The background value is exactly what the channel asks for |
| Shadows under the product | High with a fixed recipe | Light direction matches the original photograph |
| Lifestyle scene around the product | Good, with rejects | Scale, floating, props that imply extras |
| Extending the canvas for new formats | Good | Repeated patterns, invented objects at the edge |
| Colour variants from one photograph | Mixed | The colour against a physical sample |
| Upscaling | Mixed | Texture, logos and small text after enlargement |
| Redrawing the product from a new angle | Poor | Do not publish without a real photograph to compare |

## Composite or regenerate: the decision that matters

There are two ways to put a kettle on a kitchen worktop.

The first is to regenerate. You give the model the kettle photograph as a reference and ask for "this kettle on an oak worktop in morning light". The model draws a new image, kettle included. The result looks natural because everything in it was drawn together. It is also a drawing of a kettle that resembles yours. The spout may be a little longer, the handle rivets may be gone, and the brushed steel may have become polished.

The second is to composite. You cut the kettle out of the real photograph, generate the worktop and the light as a separate layer, add a shadow, and place the original pixels on top. The kettle in the final picture is the photographed kettle, pixel for pixel.

For anything a customer will buy from, we composite. Regeneration is fine for a campaign mood board and risky on a product page.

Two details make compositing work at volume.

**The cut-out.** Background removal is a segmentation problem, and good open models exist for it. [BiRefNet](https://github.com/ZhengPeng7/BiRefNet), for instance, is published under the MIT licence with high-resolution variants and separate matting models for soft edges. Hosted services do the same job per image. Whichever you use, inspect the mask and not only the result: a mask that clips two pixels off every edge makes the whole catalogue look slightly wrong on a dark background.

**Masks are advice, not a guarantee.** When you ask a generative model to change only a masked region, it may still alter what is outside it. OpenAI's [image generation guide](https://developers.openai.com/api/docs/guides/image-generation) says masking with its GPT Image models is entirely prompt-based and may not follow the mask's exact shape with complete precision. So do not trust the model to leave the product alone. After every generative step, paste the original product pixels back and compare.

### Lighting and shadows

A pasted product looks pasted when the light disagrees. If the studio lit the kettle from the left and the generated kitchen has a window on the right, the eye notices at once, even if the viewer cannot say why.

Fix this at the camera, not in the prompt. Set one lighting standard for master photographs, soft and even, from a fixed side and height, and write the same direction into every scene template. Use one shadow recipe across the catalogue: a soft contact shadow directly under the product plus a faint directional one. Consistency between products matters more to a category page than realism in any single image.

## Where does it fail?

The failures are predictable, and they sort products into groups.

**Fine detail.** Stitching, knurled knobs, mesh, perforations, jewellery settings, woven texture. Regeneration smooths or reinvents them. Upscaling is a quieter version of the same problem: a generative upscaler adds plausible detail, and plausible is not the same as present. Use upscaling to lift a slightly small original to the size a channel prefers (Google recommends product images of around 1500 by 1500 pixels or more), not to rescue a poor photograph.

**Text on packaging.** Brand names survive more often than they used to. Ingredient lists, volumes, warnings and batch marks do not. The same OpenAI guide lists precise text placement and clarity as a remaining weakness, and small print is where it shows. Any product whose label matters (food, cosmetics, supplements, electrical goods) must keep its label pixels from the photograph.

**Exact colour.** A model's idea of "sage green" is not your supplier's. Even in a composite, the generated scene can throw a colour cast that makes the product read differently. Recolouring one photograph to show variants works for plain dyed surfaces, such as a silicone case or a cotton T-shirt. It fails where colour changes the material: wood stains, metallic paint, patterned fabric.

**Reflective and transparent products.** A chrome tap shows the studio it was photographed in. Put it in a generated bathroom and the reflections describe a room that is no longer there. Glass is worse. A cut-out of a wine glass carries the old background through the bowl, and removing it means inventing the refraction. These products need a retoucher who works with the generated scene, or a reshoot against a suitable backdrop.

**Scale and grounding.** Models have no idea how big your product is. A 12 cm vase arrives the size of an umbrella stand, or hovers a centimetre above the shelf. Shopify's own guidance suggests adding sizing details to the prompt and describing how objects rest on a surface, which tells you how common the problem is.

**Sets and counts.** A set of six glasses becomes five or seven when regenerated. Count them.

## A batch workflow with human approval

Take a homeware shop with 600 products: ceramics, glassware, stainless steel cookware, textiles. It wants a white-background main image and two lifestyle images for each. This is the workflow we would build.

1. **Set a master standard.** One real photograph per product and angle, fixed lighting, neutral backdrop, a colour reference card in the first frame of each session. Everything downstream depends on it.
2. **Sort products into lanes.** Opaque matte goods go through the automated lane. Reflective and transparent goods go to a retouch lane. Anything with meaningful label text is marked "product pixels locked".
3. **Cut out and check the mask.** Run background removal, then flag masks with holes, ragged edges or a suspiciously small area for a person to fix.
4. **Write scene templates, not a prompt per product.** Eight or ten scenes per category, each with a fixed description, light direction, surface and camera height. The product's real dimensions go in as data.
5. **Generate candidates.** Three or four per product and scene, as layers behind the cut-out.
6. **Composite and run automatic checks.** Paste the master pixels, add the standard shadow, then test: is the product region identical to the master, is the measured product colour within tolerance, is the image the right size, is the background value correct for the channel. A vision model can also screen for obvious faults such as floating objects. We compare the options in [vision language models vs classic computer vision](https://westack.ai/blog/vision-language-models-vs-computer-vision).
7. **Human approval.** A person who knows the range sees the master and the candidate side by side and chooses approve, reject or send to retouch, with a reason. Those reasons are the data that improves the templates.
8. **Export with a record.** Keep provenance metadata and store what produced each image.
9. **Sample after publishing.** Review a random handful each week and watch "not as pictured" returns by product.

The record in step 8 can be small:

```json
{
  "sku": "KET-1702-ST",
  "master": "sha256:5be1...9a",
  "scene_template": "kitchen-oak-morning-v3",
  "method": "composite",
  "model": "provider/model@version",
  "checks": { "product_pixels_unchanged": true, "colour_delta": 1.4 },
  "approved_by": "reviewer-07",
  "approved_at": "2026-05-01T10:42:00Z"
}
```

When a customer or a marketplace questions an image a year later, this is what lets you answer in a minute.

Wiring this into a product information system and the shop's media library is ordinary [product engineering](https://westack.ai/services/product-engineering). The image model is a small part of the build.

## Keeping the picture honest

A product image is a statement about the product, and consumer law treats it that way. EU rules protect buyers against [misleading practices](https://europa.eu/youreurope/citizens/consumers/unfair-treatment/unfair-commercial-practices/index_en.htm), whether by giving false information or by leaving out something important, and most other markets have an equivalent. A generated scene that makes a product look larger or better finished than it is falls on the wrong side of that line, however it was made. This is general guidance, not legal advice.

Channels add their own rules. Google's [product image specification](https://support.google.com/merchants/answer/6324350) says the primary image must clearly display the exact item being sold. It also requires images created with generative AI to carry metadata saying so, and that metadata must not be stripped. Other marketplaces publish their own image policies, usually stricter for the main image than for secondary ones. Read them before a bulk upload, and check that your export step and your image CDN keep the metadata intact.

Our working rules:

- **May change:** background, surface, lighting of the surroundings, shadows, crop, canvas.
- **Must not change:** shape, proportions, colour, texture, finish, label, quantity, and anything that shows what is in the box.
- **Props must not imply extras.** A lamp base sold without a shade should not be pictured with one unless the listing says so plainly.
- **Variants must exist.** Show only colours you stock, each checked against a physical sample.
- **Scenes must agree on scale.** If the sofa is 180 cm wide, the room around it has to be drawn to match.

Wider questions of copyright, licences and disclosure are covered in our guide to [AI image generation for business](https://westack.ai/blog/ai-image-generation-for-business).

## What does it cost per image?

Vendors price per image or per token, and the fee varies with size and quality. That number is the least useful one for planning. Think in cost per approved image:

```
cost per approved image =
    generation fee x candidates generated per approved image
  + minutes of checking and fixing x what a minute costs you
  + a share of the retouch lane
```

Three things follow from it. First, the model fee is rarely the largest term. Two minutes of a person's time cost more than several generations. Second, the pass rate is the lever: a template that yields a usable image in two attempts is worth more than a cheaper model that needs six. Third, lanes have different economics. Matte opaque products may run almost untouched, while glassware costs nearly what it cost before.

Measure each lane separately, and compare against what a studio day produces for the same products, not against zero.

If volume grows to the point where you want the pipeline inside your own tooling, with queues, moderation and spending caps, the architecture is described in [adding image generation to your product](https://westack.ai/blog/image-generation-in-your-product).

## How to start

Pick 30 products in the first week: ten easy ones (matte and opaque, with no text), ten with labels and ten reflective or transparent. Photograph masters to one standard. Run all 30 through cut-out, two scene templates and the checks above, with one person approving.

Record three things per group: the share of candidates approved, the minutes of human time per approved image, and the reasons for rejection. Those make the first decision for you, which is where the automated lane ends and the retouch lane begins. In our experience the line usually sits between "opaque" and "shiny", and drawing it honestly is what makes the project pay.

After that, fix the master photography standard before touching prompts. Building the batch pipeline, the checks and the approval screen is part of our [generative AI work](https://westack.ai/services/generative-ai).

## Frequently asked questions

### Can AI replace a product photographer?

Not for the first photograph. You still need one accurate, well-lit picture of every product, because that picture is the evidence of what you sell. AI replaces the repeated work that follows: cutting out backgrounds, building room scenes, making formats for each channel and tidying shadows. Reflective, transparent and finely detailed products still need a retoucher or a reshoot.

### Are AI-edited product photos allowed on marketplaces?

Generally yes, as long as the image still shows the exact item and meets the channel's image rules. Google Merchant Center also requires images created with generative AI to keep metadata that marks them as AI-generated. Rules differ between marketplaces and they change, so read the current image policy for each channel before a bulk upload.

### Why does my product change colour in AI-generated scenes?

Because a generative model redraws pixels to suit the new scene, and warm or cool light in that scene shifts the product with it. The fix is to stop the model touching the product at all: cut the product out of the original photograph, generate only the surroundings, paste the original pixels back and check the colour by measurement.

### How should I think about cost per image?

Count the cost of an approved image, not a generated one. Take the generation fee multiplied by the attempts needed to get one usable result, then add the minutes a person spends checking and fixing it. For most shops the review time outweighs the model fee, so the saving comes from raising the pass rate, not from finding a cheaper model.
