AI generation of product images for e-commerce

AI product image generation: where the flow breaks

The assortment grows faster than the shooting pipeline. While samples are traveling to the studio and passing through lighting and retouching, some items go on sale without visuals — with a single shot from a manager or an image from a supplier.

Product cards without photos lose before prices are even compared: they are opened less often and shown less often.

Catalog photography runs into the physics of the process, not the team's skill. Each SKU must be delivered, unpacked, shot from several angles, fitted into a unified style, and reshot if the packaging or shade changes. A studio produces dozens of items per day, while the catalog adds them by the hundreds.

The bottleneck in visual content is also visible in the consistency of the shots. Some are shot on a white background, some on gray, in some places the light is hard, in others soft, and the product's proportions in the frame vary.

In a general list, this reads as if the items are sold by different stores with different levels of service.

Priorities sharpen the picture: first new arrivals and seasonal selections are covered, then top revenue items, and the rest waits in line. The wider the assortment, the longer the tail of items that live for months with incomplete visuals and lose impressions.

AI image generation removes dependence on a single shooting slot. The source frame becomes the basis, while angles, backgrounds, and variants for the requirements of different channels are assembled digitally and scaled across the entire catalog. Let's break down how this is handled.

Measurable effect: timelines, consistency, catalog coverage

We start measuring the effect before launch: we record the median time to bring a new item into the catalog, the share of items with complete visuals, the number of reshoots per hundred cards, and the share of cards with standard violations.

We capture the same values monthly — this shows not 'it became prettier,' but a concrete shift in timelines and assortment coverage.

Technically, everything rests on the visual profile: once we agree on a reference set of source images, the lighting scheme, the list of angles, and the rules for cropping and color processing.

Each product goes through this profile, and at the output a check compares the file against the marketplace checklist — background, cropping, format, absence of extra elements. What does not pass goes for correction, not publication.

The time to bring a new item into the catalog is reduced from weeks to days: shooting is needed for the base reference, while the remaining angles and scenes are assembled from it.

Stable Diffusion paired with a mask along the product contour keeps the geometry unchanged — cards do not diverge in proportions and shade.

Assortment coverage grows thanks to items that were previously not reached: the long tail, rare colors, seasonal variants.

The share of items with complete visuals rises from part of the catalog to almost the entire matrix, standard violations fall to a few percent, and reshoots fall by a multiple: changing the background for a promotion is done from source images without a studio visit.

Image workflows: photography, 3D, generation

No single format always wins: the choice depends on the product category, the frequency of catalog updates, and how critical physical accuracy is.

Studio product photography provides reference color rendering and material texture, 3D product visualization covers angles and configurations not available in the warehouse, and AI image generation provides speed and volume where scene variability is needed.

In practice, formats are combined: one reference series is shot, and from it a model is assembled or synthesis is configured for a specific category.

The catalog scales without repeated studio visits, but the quality of the reference frame remains the boundary — a credible visual cannot be assembled from a blurry shot.

The common limitation of content formats is one: the image must correspond to the real product. Photo-based synthesis must not draw in properties the item does not have — otherwise the card is rejected on the marketplace and returns grow.

Therefore, we determine in advance which scene elements can be changed and which are rigidly tied to the source frame.

How we run the project: from catalog audit to handover

The work proceeds in waves: first we analyze the catalog and fix the rules, then we configure generation for categories and test it on a pilot sample of items, and only then do we launch the entire assortment.

This way an error in the rules is caught on dozens of cards rather than on the entire catalog, which would then have to be redone.

1. Catalog audit. We export the items, tag them by category, and note discrepancies: different backgrounds, angles, cropping, incomplete sets. The result is a map of problems and scope of work.
2. Brand rules. We agree on background, palette, angles, product position in the frame, shadows, logo, and prohibited techniques. We document them in text and references — acceptance is based on the document, not on taste.
3. Pilot sample of items. 20–40 cards from different categories and the most difficult cases: gloss, transparent items, small text on packaging, repeating patterns, sets. This shows where the rules do not work.
4. Category-specific setup. Each group gets its own set of rules: one angle and background for jewelry, another for furniture, a third for clothing. A single template for the entire catalog produces defects in complex categories.
5. Checking and corrections. We compare frames with the real product by shape, color, and details, calculate the defect rate, and refine the rules until the result becomes stable.
6. Acceptance of the result. We agree on criteria in advance: compliance with brand rules, product recognizability, absence of distortions, unified catalog appearance. Disputed frames are resolved before signing.
7. Handover. We provide the rules, references, the procedure for launching new categories, and rollback to original photos. After that, the catalog is replenished by your team.

Deliverables: rules, regulations, access, templates

The deliverables are not a folder of finished images, but a working toolset that the client's team uses on its own.

The client receives a repository with settings for specific product categories, regulations, and documentation: new seasonal assortment is run through the same rules without our involvement, rather than 'however it turns out.'

All artifacts are versioned and reviewed when changed. This means a configuration edit can be rolled back, and the difference between batches can be explained from history rather than from the performer's memory.

• Category templates: parameter sets for clothing, footwear, small electronics, cosmetics — accounting for material, gloss, and typical shooting defects.
• Angle and background matrix: how many views per card, which background for each placement channel, how complex-shaped products are processed.
• Catalog publication regulations: the sequence of steps from uploading source images to publishing, checkpoints, and those responsible at each.
• Access rules and roles: who changes templates, who only launches generation, who approves a batch — with an action log.
• Image rights document: origin of source images, terms of use, and what to do if a product is discontinued.
• Batch processing scripts: mass catalog run, restart after failure, automatic validation of results before publishing.
• Batch acceptance checklist: rejection rate, color and texture match with the reference, shadow consistency across the entire sample.
• Team instructions: analysis of typical failures, manual finishing of disputed frames, and the procedure for transferring work to a new employee.

Describe your categories and current shooting process — we will assemble a package for them and show how it fits your team.

Case study: a catalog of 40,000 items without a full reshoot

Before: visuals covered for a third of the catalog

The catalog grew faster than the shooting pipeline: out of 40,000 items, about a third had full-quality shots, while the rest were amateur photos from suppliers and images from other cards.

Preparing an item took two to three weeks — studio, retouching, coordination with the supplier, and reshooting when there were comments.

In categories such as fasteners and consumables, up to half the cards went out with only one angle: the buyer could see neither scale nor details.

What we did

We divided the assortment into groups by surface type, packaging shape, and the sales channel's requirements for background and frame margins.

For each group, we assembled a reference base from physical samples and configured generation tied to it: color and texture were verified with color patches, not by eye.

Each frame passed an automatic check — white background along the contour, intact shadows, a unified angle within the category; complex items with engraving and small text on the label were finished by a retoucher in Adobe Photoshop.

Metrics before and after

The catalog stopped depending on the studio schedule: new suppliers enter the assortment without reshoots, and visual preparation costs do not grow with the number of items.

Consistent frames within a category remove the cause of returns for 'not what was in the picture' — we include metrics for the remaining assortment sections in the same plan.

Retouching, 3D visualization, or synthesis: what to choose for a catalog?

There is no direct answer here: the visual format is chosen based on four criteria — product type, marketplace requirements, frequency of line updates, and availability of source images.

If source images exist and the product is one-off, it is more accurate and faster to finish the retouching of product photos. When the line updates frequently and hundreds of items differ by color or configuration, manual processing of each frame becomes a bottleneck.

Retouching maintains documentary accuracy — color, texture, seams, markings. Its limit is scale: shooting, selection, color correction, and processing are repeated for every frame, while marketplaces require a set of angles per SKU.

3D product visualization pays off where geometry and materials repeat: the model is assembled once, and then a new frame is a change of angle or environment. The limitation is that a three-dimensional sample and approved materials are needed.

Synthesis is used when there are few or no source images, but variations are needed for different marketplaces and usage scenes.

It is faster than other paths, but requires references and selection: small details and branded elements are generated inaccurately, and the final frame is finished manually.

Practical criterion: the higher the frequency of line updates and the number of items, the more 3D and synthesis win; the stricter the requirements for documentary accuracy, the more important retouching is.

More often the scheme is combined — the base frame is shot or assembled in 3D, and variations are obtained through synthesis: the labor intensity of frame preparation drops by a multiple.

Technical concerns: quality, rights, integrations

Objections to generation usually come down to three points: whether the buyer will distinguish a synthesized frame from a live shoot, who owns the rights to the result, and how frames get into the catalog without manual assembly. We address each one from an engineering standpoint.

Is it legally clean to use such images?

We train the model on our own stock and materials with a confirmed license. The contract records the transfer of rights to the result and the absence of restrictions on terms and placement channels.

For products with branded packaging, we separately check that the frame contains no third-party logos or recognizable elements.

How do frames get into the catalog?

Generation runs in batches from an export from the accounting system: files are named by SKU, proportions and background are set by category rules, so the listing does not 'fall apart' because of different shooting.

Missing angles are created automatically when a new SKU appears — without rerunning the entire catalog.

Let's discuss the task and come back with an estimate and plan

To come back with an estimate and work plan rather than general words, we need a snapshot of your catalog: volume is calculated from actual data, not by eye. Below is what actually affects the implementation scheme and timelines.

• Catalog export: SKUs, categories, current photos — it shows where generation replaces shooting and where frame finishing is needed.
• 10–15 products from different categories, including difficult-to-shoot ones: gloss, glass, textiles, small hardware.
• Marketplace requirements: frame size, background, angles, infographics — output templates are configured for them.
• What already exists: source images on a white background, 3D product models, or only photos 'as shot.' This determines what we train on your data and what we assemble from a template.
• Work rhythm: how many new items per month and how often cards are updated — a one-time run or a constant stream.
• Who accepts the result and by what criterion — so that acceptance is formal, not 'like it / don't like it.'

In response, we send an implementation scheme: where the model is trained on your photos, where template generation is sufficient, how product compliance is checked, and what we do with a questionable frame.

The estimate and work plan go together — with a breakdown by stages and the sample size for acceptance.

If some data is missing, this does not block the start: describe the task in free form, we will clarify the rest in a conversation and come back with an estimate and plan.