AI product photos on Russian marketplaces: the rules from October 1, 2026

The short version: what changes on October 1, 2026

If you have no time to read, here it is in seven points. Each one is unpacked below with law numbers and amounts.

• Federal law 295-FZ of August 4, 2026 takes effect on October 1 and adds article 14.70 to the administrative code, written specifically for product listings on platforms.
• A seller faces 5,000–30,000 ₽ (sole traders and officers) or 30,000–70,000 ₽ (companies). A repeat offence goes up to 50,000 ₽ and 200,000 ₽ respectively.
• The offence is about paperwork, not pictures: goods without mandatory track-and-trace marking, without a conformity document, without state registration, or withdrawn from circulation. Platforms will check registry links and expiry dates algorithmically.
• There is no mandatory AI-content labelling in Russia. Law 243-FZ of July 26, 2026 makes the label voluntary for authors; a separate mandatory-labelling bill was rejected by the State Duma in spring.
• Platforms with more than 500,000 daily Russian users — Wildberries, Ozon and Yandex Market all qualify — must provide a labelling tool. That obligation starts on March 1, 2027.
• Wildberries has been showing an automatic “Looks like AI” tag on review media since July 6, 2026. Tagged content is neither hidden nor removed.
• The real exposure for AI imagery is not in the new rules at all — it is part 2 of the long-standing article 14.7: misleading buyers about a product’s properties, up to 500,000 ₽ for a company.

What 295-FZ introduces: the new article 14.70

The autumn package for the platform economy is three documents. The substantive rules come from laws 289-FZ (platform economy) and 290-FZ (fundamentals of trade), both of July 31, 2025; the penalties come from law 295-FZ of August 4, 2026. All three start working on October 1, 2026.

For a seller the key novelty is article 14.70, “Violation of the requirements for posting offers to sell goods”. It punishes not the quality of your photography but the fact that a listing offers something it may not: goods withdrawn from circulation, goods without mandatory marking, without conformity documents, or without state registration — which pulls in pesticides, supplements, medicines and medical devices.

• First offence: 5,000–30,000 ₽ for sole traders and officers, 30,000–70,000 ₽ for companies.
• Repeat offence: 10,000–50,000 ₽ for sole traders and officers, 70,000–200,000 ₽ for companies.
• Article 14.69 was extended separately and targets the platform itself, up to 500,000 ₽ — including for blocking a seller’s account and for discounts funded out of the seller’s pocket.
• A listing must carry full product information: permits, licences, certificates, marking data, warranty terms and the returns procedure.
• None of this is checked by eye. Platform algorithms verify that registry links exist, that documents are still valid, and that a document was issued to this seller for this product.

AI labelling: what was actually passed, and what was not

This is where most of the confusion lives, so here it is by date. In March 2026 the digital ministry put forward a bill making labelling mandatory for any generated photo, video or audio. On April 15, 2026 the relevant State Duma committee rejected it.

What ended up in force is federal law 243-FZ of July 26, 2026, “On supporting the development of artificial intelligence technologies”. It works the other way round from what most expected: authors carry no duty to label their own content, and there is no fine for not labelling. The duty sits with platforms that have more than 500,000 daily users from Russia — they must give users a way to add the label. That takes effect on March 1, 2027. Clarifying amendments to the same law were tabled on August 17, 2026.

The practical takeaway for a seller: today you are not required to state that a background was replaced by a model. But the labelling tool will appear on Wildberries, Ozon and Yandex Market, so decide in advance which content you are willing to label — and which should never need a label at all.

The “Looks like AI” tag on Wildberries, precisely

Since July 6, 2026 the Wildberries mobile app shows a “Looks like AI” caption in the bottom-left corner while a media file is open. It is applied automatically: the algorithm looks for signs of substantial editing or generative processing.

Two details usually get lost in the retellings. First, the tag launched on reviews — buyer photos and videos, not seller listings. Second, Wildberries stated explicitly that it will not hide or delete content because of the tag: it does not mean the review is untrue, it just helps a shopper judge what they are looking at before buying. The feature is still rolling out to part of the app audience.

Treat it as a signal of direction rather than a ban today. The detector already runs and is already visible to buyers. By the time the 243-FZ provision switches on in March 2027, the labelling plumbing will be in place.

Where the red line runs: article 14.7, part 2

The most expensive rule in this whole story is neither new nor about neural models. Part 2 of article 14.7 covers misleading consumers about a product’s properties or quality. The fine is 3,000–5,000 ₽ for individuals, 12,000–20,000 ₽ for officers and 100,000–500,000 ₽ for legal entities.

The law does not care what you edited the photo with — Photoshop, a generative model or good lighting. Only one thing matters: does the image match what the buyer receives. A sweater given a pattern it does not have; trainers with a panel recoloured in generation; a set shown as four items and shipped as two — that is article 14.7 regardless of the tool.

Which is why a safe AI pipeline always starts from a real photograph of the real item. Not “generate me a thermos”, but “take my thermos and remove the kitchen table behind it”.

Shot on a phone

Ready to publish

The same thermos: background, light and cleanliness change — shape, colour and cap stay exactly as they are in real life

What you may do with a model

The dividing line is not “AI versus not AI” — it is “you showed the product” versus “you showed something else”. In practice it breaks down like this.

Allowed — anything that does not change the object itself:

• Removing or replacing the background, cutting the item off a table and placing it on white.
• Evening out the light, killing blown highlights, noise, reflections and lamp glare.
• Cleaning fingerprints and dust off metal and glass that a new unit would not have anyway.
• Building a lifestyle scene: the same item in a kitchen, an interior, outdoors — as a usage scenario, not as a substitute for the product shot.
• Assembling infographics from the real specifications already in the listing: volume, material, dimensions, contents.
• Drafting the description and the spec list with a model, then checking every number against the product documentation.

And what you may not — whatever the tool

Every item below is either a moderation rejection, or an article 14.7 case, or both.

• A fully generated product standing in for a real photograph. Worst of all in clothing, footwear and jewellery, where buyers read texture.
• Changing colour, texture, pattern, shape or contents relative to what is actually in the warehouse.
• Invented details: a zip, a pocket, a panel or a strap the item does not have.
• Generated “customer photos” and reviews — on Wildberries, review media with generative traces is already tagged automatically.
• A model wearing a garment that was never put on them, when the real fit differs from the rendered one.
• Marketing copy, prices, discounts and third-party logos on the main image — a platform rule in force since 2025, carrying a 25,000 ₽ Wildberries fine, and nothing to do with generative tools.
• Food rendered from scratch. It is the category where buyers spot the gap instantly and put it straight into the reviews.

The photo requirements nobody repealed

The new rules do not replace the platforms’ technical requirements — moderation still checks them, and they are where listings assembled by eye usually fail.

Wildberries expects a vertical 3:4 frame, at least 700×900 pixels (900×1440 is the working standard), a white background for the product shot and the item filling at least 70% of the frame, centred. Ozon wants the same vertical ratio and a clean background on the main slide. Watermarks, advertising captions and third-party logos are banned on both.

Generation has a systemic weakness here: a model knows nothing about 3:4 or about 70% coverage. Ask for a “beautiful product photo” at 16:9 and moderation will bounce the card — not because of AI, but because of geometry. Set the aspect before you generate rather than cropping afterwards.

One real thermos on three backgrounds: white for the main slide, kitchen and outdoors for the scenario slides

How to build a listing safely: four steps

Order matters more than speed here — almost every problem comes from checking the paperwork after publishing.

Step one: photograph the item yourself. A phone, a window and a sheet of white paper are enough; what matters is not the studio but that the frame holds your physical unit with all its quirks. That frame becomes the anchor — everything after it is editing, not invention.

Step two: fix the background and the light. Nano Banana Pro holds the object best while the surroundings change, GPT Image 2 follows a long brief most literally, Seedream 4 and FLUX.2 give the most studio-like render. Check the result against three things: colour, shape, contents.

Step three: assemble the documents before you publish. Certificate or declaration, marking code, registration data, validity dates, and matching legal details. From October 1 this is exactly where article 14.70 lives — not in the pictures.

Step four: check geometry and text. 3:4, item no smaller than 70% of the frame, not a single caption on the main slide. Infographics start from slide two and only from real specifications.

Which model for which job, and what it costs

In Twin AI every model runs off one credit balance — no separate subscription per tool and no foreign card. Cost of one generation at default settings:

• FLUX.2 Pro — 200 credits. A fast way to run many background variants for one product.
• Wan 2.7 Image — 200 credits. Open weights, solid on object geometry.
• GPT Image 2 — 200 credits. Follows a long brief most precisely: “the same thermos, white background, contact shadow, 3:4 frame”.
• Seedream 4 — 200 credits. Dense studio-style rendering for product shots.
• Qwen Image — 200 credits. Strong on legible lettering, for infographics from slide two onwards.
• Nano Banana 2 — 200 credits. Fast background swap with the object preserved.
• Nano Banana Pro — 200 credits. The best object retention under a full environment change — the workhorse for listings.
• The /compare mode sends one prompt to several models at once: for a hundred SKUs, comparing once is cheaper than redoing a batch.

Checklist before October 1, 2026

Walk it once; after that it folds into the normal upload routine.

• Confirm every SKU has a valid conformity document attached to the listing and that it has not expired.
• Match the legal details: the document must be issued to your company or sole trader entity, for this exact product.
• Check mandatory marking codes wherever the category requires them.
• Re-inspect photos for divergence from the real item: colour, pattern, contents, invented details.
• Strip marketing copy, prices and non-brand logos from the main slides.
• Check geometry: 3:4, at least 700×900, item no smaller than 70% of the frame.
• Decide in advance which content you will mark as generated once the platforms ship the labelling tool.

Common seller mistakes in 2026

What usually goes wrong for sellers who moved their listings onto generative tools:

• Assuming the new fines are about AI. Article 14.70 is about documents; pictures are governed by article 14.7, which has always been there.
• Waiting for mandatory labelling instead of preparing for voluntary labelling. The duty sits with the platforms, the tool is coming, and deciding what to label will be your call.
• Generating a whole product “just to test” and leaving it in production. A review photo of the real item next to your listing is the most expensive proof of mismatch there is.
• Building infographics from invented specifications. A model will happily add “keeps heat for 24 hours” on its own — and you are the one answering for the number.
• Generating at 16:9 and cropping to 3:4. The item either leaves the frame or drops below 70% coverage, and moderation bounces it.
• Reusing one background everywhere. White is mandatory on the main slide, while scenario shots sell better — and belong from slide two.