AI Product Photography: A Practical Guide for Ecommerce

A shopper lands on your product page and does one thing before reading a single word. They look at the picture. Baymard Institute's usability testing found that 56% of people start by exploring the product images before the title or the description, and 42% try to judge an item's real size from those images alone. The photo isn't decoration. It's the pitch.

AI product photography can make that pitch for pennies. A studio day runs $1,000 to $3,000; a generated image costs cents. That gap is why every marketplace seller is now asking the same question, and why the real answer comes with a catch.

Here's our take: use AI to edit your real product photos. Never use it to invent the product. Get that one distinction right and this becomes the cheapest catalogue upgrade you'll make this year. Get it wrong and you're shipping fake products into a returns pile, with three regulators now watching.

The short version:

• Edit, don't invent. Start from a real photo of your actual SKU. Use AI for backgrounds, lighting and props, never for the product itself.
• Use a prompt formula. Subject, context, lighting, style, camera. One choice per slot.
• Fix your style once. Reuse a saved theme across the whole catalogue so 50 SKUs look like one brand.
• Meet the specs. Amazon wants 1600px for zoom. Most AI models cap at 1024px, so upscaling isn't optional.
• Keep the metadata. Amazon, Google and the EU now have dated rules on AI imagery. Don't strip the tags.

When AI product photography earns its place (and when it doesn't)

AI photos earn their place where the scene is the variable and the product is fixed. A candle that needs to sit on ten different surfaces. A backpack that should appear on a trail, on a train, and on a desk. You shoot the product once, cleanly, then let the model rebuild the world around it.

Keep the camera for anything where colour or scale accuracy drives returns: apparel fit and drape, jewellery detail, cosmetics colour matching, food, electronics with visible ports and text, and regulated goods. The test is simple. If "it looked different in person" is a common complaint in your category, stay edit-only or shoot real.

Etsy draws its own line. Its July 2024 Creativity Standards treat AI-enhanced photos of a genuinely handmade item as a tool, which is fine, but the process and packaging shots that prove the item is handmade have to be real. Read the room before you generate one.

The one rule: edit your product, never invent it

The whole game is whether your product's actual pixels survive.

Generate a product from a text prompt and the model hallucinates it: wrong logo, invented ports, garbled text on the packaging. Edit a real photo and the geometry, colour and logos stay intact. Reference-conditioned editors have made that edit workflow genuinely reliable. Black Forest Labs' FLUX.1 Kontext, released May 2025, preserves a reference object across new scenes. Google's Gemini 3 Pro Image holds identity across up to five subjects. ByteDance's Seedream 4.0 accepts up to six reference images at once.

Think of it like a stunt double. Great for the explosion in the background. You still want the real actor's face in the close-up.

The catch, and this comes from the model-makers themselves: even the good editors drift. Black Forest Labs says so in its own docs, that stacking multiple edits on one image can reduce quality and introduce artefacts. Glass and chrome warp first, so watch reflective and transparent products closely. Which editor holds a product best is its own question; we ran the same brief across three of them in Nano Banana vs GPT Image vs Recraft. Start from a real photo, and let AI touch everything except the product.

The mockup prompt formula that actually works

Vendor and practitioner guidance converges on one slot-based structure: subject, context or surface, lighting, style, camera. Make one deliberate choice per slot and revise the slot responsible for whatever looks wrong.

Camera vocabulary works because the models learned from captioned photographs, so the terms map onto real optical clusters. A focal length sets perspective: 50mm looks natural, 85mm compresses and flatters. An aperture sets depth of field: f/2.8 blurs the background, f/8 keeps the whole product sharp. Lighting terms set mood: softbox, three-point, high-key, seamless backdrop. The caveat most guides skip is that these are directional imitation, not physics. Stack a conflicting lens and lighting rig and the output degrades. One dial at a time.

For product work specifically, a 15 to 30 degree downward angle is the catalogue default. Flat-lay and knolling go top-down. Hero shots go low and look up.

Here's a template you can paste and fill:

A [product], on a [surface/context], lit by [lighting], in a [style] look, shot at [focal length] and [aperture], [angle].

Keep a 50-SKU catalogue looking like one brand

The hard problem at scale isn't one good image. It's 50 that look like they came from one shoot.

Two things drift: the style between generations, and the product's own colour. Practitioners guardrail both with a small library of approved scenes reused widely rather than a fresh idea each time, a strict accurate-colour rule, and a saved style reference instead of retyping the description. Seedream 4.0 can output up to nine consistent images in one go; FLUX Kontext holds a style across iterative edits.

This is where a saved theme does the quiet work. JammyJar's Themes feature extracts a look from a reference image and holds it across every generation, so you set the shelf, the light and the palette once and apply them to the whole range. Extract the look, reuse it, and your tenth SKU matches your first. Set a theme once, and stop redecorating for every product.

Resolution, and the specs marketplaces enforce

Upscaling is a mandatory step in the pipeline, not a nice-to-have.

Here's why. Recraft's V4.1 family and GPT Image Mini both cap output at 1024px, and Amazon wants 1600px or more to enable zoom. So a 1024px generation feeding an Amazon listing has to be upscaled before it goes live. Some models ease this: Gemini 3 Pro Image outputs 2K and 4K natively, and Seedream reaches 2048px. Know your model's ceiling before you promise a marketplace anything.

The rest of the specs are quieter but just as enforced. Work in sRGB, not CMYK, or your screen colours come out wrong. Amazon takes JPEG, TIFF, PNG and GIF, but not WebP, so save JPEG for the listing. For your own product pages, use WebP or AVIF instead: the smaller files load faster, which your Core Web Vitals will thank you for. Google Merchant Center recommends roughly 1500 by 1500px.

If your export step can't hit these formats and sizes cleanly, fix the export step before you fix the photos.

What Amazon, Google and the EU actually require

This is the part the thin guides skip, so here are the dated facts.

Amazon. The main image sits on pure white (RGB 255,255,255), with the product filling at least 85% of the frame, 1000px to enable zoom and 1600px+ recommended. Images must accurately represent the product, and AI editing is allowed as long as it doesn't mislead. As of 22 July 2026, Amazon requires sellers to tag any image containing a photorealistic AI-generated person with the IPTC keyword contains-synthetic-performer, prompted by New York's synthetic-performer law that took effect 9 June 2026 with penalties starting at $1,000. It exempts fictional characters and real people merely altered with AI.

Google Merchant Center. Since February 2024, AI-generated images must keep their IPTC DigitalSourceType metadata: TrainedAlgorithmicMedia for a fully generated image, CompositeSynthetic for a real photo with an AI-generated background. Google's own help documentation is blunt about not removing those tags, and stripping them risks disapproval.

The EU AI Act. Article 50's transparency obligations apply from 2 August 2026, with a grace period to 2 December 2026 for systems already on the market. They require machine-readable marking of synthetic image content. Penalties reach €15 million or 3% of worldwide turnover.

And for US sellers, the FTC's line is short: there's no AI exemption from the laws already on the books. A misleading image is a deceptive practice however it was made. The wider question of what you're actually licensed to sell is in can you use AI images commercially. The practical takeaway: keep the metadata your tools embed, and your compliance is mostly automatic.

The 60-second review checklist before you publish

Run every AI-edited image past this before it reaches a listing:

• Logo: intact, right colours, not warped or re-spelled.
• Colour: matches the real product on a calibrated screen.
• Text: any packaging wording is legible and correct.
• Geometry: ports, seams, handles and proportions are real, nothing invented.
• Scale: the shot lets a shopper judge true size (Baymard found 42% try to).
• Metadata: DigitalSourceType kept for Google, synthetic-performer tag added if there's an AI person.

The maths is unkind to shortcuts. One fake image that slips through costs you a return, a suppressed listing and a dented account metric, all of which run to more than a re-shoot. Check the six, then ship.

What it costs against a studio day

The cost story is real, and vendor copy oversells it. Traditional product photography runs roughly $25 to $500+ per finished image. A mid-tier studio day of $1,000 to $3,000 yielding 30 to 50 images lands around $40 to $67 each before retouching, with jewellery and reflective goods at the top of the range. AI per image is cents to low dollars: Recraft V4.1 Pro sits near $0.21, Gemini 3 Pro Image around $0.13 to $0.24, GPT Image models roughly $0.02 to $0.19.

The catch: those AI numbers ignore re-generation cycles, upscaling and human review, plus the cost of a bad image reaching a customer. And here's the part the tool blogs won't tell you. Nearly every "AI images lifted sales 40%" figure is vendor-sourced, and there's no rigorous independent A/B test of AI versus real product images in the public domain. So treat conversion promises with suspicion. The defensible wins are marginal cost and speed at catalogue scale, not some magic uplift.

Trust cuts the other way too. Gartner's March 2026 survey, run by Senior Principal Analyst Emily Weiss across 1,539 US consumers, found half say they'd prefer to buy from brands that avoid GenAI in consumer-facing content. Meltwater and YouGov put 86% wanting disclosure. Peer-reviewed work by Bui in 2026 finds disclosure tends to lower purchase intent, though Shi and Jiang show the same year that it's double-edged, raising perceived novelty even as it dents authenticity, with the net depending on the product. It's contested, not settled. We'd bet the accurate photo wins the return-rate fight, but nobody's run the closed-loop test to prove it. Which is the other reason to edit real photos: an edit of your actual product is accurate, and accuracy is the one thing you never have to disclose your way out of.

AI product photography FAQ

Can I use AI product photos on Amazon and Etsy?
Yes, with conditions. Amazon allows AI-edited images if they accurately represent the product and don't mislead, and needs 1600px+ for zoom. Etsy's July 2024 rules treat AI-enhanced photos of a genuinely handmade item as a tool, but require the proof shots, process and packaging, to be real. Generating a fake product for either is a fast route to a suppressed listing.

Do I have to disclose that an image is AI-made?
For metadata, often yes. Google Merchant Center has required the IPTC DigitalSourceType tag since February 2024. From 2 August 2026 the EU AI Act requires machine-readable marking of synthetic content. Amazon has required a synthetic-performer tag on AI-generated people since July 2026. Keep the metadata your tools embed rather than stripping it, and you're most of the way there.

Will AI images hurt customer trust?
They can, if they read as fake. Gartner's March 2026 survey found half of US consumers prefer brands that avoid GenAI in consumer content, and Getty and MarketCast reported that 76% of people can't reliably tell AI images from real ones anyway. The safer path is editing real photos, which stay accurate whether or not anyone notices.

Which AI model is best for product photos?
It depends on the job. Reference-preserving editors like FLUX.1 Kontext, Gemini 3 Pro Image and Seedream 4.0 are built to keep a real product consistent across scenes. For high-resolution output without a separate upscaling step, Gemini 3 Pro Image reaches 4K. Match the model to the task rather than committing to one for everything.

Do AI product photos need upscaling?
Usually. Several popular models, including the Recraft V4.1 family and GPT Image Mini, cap at 1024px, while Amazon wants 1600px+ for zoom. Generate first, then upscale to the marketplace spec before you publish.

So shoot the product, then rebuild the room

Your shopper still looks at the picture first, before the title, before the price. Give them a real product in a better room, not a convincing fake in a perfect one. Edit, don't invent: that's the whole rule, and it happens to be the one that keeps Amazon, Google and the EU off your back.

Pick your three best sellers, shoot each one cleanly this week, then let a model handle the backgrounds, the props and the light. JammyJar keeps the models, the saved themes and the export formats in one place, so you spend the afternoon editing instead of tab-hopping between five tools. Start jamming, it's free.