How long does one AI product photo take?

OpenAI says a complex prompt on its image models can take up to 2 minutes to render. Choosing takes longer. In our work a usable photo took about 8 renders, so most of the time goes into reviewing them.

Do I have to tell customers a photo was made with AI?

It depends on where you sell and advertise. Google Merchant Center requires AI-made product images to keep their AI metadata, and its help center notes that AI rules in the EU, India and New York require disclosures or labels on certain ads with AI-made or AI-edited images. See do you need to label AI images in ads.

Can AI make photos of a product that does not exist yet?

It can draw an idea. Without photos of a real product, though, the model invents every detail. Treat those images as concepts, and photograph the first real sample before you sell from pictures of it.

What is AI product photography?

Checked September 24, 20266 min read

Short answer

AI product photography means making product photos with an image model instead of a camera. You give the model photos of your real product and describe a scene, and it draws a new picture with your product in it. The model redraws the product each time, so labels, logos and shapes can change, and every result needs checking against the real thing before you use it.

How AI product photography works

An image model is software that learned what pictures look like from a large collection of sample images. You give it a written request, called a prompt, and one or more reference images: photos of your product, and sometimes of a person, a place or a look you like. It returns a new picture, called a render. Every pixel of a render is drawn fresh, including the product.

That makes it different from editing a photo. When you retouch a picture or remove its background, the product’s pixels still come from your camera. When a model generates a scene, the product is the model’s copy of what it saw in your reference, and copies drift. Google Merchant Center sorts images the same way, with separate metadata labels for images made by a model and for composites that include made-up parts.

Method	Where the product comes from	Example	Risk to accuracy
Edit your photo	Your camera	Background removed and replaced with white	Low
Composite	Your camera, placed into a generated scene	Your cut-out bottle on an AI-made kitchen counter	Low for the product; light and shadow may not match
Generate from references	The model’s redrawing of your photos	A render of your bottle on a beach	Higher, since text, logos, colors and shape can change

Three ways a product can end up in an image, from least to most redrawn.

Generated images also carry marks. Google’s developer docs say every image its Nano Banana models make includes an invisible SynthID watermark, and Google Merchant Center tells sellers not to strip the metadata that marks a product image as AI-made.

Four ways brands make AI product photos in 2026

Tools fall into four groups by how you work with them. The groups overlap, and none suits every job. For picks by task, see the best AI product photography tools.

Way	Examples	What you do	Good for	Watch for
Product photo apps and store tools	Photoroom, Google Pomelli, Google Product Studio	Upload a product photo, then pick a template or describe a scene	Fast catalog and social images	Results follow the app’s templates and scene options
Chat assistants	ChatGPT (Images 2.5), the Gemini app (Nano Banana 2)	Attach photos and ask for an image in a chat	Trying ideas, single images	Keeping a set of images consistent with each other
Node canvases	Figma Weave, Flora	Wire inputs, AI steps and outputs into a workflow you can run again	Agencies and teams that repeat the same job	Setup time and skill; you pay for credits or model use
Brief-in tools	Overs	Give a product photo, brand details and what the set is for; the tool plans and makes the shots	A planned set in one brand style	You still check every image against the product

Examples checked on September 24, 2026.

Some specifics, as of September 2026. Google Labs added Photoshoot to Pomelli on February 19, 2026. You pick a product and a studio or lifestyle template, and it makes branded shots with Google’s Nano Banana model, for free. Photoroom’s AI Shot List makes studio, lifestyle, ghost mannequin, flat lay and try-on versions from one product photo, with a preview before any credits are spent. OpenAI rolled out ChatGPT Images 2.5 to every ChatGPT plan on September 8, 2026. Figma Weave’s paid plans start at $24 a month and Flora’s at $18.

What does AI do well, and what does it get wrong?

The companies that make the models publish their limits. OpenAI says its image models can struggle with precise text and with keeping recurring characters and brand elements the same across images. Google says Nano Banana Pro may not render small text, fine details and accurate spellings perfectly. The tries-per-photo numbers below come from four campaigns our agency made in 2026.

Job	How well AI does it	Why
New scenes for a product you have photographed well	Well	The scene is invented anyway; only the product has to stay true
Many versions of one idea: sizes, seasons, backgrounds	Well	Each new version costs cents instead of a reshoot
The product alone, with no words in the image	Well	Our product-only supplement ads took about 3 renders per usable photo
Physics: pours, crumbs, things breaking	With work	Our brownie ads with cocoa in the air took about 12 renders per usable photo
The same person in every photo	With work	Faces drift; our surf shoot on one model took about 14 renders per usable photo
Small text on labels	Poorly	OpenAI and Google both list small or precise text as a limit
Exact fit of clothes on a body	Poorly	The model decides how fabric falls; it has never seen your size chart
A label or logo that must match to the pixel	Not by prompting	OpenAI’s own guide says to composite any area that must stay pixel-identical

For the same question shot by shot, see can AI replace a product photo shoot. For the fixes that make renders pass as photos, see how to make AI product photos look real.

What you need to start

Sharp photos of the product from every side you will show, plus the label straight on. See how to photograph your product for AI.
A few lines on your look: colors as hex codes, the kind of light, the camera distance, and what never appears.
A list of the images you need and where each one will run. See what product photos an online store needs.
Someone with time to check every image against the real product before it goes live.
A budget for tries. In our four campaigns, one usable photo took about 8 renders.
An account with a tool or a model provider. In September 2026, published model prices ran from $0.006 for a low-quality square draft on OpenAI’s GPT Image 2 to $0.24 for a 4K image on Google’s Nano Banana Pro.

For the full cost picture, with subscriptions and credits, see how much AI product photos cost.

Words you will see

Word	What it means
Render	One image the model returns. Most renders are not used.
Keeper	A render someone chose to use.
Reference image	A photo you give the model to copy from: the product, a person, a place or a style.
Prompt	The written request: what to show, where, and in what light.
Image model	The AI that draws the picture, such as Google’s Nano Banana 2 or OpenAI’s GPT Image 2.
Packshot	The product alone, usually on a white background.
Lifestyle shot	The product in a real-looking place, often in use.
On-model	Clothing or accessories shown on a person.
Character sheet	One image of the same person from several angles, used to keep that person the same in every photo.
Composite	A real product photo placed into a generated scene.
Mask	A marked area that tells the model which part of an image to change. OpenAI says its models treat a mask as a guide.

Questions people also ask

How long does one AI product photo take?: OpenAI says a complex prompt on its image models can take up to 2 minutes to render. Choosing takes longer. In our work a usable photo took about 8 renders, so most of the time goes into reviewing them.
Do I have to tell customers a photo was made with AI?: It depends on where you sell and advertise. Google Merchant Center requires AI-made product images to keep their AI metadata, and its help center notes that AI rules in the EU, India and New York require disclosures or labels on certain ads with AI-made or AI-edited images. See do you need to label AI images in ads.
Can AI make photos of a product that does not exist yet?: It can draw an idea. Without photos of a real product, though, the model invents every detail. Treat those images as concepts, and photograph the first real sample before you sell from pictures of it.

Where Overs fits

Overs is a brief-in tool. You give it one product photo and your brand details, it plans the shots and shows an estimated cost, and it makes the photos after you say yes. The plan is free for 40 photos a month; the AI itself costs a few cents a photo through your own OpenRouter account, a service that gives you one account and one bill for many AI models.

Free for 40 photos a month. The AI that makes the photos is billed separately, on your own key, with no markup from Overs.

Start free How it works

Sources

Checked on September 24, 2026. Prices, specs and rules change; follow the links for the current versions.