From a simple photo taken on your phone, artificial intelligence can already swap the background, place the product in an attractive scene and create versions for Christmas, Black Friday or Mother’s Day in minutes. What it still cannot do reliably is copy your product faithfully: labels, shape, texture and size often come out wrong. The rule that solves almost everything is to use AI to edit the real photo, never to invent the product.
The topic is back in the news this week. On 4 November 2026, OpenAI launched ChatGPT Images 2.6, a new version of the image generator built into ChatGPT. A few days later, Artificial Analysis, an independent company that compares AI models, placed the two models of the new version first and second in its image generation and image editing rankings. For anyone running a small business that sells online, the practical question is just one: can you make your shop’s photos without hiring a photographer? It depends on what the photo has to prove. Below is what already works, where it still fails and how to do it without complications.
What changed this week
Images 2.6 reached every ChatGPT plan, on computer and phone. Three new features matter to people who sell. The first is the saved reference image: you store a photo of a character, a mascot or a person and reuse it in later creations without uploading it again. The second is short GIFs, those looping animations that work well in stories and ads. The third is the promise of keeping the product and brand details more consistent from one edit to the next, which has always been the weak point of these tools.
For developers, OpenAI added two models to the API: Flare, the faster one, and Sunburst, more detailed and with more consistent edits. The company says both generate images 50 percent faster than the previous version, but did not say what baseline it used for the comparison.
It is worth understanding what the ranking measures. Artificial Analysis shows two images made from the same request and asks people which one is better. It is a measure of preference, not of fidelity. A photo can win the vote for being beautiful and still show a bottle with the wrong label. In the editing ranking, in fact, the gap to competitors was within the margin of error: in practice, a statistical tie with the best rivals.
What AI already does well in product photos
Swapping and cleaning the background. This is the safest use and the one that saves the most time. You photograph the product on the kitchen table and ask for a plain white background, a background in your brand colour or a light wooden surface. Shadows and reflections come out consistent most of the time.
Building a scene (the so-called lifestyle shot). The mug on the breakfast table, the sneaker on the pavement, the lit candle in the living room. These photos sell because they help customers picture the product in use, and they used to require a production, a location and props.
Seasonal versions. With Black Friday at the end of November and Christmas right after, you can create the same product with decorations for each date without redoing the photo shoot. For anyone who posts every week, this solves the “I have no new photos” problem.
Adapting formats. The same photo as a square for the feed, vertical for stories and horizontal for the website banner, with AI filling in the missing space around the product instead of cropping it.
Supporting images and short animations. Backgrounds for promotional posts, illustrations for the blog and, now, a GIF of the product spinning or the packaging opening. None of this needs to be accurate to the millimetre, which is why AI does well.
Where AI still fails
Labels, text and logos. Small lettering is where AI stumbles the most. It redraws the label, swaps words, invents ingredients and distorts the logo. Even in the new versions, OpenAI itself warns, in the documentation for markup editing, that the tool does not guarantee that areas outside the marked zone stay untouched, and recommends checking faces, labels and lighting.
Shape and proportion. The bag strap gets longer, the bottle gains a different curve, the ring changes thickness. When you ask for a new scene, AI sometimes “recreates” the product instead of just moving it.
Real size. AI does not know your vase is 15 centimetres tall. It may place it next to a sofa as if it were half a metre. For the customer, that is a false promise.
Texture and material. Fabric, leather, brushed metal, glass and food have details that AI tends to make prettier and smoother than they are. The photo looks attractive, but it is not what arrives at the customer’s home.
Exact colour. The blue of your item can turn into a different blue. In fashion and home décor, colour differences are one of the most common reasons for returns.
Catalogue consistency. Across 30 products, keeping the same light, the same angle and the same background still takes work. Each image comes out slightly different, and the shop window looks like a patchwork quilt.
The rule that avoids trouble: edit the real photo
Almost all of the errors above appear when you ask AI to draw the product from a description. They drop sharply when you provide a real photo and ask it to change only what is around it. The step-by-step that works:
1. Take a good base photo. Window light, no flash, the whole product in frame, a neutral background and a steady phone. AI improves the surroundings, but it does not invent details the photo does not show.
2. Ask for one change at a time. First the background, then the light, then the scene. A request with five changes at once is where the product usually gets altered without you noticing.
3. Say what must not change. Example request: “keep the product exactly as it is, without changing the label, colour, shape or size; only replace the background with a light marble countertop with soft morning light”.
4. Compare side by side. Open the original and the edited photo together and check label, colour, seams, corners and proportion. Zooming in helps spot swapped letters.
5. Save the request that worked. Once you get the style right, use the same text for every product. That is what gives the shop window a catalogue look.
What marketplaces and the law expect
Marketplaces do not ban AI-made images, but they require them to represent the real product, and they are stricter with the main photo. Amazon asks, for the main image, for a pure white background and the product filling most of the frame, with no scenery. Mercado Libre also treats the main photo with tougher rules than the others and has a tool that assesses the quality of listing images. As of the date of this article, we found no official Mercado Libre rule specific to AI-generated images. So the safe path is the same on any platform: a real main photo with a clean background; scenes and seasonal versions in the following photos.
There is also the legal question. Consumer protection laws generally treat as misleading any advertising capable of deceiving buyers about a product’s characteristics, quality or quantity; in Brazil, for example, this is set out in the Consumer Defense Code. A photo that shows the product bigger, shinier or with a finish it does not have falls into that risk, besides generating returns and negative reviews. The question to ask before publishing is simple: if the customer receives exactly what is in the photo, will they be satisfied?
When a photographer, or 3D, is still worth it
AI handles the day-to-day well for anyone with few products who posts frequently. It is still worth hiring a professional when:
The detail is what sells. Jewellery, watches, textured cosmetics, food and clothes with drape need faithful photos. Here, AI comes in afterwards, for background and format, not in place of the shoot.
The catalogue is large and needs to be uniform. A well-planned shoot, with the same light and the same angle for every item, is still faster and more consistent than correcting image by image.
The product does not exist yet or has many variations. A launch before production, a piece of furniture with ten finishes, a technical part that needs to be shown from the inside: in these cases, the usual path is to model the product in 3D once and generate every image from the model. We explain how this works and what drives the price in our article on CGI rendering and why quotes vary so much. The same model then serves for video, a viewer on the website and augmented reality, and it is the kind of work we do in 3D modeling and animation.
How much it costs
For ChatGPT users, the image generator is available on every plan, including the free one, but with a tighter usage limit. Anyone creating images every week will probably need a paid plan. Through the API, which is the way to generate in volume inside your own system, OpenAI charges by data volume rather than per image. Estimates based on the official price table point to something between 4 and 5 US cents per square image at medium or high quality. The biggest cost is not the image, it is the time of whoever requests, checks and corrects it.
Checklist before publishing
Before uploading the edited photo to your website, marketplace or social media, check:
✔ Are the label, logo and any text on the product identical to the original?
✔ Does the colour match the real item, seen in natural light?
✔ Does the product’s size make sense relative to the objects in the scene?
✔ Are shape, straps, corners, seams and finish the same?
✔ Does the main photo show the real product, with a clean background?
✔ Would the customer be satisfied if they received exactly what they see?
If any answer is “no”, redo the edit from the original photo, asking for one change at a time.
Where to start this week
Pick your best-selling product and take three good base photos in window light. Use AI to create, from them, one version with a white background, one lifestyle shot and one for Black Friday. Post the lifestyle shot on social media, compare the results with the photos you used before and only then extend it to the rest of the catalogue. If you would rather leave this production to people who do it every day, take a look at our social media image production.
Frequently asked questions
Can I use AI-made product photos on Mercado Libre and Amazon?
Yes, as long as the image represents the real product. Both platforms are stricter with the main photo, which should show the real product on a clean background. The safest approach is to keep AI-made scenes and seasonal versions for the listing’s following photos.
Can AI create the product photo without me photographing anything?
It can create a beautiful image, but rarely a faithful one. Label, shape, colour and size often come out different. For selling, results are much better when you provide a real photo and ask AI to change only the background, the light or the scene.
Which tool should I use for AI product photos?
ChatGPT’s image generator, which got the Images 2.6 version in November 2026, leads the independent generation and editing rankings. Tools from other big companies come very close in editing. More important than the tool is the method: start from the real photo and ask for one change at a time.
What is a lifestyle shot?
It is a photo that shows the product in use or in an everyday setting, like a mug on the breakfast table or a sneaker on the street. It helps customers picture the product in their own life. AI can already build this kind of scene from a simple photo of the product.
Can using AI-edited photos cause problems with customers?
Yes, if the image shows something the product does not have, such as a bigger size, a different colour or a better finish. That leads to returns, negative reviews and, under consumer protection laws such as Brazil’s Consumer Defense Code, it can be considered misleading advertising. Always compare the edited photo with the original before publishing.
When is it still worth hiring a photographer?
When the detail is what sells, as with jewellery, cosmetics, food and clothing, and when the catalogue is large and needs the same light and angle for every item. For products that do not exist yet or have many variations, modelling in 3D and generating the images from the model usually works better.