Make a Product Photo

How it works

Your photograph is scaled so its total area is around one megapixel, keeping the original proportions. An image-editing model reads that copy along with a fixed instruction: create a clean commercial product photograph on the scene you selected, preserving the exact product shape, label, logo, color, quantity, materials, and proportions from the reference, improving composition and lighting without inventing product features or accessories.

The second instruction is unusually specific, and it is worth reading as a list of expected failures rather than a formality. It rules out changed product shape, changed logo, misspelled text, extra products, extra accessories, distorted packaging, and watermarks.

The model then generates a new photograph at high strength.

Two jobs pulling in opposite directions

There is a tension built into this workflow, and being clear about it is more useful than pretending it is solved.

The strength is set high. That is what allows the scene to be genuinely rebuilt — a new surface, new lighting, real shadows, a background that belongs with the object rather than sitting behind it. At a low strength you would get your original photograph with the colors pushed around, which is a filter and not worth a GPU.

But high strength applies to the whole frame. The model is not given a mask that says the product is off limits and the rest is free. It is given an instruction that says so, in words, and instructions are guidance rather than enforcement.

So the graph asks for maximum reinvention of the scene and zero reinvention of the object, using one control that governs both. Most of the time the wording holds, because the product is the salient thing in the reference and the model follows salience. When it does not hold, it fails on the product, which is the only part that had to be right.

What you are allowed to invent

The useful way to think about this is to separate the picture into two categories.

The scene is yours. Backdrop, surface, lighting, shadow, reflection, depth of field, color grade, the styling of the setting. Nobody buys a backdrop, and making one up is what studio photography has always been.

The product is not. Shape, proportion, color, material, finish, the wording and placement of the label, and above all the quantity — a two pack that arrives as a picture of three is a specific and expensive kind of mistake.

And the model will rewrite your label anyway, so check it every time. It cannot spell. Given a plain tube on 13 August 2026 it returned a clean packshot with MnocIntiget printed across the front — confident, well-set, and not a word. Any lettering on the packaging is at risk: the brand name, the variant, the weight, the claims. This is the image model's limitation rather than a setting, and it is the single most likely way a picture from this page becomes one you cannot publish, because everything else in the frame will look right. If the label carries text that matters, keep the original product area or set the type yourself afterwards.

The lifestyle preset sits awkwardly across that line. Props are scene, so inventing them is fair. But a prop next to a product reads to a buyer as something that comes with it, and a picture that implies an accessory you do not ship has crossed from scene into claim without changing anything visually.

The starting photograph matters more here

For most pages on this site the advice is to use a good photograph. Here it is more specific than that.

The reference is the only description of the product the model receives. If your label is out of focus, the model has to guess at the wording. If the color is off because of the light you shot under, the generated version inherits the wrong color and renders it confidently under better light. If part of the object is hidden, it gets invented.

None of that is visible in the result, because the result looks like a competent photograph either way. A confident render of a wrong shape and a confident render of a right one are indistinguishable without the reference beside them.

So the reference should be sharp, evenly lit, and complete, even if it is unattractive. Ugly and accurate is the correct input. The picture is the part this page is for.

Publication gate

This page ships once the workflow has been run on both presets against a labeled bottle, a multipack, and a product with a reflective surface, with each output compared to the reference for shape, color, count, and label wording.