Will a newer model stop the drift?

It helps. OpenAI said ChatGPT Images 2.5, released September 8, 2026, is better at preserving the subjects in reference photos. OpenAI and Google both still list text and consistency as limits, and Photoroom now offers enterprise customers a contract where they pay only for outputs that pass quality criteria they set. Check every image either way.

Why is my logo right in one image and wrong in the next?

Each render is drawn fresh, so each one is a new chance to drift. Keep the prompt and references identical across the set, and check the logo in every image, including the tenth.

Does a longer prompt keep the product more accurate?

Not by itself. In our 2026 campaigns, kept prompts ran about 140 to 200 words on Nano Banana models and about 300 on GPT Image. On the surf job, the longest character-sheet prompt was not kept; the kept one was four short lines with five well-chosen references. Length helps when it names what must not change.

Why does AI change my product, and how do I stop it?

Checked September 24, 20266 min read

Short answer

Because the model draws your product again in every image. It studies your reference photo and redraws the whole picture, product included, so small things drift: label text, logo shapes, proportions, cap colors, stitching. Better references, one clear job for each reference, a written list of what must not change and fewer props touching the product all cut the drift. When a label must be exact, place your real photo of it into the image afterward.

What the model does with your photo

An image model takes in your prompt and reference photos and produces a new picture, pixel by pixel. The product in that picture is the model’s drawing of your product, fitted to the new light and angle. The lighting can come out right while the fourth letter of your brand name comes out wrong.

The companies that make the models say as much. OpenAI lists keeping "recurring characters or brand elements" the same across images as a known limit of its GPT Image models, and says a mask, the marked area you want changed, is guidance the model may not follow exactly. Google says Nano Banana Pro may not render small text, fine details and spellings perfectly. OpenAI’s prompting guide adds that when an area must stay pixel-identical, you should composite it instead of relying on the prompt.

The changes you will see most

What changes	What it looks like	Usual cause	First fix
Label text	Swapped letters, extra words, made-up small print	No straight-on label photo, or text too small to copy	Add a straight-on label reference; composite if it must be exact
Logo shape	Rounded corners, wrong weight, merged letters	Logo small or at an angle in every reference	A close-up of the logo as its own reference
Proportions	Bottle taller, box deeper, neck thinner	One angle only, or a reference shot close up at 1x	Front and side references shot at 2x, and real dimensions in the prompt
Color	Red turns orange, a matte black cap turns glossy	Colors named in words only, or mixed light in the reference	Name each color with a hex code and its finish; reshoot in one light
Extra details	An added pocket, a second button, a pump on a jar that has a lid	The model fills gaps from similar products	List every part and end with "nothing added and nothing removed"
Pattern and fabric	Stripes change width, a print gets simpler, a weave goes smooth	No texture close-up	Add a texture reference and describe the weave
Hardware	Zip pulls, buckles and clasps change shape	Parts too small in the reference to read	A close-up of each part
Count	Five gummies where the pack shows four	The prompt does not state the number	State the count and check it

Seven fixes, in order of payoff

Start at the top. The first fixes are cheap and change every render after them; the last ones take more work per image.

Better references. Most drift starts with a photo that never showed the detail: a label at an angle, a back that was never photographed. Reshoot the missing views first. See how to photograph your product for AI.
One job per reference. Give each image only the photos it needs, product first, and say what each one is for, for example "image 1 is the product, image 2 is the label". Google and OpenAI both tell users to give each reference a role. The character sheet we kept on a surf campaign used 5 references and a short prompt, after attempts with up to 8 references failed.
Say what must not change. Google’s documentation advises describing critical details in great detail when you want them kept through an edit. Name the parts, colors and text, then close with "nothing added and nothing removed", a line from one of our kept prompts.
Fewer things touching the product. Hands, pours, props leaning on the pack and overlapping objects make the model redraw the edges where they meet. Start with the product alone, then add contact.
Pick a model built for keeping products. OpenAI recommends GPT Image 2.5 Sunburst "where editing precision matters most", and Google says Nano Banana 2 keeps up to 10 objects at high fidelity. Run the same shot ten times on two models and count the keepers; see which AI image model is best for product photos.
Edit only the broken part. Mask the label and ask for a fix, or ask a chat assistant to change only that area. Then compare the whole image with the last version, because OpenAI notes repeated edits can still change details you meant to keep.
Composite the real product. Cut the product out of your own photo and place it into the generated scene, then match light and shadow. The label in the final image is then your own photo of it.

A product lock you can paste

Put this block at the end of every prompt in a set, filled in from your own product. Keep the wording identical from prompt to prompt, so every render gets the same instructions.

Product lock (fill in the brackets)

The product is shown in image 1. Keep it exactly as shown.
Shape: [height] tall, [width] wide, [straight sides / rounded shoulders / tapered neck]
Colors: [part] [color name] [hex], [part] [color name] [hex]
Finish: [matte / gloss / frosted] on [part]
Text: the label reads "[exact words]" at [position]. Do not add, remove or respell any text.
Parts: [every visible part: cap, pump, zip, buttons, seams, stitching]
Nothing added and nothing removed. No text, logos or watermarks that are not on the product.

Excerpt from a kept prompt, GPT Image 2, trousers on a model (lightly edited: brand names removed)

The trouser must match INPUT IMAGE 2 exactly, navy smoke houndstooth pattern in a textured woolen weave with a matte finish and substantial drape, flat front with no pleats, angled slash front pockets, standard waistband with belt loops and an extended tab closure, a single dark button at the waist closure, no rivets, minimal tonal visible stitching, sharp center front crease, straight to mildly tapered leg with a plain finished hem, nothing added and nothing removed.

INPUT IMAGE 2 in that prompt was a flat lay of the trousers. The job it came from needed about 3.6 renders per usable photo, the second-lowest of our four 2026 campaigns.

When should you stop prompting and composite?

Composite from the start when:

The text is regulated or legal: ingredients, dosage, allergens, warnings, certification marks.
The image is a marketplace main image, where the product has to match what ships.
The logo is the point of the image.

For everything else, set a limit before you start, such as three tries on the same flaw, and switch to a composite when you reach it.

A composite is a cut-out of your real product placed into the generated scene, with a shadow added where it meets the surface. Plan the scene around the product photo you have. The light in the scene has to come from the same side as the light in your photo, or the product looks pasted in. For packs and labels in detail, see AI photos for packaged goods.

Questions people also ask

Will a newer model stop the drift?: It helps. OpenAI said ChatGPT Images 2.5, released September 8, 2026, is better at preserving the subjects in reference photos. OpenAI and Google both still list text and consistency as limits, and Photoroom now offers enterprise customers a contract where they pay only for outputs that pass quality criteria they set. Check every image either way.
Why is my logo right in one image and wrong in the next?: Each render is drawn fresh, so each one is a new chance to drift. Keep the prompt and references identical across the set, and check the logo in every image, including the tenth.
Does a longer prompt keep the product more accurate?: Not by itself. In our 2026 campaigns, kept prompts ran about 140 to 200 words on Nano Banana models and about 300 on GPT Image. On the surf job, the longest character-sheet prompt was not kept; the kept one was four short lines with five well-chosen references. Length helps when it names what must not change.

Where Overs fits

Overs sends each photo only the reference pictures its shot plan names, in order, and leaves out any it does not have instead of guessing. A quality check runs after the prompts are written and before the render, and on Pro and Team plans a reviewer compares each finished photo with the references and can retry on its own, up to a limit you set.

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.