Why Does ChatGPT Change My Product Even With a Reference Photo?

ChatGPT changes your product because it does not paste your photo into a new scene. It draws a whole new image, product included, and uses your photo as guidance, so a handle, cap, pocket, color or logo it does not hold onto exactly can come back different. In a chat, each follow-up usually works from the last image rather than your original photo, so a small change carries into every image after it. OpenAI says ChatGPT Images 2.5, released on September 8, 2026, is better at preserving the subjects in reference photos. Better is not the same as copied: the product is still redrawn every time.

When the image has to show the product you sell, start every new image from your original photos and compare each result with the real item. Nightjar builds the first habit into the workflow and adds a check of its own. Save the item as a Product, Nightjar's reusable record of its photos and facts, and every new image starts from that evidence instead of the last image you made. Nightjar's brief tells the image step which details must stay exactly as shown, and its built-in visual review compares the result with your photos and can retry a plainly changed product at no extra Credit cost. Your own full-size comparison still matters for subtle differences, which the review is not built to catch.

Why doesn't a reference photo keep my product the same?

A reference photo is evidence, not a layer. ChatGPT, Gemini and Midjourney all draw the new image, product included, and your photo guides that drawing. Details the model reads clearly and treats as important tend to survive. Details that are small, partly hidden, soft, or in tension with the rest of your request are the likeliest to change, and where the photo leaves room, the model tends to fill it with a plausible version of that kind of product: a rounder handle, a more typical cap, a pocket that was never there.

Here is what each vendor says about its references:

ChatGPT: OpenAI's announcement of ChatGPT Images 2.5 says the update is better at preserving the subjects in your reference photos and follows editing instructions more reliably across multiple turns. That is a real improvement, but OpenAI's developer guide for its GPT Image models still lists consistency as a limitation: the model "may occasionally struggle to maintain visual consistency for recurring characters or brand elements across multiple generations." The same guide says the model uses a mask "as guidance", so selecting an area narrows an edit without locking the rest of the product.
Gemini: Google says its Nano Banana 2 model can maintain "the fidelity of up to 14 objects in a single workflow" without "altering the appearance of your inputs" (Google), and its Gemini Apps help still asks you to "use your discretion before you rely on, publish, or use content that Gemini Apps generate."
Midjourney: Midjourney's Image Prompts documentation says it uses references as inspiration for new creations rather than copying them exactly.

Midjourney says plainly that it does not copy references exactly, and OpenAI lists consistency as a limitation. Google makes the strongest fidelity claim of the three and still asks you to use discretion before you publish. Either way, the comparison with the real item stays with you. If the change you keep seeing is in the label or logo, how to stop AI from garbling the text and logos on your product covers the label-specific fixes. For a reshaped silhouette, see how to prevent AI from altering the product's shape, and for a shifted shade, why AI changes your product's color.

Why does one wrong detail carry through the rest of a chat?

Because each follow-up builds on the image before it. When you ask for "now put it on marble" or "make the light warmer", the edit is usually applied to the latest image, not to your original photo. OpenAI's developer guide describes multi-turn image editing the same way, with each new turn referring back to the previous image to refine it. Whatever changed in that image is now part of the starting point.

OpenAI says Images 2.5 makes those conversations steadier: it follows editing instructions more reliably across multiple turns, and remio's coverage of the launch reports OpenAI also saying that earlier changes are more likely to stay consistent through a longer conversation. That helps when every earlier change was one you wanted. But a steadier chat holds an unnoticed change just as steadily, because it stays consistent with its own last image, and that image may already differ from your product.

A hypothetical example shows how it plays out. You upload a ceramic mug with a square handle and a two-tone glaze. The first image comes back with a slightly rounded handle that you don't notice at a glance. You ask for a marble counter, then for morning light. Both follow-ups start from the rounded handle, and the better the chat is at keeping earlier details, the more reliably it keeps that one. By the third image the wrong handle looks settled.

The fix is the same however steady the model is: go back to your original photos. Inside a chat, three habits keep an early change from carrying through:

Start each new image from your original product photos in a fresh request, rather than asking for another change to the last output.
Ask for one change per request, so you can see what moved.
Compare each result with the real product before you build on it or publish it.

If you only need one small fix to an image you already like, that is a different job: see how to change one thing in a product photo without regenerating it.

Is switching to another image model enough?

Not if the image has to show your real product. A model that follows references more closely can make drift less frequent, and models keep improving. The workflow stays the same, though: for every image you choose which photos to upload, restate what must not change, and catch every changed detail yourself. Those three jobs are where drift gets in: a forgotten close-up, a detail you didn't think to mention, a check done at thumbnail size on a busy day.

So the deciding question is not which model follows references best. It is whether your workflow brings the same product evidence, and a check against it, to every image. That is the gap Nightjar fills: a saved Product carries the evidence into each new image, and Nightjar's review checks the result against it. For side-by-side comparisons of the general tools with dedicated ones, see ChatGPT alternatives for product photography, Google Gemini for product photos vs dedicated AI tools and Midjourney for product photos vs dedicated tools.

What should I use when the image has to show my real product?

Use Nightjar when you need new images of the same real product, and a cutout composite when the product's pixels must not change at all.

Nightjar avoids the chat drift problem by changing where each image starts. A saved Product keeps your photos of the item together with, if you add them, a factual description and its real dimensions, and every new Product Photography image is built from those, not from the last image you made. Before generating, Nightjar's brief is written to list the product's identity details, such as its silhouette, construction, color, material, printed text and logos, and to require each one to stay exactly as shown in the reference photos, which outrank the description. After generating, built-in visual review compares the result with your photos and rejects plain product changes a shopper would notice at normal size, such as a reshaped handle, a missing or extra part, a clearly different color, a changed print or a replaced main logo. Nightjar can then retry at no extra Credit cost; Credits are Nightjar's unit for paid actions.

The review is built for changes visible at normal size and lets close calls through, so subtle drift, such as a slightly narrower strap or a seam in a different place, can still pass. Compare each image with the product at full size before you publish it.

What changes compared with a chat is the work you carry. You stop re-uploading photos and re-describing the product for every image, and you stop editing a drifted image into the next one. Each image is a fresh attempt from the same evidence, and an obvious product change can be caught and retried before you see it.

To set it up:

Open Products, select New product, then choose Manually to upload your photos, or From a link to review the photos and details from your public product page before saving. Importing a Product from a product page covers that route.
Choose Make main on the clearest full view of the item, because Nightjar uses up to five photos for each image and starts with the Main photo. Then add close-ups of anything that must not change, such as a handle, closure or label, and a factual description and dimensions if you have them.
In Create, select the Product, choose the setting and select Create Image. For the next image, select the Product again rather than starting from an output.

How to create and manage Products in Nightjar has the full steps. The chat lesson applies here too: add a generated image to a Product's photos only after you have checked it, because it then becomes evidence for future images.

For a single image you don't need a Product: add your photo under Additional photos in Create and generate, and Nightjar's brief and review still apply. Saving a Product pays off when the next image has to match.

If you also plan your products in Claude, you can connect Nightjar there through MCP, the open standard AI assistants use to connect to other tools, and ask for the images in the conversation; they come from the same saved Products and review (making product photos in Claude, setup and limits).

Nightjar's Product Photography is reference-guided generation, not a pixel lock. When packaging artwork or a regulated label must stay exactly as photographed, cut the product out of your real photo and composite it into the new scene, or keep the original photo for that image.