GPT Image 2.5 prompt — a transparent product cutout with clean alpha
Real alpha, not a painted checkerboard. The prompt has to say both.


| Mode | Model | Quality | Size | Ratio | Credits | Background |
|---|---|---|---|---|---|---|
| Image to Image | Think | Standard | 1K | 2:3 | 10 cr | PNG |
Published by OpenAI as an example for this model — not generated on this site. Model: gpt-image-2.5-sunburst. Source: OpenAI The published example ran at 1024x1536, a legal size we do not currently offer; the frame here is the same shape at our nearest size.
A shampoo bottle lifted off its background onto true transparency. This is the single most requested edit in e-commerce and it has one specific trap: asked for a transparent background, the model will sometimes paint the grey-and-white checkerboard that *represents* transparency in design software. The prompt forbids it by name, and so should yours.
The prompt, nothing cut
Extract the product from the input image and isolate it on a fully transparent background. Output: centered product, crisp silhouette, no halos/fringing. Preserve product geometry and label legibility exactly. Add only light polishing. Do not add a solid backdrop, checkerboard, scenery, or shadow. Do not restyle the product; remove the background and preserve clean alpha transparency.
This exact prompt, at these exact settings
The mode and all five settings are locked to this template, so there is nothing to line up by hand. Press Run this prompt above and the panel opens filled in — prompt, mode, quality, size and frame — and your first image needs no account. If your plan sits below what this ran at, the line under the panel says so.
This example was published at Think · Standard · 1K. A free run is Fast · Standard · 1K, so expect less fine detail and softer small text. The composition is the same; the finish is not.
What is safe to change
Most libraries publish this list and stop, which is why so many copied prompts come back worse than the original: the swappable nouns are the safe half. The three sections under it are the other half — the clauses that are doing the work, the ones that break it if you touch them, and why it is written in this order.
The edge quality clause
"Crisp silhouette, no halos/fringing" targets the exact artefact that makes a cutout unusable — the pale outline left where the old background bled into the edge. Hair and glass are where you will see it.
The preservation clause
"Preserve product geometry and label legibility exactly" stops the bottle getting subtly redesigned and the label getting re-typeset. For a real product, this is the line that decides whether the asset is usable.
"Add only light polishing"
A dial. Remove it and the model does more; strengthen it to "no polishing" and it does less. Useful when the source is already a clean studio shot.
Three ways to break it
Each one is a real failure with a reason attached. A rule without a reason is not usable.
01Omitting "checkerboard" from the prohibition list
The model has seen transparency drawn as a checkerboard in a vast number of design screenshots. Asked for transparency without ruling it out, it will sometimes deliver an opaque image of a checkerboard. This is the defining GPT Image gotcha.
02Saving as JPEG
No alpha channel, and no warning — it flattens to white silently. Use PNG or WebP, and skip output compression on PNG.
03Judging the result in a light-background viewer
A white-backed image and a transparent one look identical against white. Open the alpha channel, and look specifically at hair, glass, shadow edges and any semi-transparent packaging.
Inside the Cut a product out onto real transparency prompt
Background removal is the most-performed image edit on the internet and there are a hundred tools for it. What makes it worth a row here is that GPT Image 2.5 has a failure mode none of those tools have, and it catches almost everyone once.
The checkerboard. In design software, transparency is *displayed* as a grey-and-white chequered pattern. Screenshots of that convention are everywhere in the training data, reliably captioned with the word transparent. So a model asked for a transparent background has two ways to satisfy the request: return an image with an empty alpha channel, or return an image that looks like the pictures captioned "transparent background". The second one is fully opaque. It is a picture of transparency. OpenAI names this directly in its guidance — a drawn checkerboard is not transparency — and the prompt closes it with one word in a prohibition list.
Once past that, the two clauses that decide whether the asset is usable are the edge clause and the preservation clause. "Crisp silhouette, no halos/fringing" targets the pale outline that appears when the old background bleeds into the subject's edge, which is the artefact that makes a cutout look cheap when it is placed on a dark surface. And "preserve product geometry and label legibility exactly" is what stops a generative model doing what generative models do — improving things. A slightly redesigned bottle with a slightly re-typeset label is not your product.
That is also the honest caveat for anyone doing this at commercial volume. This is generation, not masking. The pixels are new. For most uses that is fine and often better, because the edges are cleaner than a mask would give you. For a regulated label or an exact product shot, verify the text and the silhouette against the original before it ships.
And check the alpha channel rather than the preview. Against a white page, an opaque white background and a transparent one are indistinguishable.
Cut a product out onto real transparency — common questions
- I asked for transparent and got a checkerboard. Why?
- The model painted the checkerboard as a picture of transparency. The image is opaque. Add "checkerboard" to the prohibition list, as the official prompt does.
- How do I check the transparency is real?
- Open the alpha channel, not the preview. Look at hair, glass, shadow edges and semi-transparent packaging — that is where fringing survives.
- Is this masking or regeneration?
- Regeneration. The pixels are new, which is why the edges are often cleaner than a mask — and why you must verify labels and geometry against the original.
- Which output format?
- PNG or WebP. JPEG has no alpha and flattens silently. On PNG, leave output compression off.
Written and maintained by Andy SwiftPublished Last updated





