GPT Image 2.5 Image to Image
Change one thing. Keep the rest.
Upload the picture, tick what must not move, and describe the single change you want.Every edit re-renders from your original file, never from the last result — so a fifth edit is as clean as the first. The tick list becomes OpenAI's own change only X phrasing before the request leaves, and the receipt under the result names the model that ran.
Worked edits · original included
Before
AfterModify the messy bed to be neatly made with a smoothed-out white comforter. Arrange the pillows neatly and symmetrically against the headboard, and remove the pillows on the floor. Keep the room structure, bed frame, bedside tables, lamps, lighting, open door, window, and rug identical.
- Fast
- Standard
- 1K
- 3:2
- PNG
The room, the portrait and the dog are OpenAI's own published before/after; the three drawing-to-photograph pairs are reference material held for this page. Prompts are written in the change only X shape the tick list writes for you. The pairs carrying Full prompt come from the prompt library, written up with what they cost and what breaks them.
- Free preflight
- Tick what must not move
- Never chained
- Nothing is overwritten
- Model-ID receipt
- Failed runs are free
1 free image, no card, no login · failed runs cost nothing · your original is never overwritten
What's on this page about GPT Image 2.5 image to image
What GPT Image 2.5 image to image actually does to your photo
Image to image — also written img2img, and sometimes called picture-to-picture or photo-to-photo AI — takes a picture you already have and uses it as the anchor. Your prompt rewrites only what it names. Composition, pose, perspective and framing survive the trip. It is not a filter, and it is not a text prompt with a mood board attached: the source photo is doing most of the work and the sentence is just steering.
Starts from
- Image to Image
- The photo you already have
- Text to Image
- A blank canvas
What survives
- Image to Image
- Composition · pose · perspective · framing
- Text to Image
- Nothing — everything is invented
What changes
- Image to Image
- Only what your prompt names
- Text to Image
- All of it
Use it when
- Image to Image
- You've already got something worth keeping
- Text to Image
- You're building an idea from nothing
Every edit starts from your original, not from the last result
Chained editing is where quality goes to die. Each pass re-encodes the pass before it. Edit four times and you are four generations away from your source, carrying every compression artifact the previous three picked up. A ChatGPT Plus user described the result on the OpenAI forum: "The effect is very noticeable with just 2 edits to the same image" — and he ended up breaking out Photoshop.We don't chain. We keep your uploaded file, merge everything you have asked for into one instruction, and re-render from the source. Edit number four is computed from the same pixels edit number one was.That is a claim about how we call the API, not about the model — and there is a reason to believe it matters. An OpenAI community moderator, replying to that same complaint: "the more times I edit the same image, the more the quality tends to degrade, especially when editing through ChatGPT. I haven't noticed the same degree of quality loss when using the Image API for edits, though."
Two controls make it visible: Back to original throws away everything and returns you to the file you uploaded. Merge my changes folds every request you have made into one prompt and runs it once, from the source.
Six ways in, and each wants a different sentence
Most disappointing image to image results aren't a model problem. They're a mismatch: you asked for a background swap using the phrasing of a restyle, and the model did what you actually said.
Change one thing
- Use it when
- One element is wrong — a product, a background, a line of copy
- How to phrase it
- Write
change only X, then list what must survive. This is the mode most people should be in and aren't. No mask to draw and no inpaint region to select — where Photoshop wants a selection, this wants a sentence.
Swap the background
- Use it when
- New scene behind the same subject
- How to phrase it
- Describe the new background and the light direction, then say the subject's lighting must match it. Skip that and you get a cutout floating in someone else's photo.
Restyle
- Use it when
- Oil painting, pixel art, anime, film stock
- How to phrase it
- Say whether you want it repainted or tinted. The model repaints by default, which is usually not what people mean. A restyle is not a retouch: ask to repaint and you get a new surface, ask to retouch and you get the same photo, cleaned.
Combine references
- Use it when
- Two to five images into one
- How to phrase it
- Every image needs a stated job. Unlabelled references get averaged.
Cut out to transparent
- Use it when
- You need a PNG with a real alpha channel
- How to phrase it
- Native on the model — but not exposed here yet. Ask for a flat single-colour field and key it out.
Sketch → finished
- Use it when
- You drew it badly on purpose
- How to phrase it
- Say the sketch is a structure reference, not a style reference, or it'll render your pencil lines.
How to edit an image with a prompt
Drop your image in
Every image to image edit starts with a file, not a sentence. The preflight check runs in your browser the moment you drop one — format, size, dimensions, ratio, and whether the file already carries transparency. Nothing uploads, nothing gets charged.
1Tick what must not change
Identity, camera angle, text, lighting, product geometry. This takes four seconds and it is the difference between an edit and a re-imagining.
2Describe the one change
One thing at a time. That is OpenAI's own guidance and it holds up: two simultaneous requests is where drift starts.
3Generate, then check the receipt
About nine seconds. Not right? Go back to the original and adjust — don't stack another edit on top of a result you don't like.
4
Standard at 1K usually comes back in about nine seconds. Editing takes a little longer than generating. If it fails, times out, or changes more than you asked for, you keep your credits.
What GPT Image 2.5 image to image keeps when you tick a box
"Existing image → preserve it → make one requested change → preserve everything else."
That is not our tagline. It is a paying ChatGPT user writing out what he wanted, in the post where he explained why he had asked for his money back. He also described the failure precisely: "I can provide an exact screenshot and ask ChatGPT to recreate that environment at a higher visual quality while preserving the camera angle, geometry, proportions, furniture placement… Instead, it frequently decides to 'interpret' the image."
OpenAI's own prompting guide already has the answer: write "change only X" and then list everything that must survive. It works. Almost nobody types it, because it is four sentences of boilerplate every single time. So we made it five checkboxes.
Expanded by default, on desktop and on a phoneTick the boxes. We write the sentences.
Person's identity →
keep the same face, hair and skin toneComposition & camera angle →keep the camera angle, framing and geometry unchangedText & typography →keep all existing text exactly as it isLighting →keep the existing light direction and intensityProduct geometry & labels →keep the product's shape, proportions and label artwork identicalPlus a free-textalso keep…for the red bottle and the gold foil label. The panel shows you the assembled instruction before you spend anything.
Always sentOne line goes in whether you tick anything or not
Do not add text, logos or watermarks.Unprompted captions, invented logos and fake watermarks are the most common thing nobody asks for and everybody gets. It costs nothing to forbid and it removes most of them.
Two clicks, and they are the whole differenceThe same photo, run twice
One original, one prompt — "change the background to a sunlit marble countertop." The only difference between the two runs is whether any boxes were ticked.Nothing ticked: background changed — and the label typeface changed, and the highlight on the bottle shoulder moved about 40px. Product geometry and lighting ticked: background changed, label and highlight line up pixel for pixel.Neither run is the model misbehaving. In the first one, nothing told it the label was load-bearing — and that, more than any model choice, is what decides whether image to image works for you.
Signed-in history stays until you delete itNothing you make gets overwritten
"One image is shown — that you have to copy and save fast. Then it is replaced irrecoverably." That is a real complaint about a real product, and it is why people keep a folder of screenshots.Every generation and every edit here is its own version. Open the version strip under the canvas, click any earlier one, branch from it. Nothing overwrites anything.This is the actual reason to make an account here. Not to unlock anything — so you stop being one misclick away from losing an hour.
- Three free re-runs a day on Pro
If an edit changes more than you asked, the retry is free
We compare the result against your original. When the changed area is bigger than what you asked for, we say so and offer a free re-run — three a day on Pro, no ticket, no email, no explaining yourself to support.An edit that ignores your preserve list is our problem, not your credit balance.
Frames on this band are OpenAI's own published GPT Image 2.5 examples, from the Images 2.5 announcement of September 8, 2026. They are cited, not ours — our own runs replace them.
GPT Image 2.5 image to image: one change, and everything else held
Five edits from OpenAI's GPT Image 2.5 prompting guide, each with the prompt that produced it. Look at what did not move — that is the whole argument of the checkboxes above, and it is the half a marketing screenshot never shows you.
prompt In this room photo, replace ONLY the white chairs with chairs made of wood. Preserve camera angle, room lighting, floor shadows, and surrounding objects. Keep all other aspects of the image unchanged.
A real room, with real contact shadows under the chair legs and real light falling across the floor.
The chairs are wood. The camera has not moved, the floor shadows still land where the light says they should, and nothing on the counter changed.
This is the prompt shape the preserve list writes for you: name the one thing to change, then name what must survive.
prompt Remove the flower from man's hand. Do not change anything else.
The hand is still doing what a hand holding nothing does. That is the part object removal usually gets wrong.
Eleven words of prompt. The clause that does the work is the second sentence, not the first.
prompt Extract the product from the input image and place it centred on a flat pure-white field. Crisp silhouette, no halos or fringing, no shadow on the floor. Preserve product geometry and label legibility exactly. Do not add scenery, a gradient, or a checkerboard.
The prompt forbids a checkerboard by name. It has to, because the model paints one when it is asked for transparency it cannot deliver.
The label stayed legible and the silhouette has no halo. Those two are the whole job — and on an even field a one-click key-out gets you the rest of the way, which is what the transparency section below is about.
prompt Edit the image to dress the woman using the provided clothing images. Do not change her face, facial features, skin tone, body shape, pose, or identity in any way. Replace only the clothing, fitting the garments naturally to her existing pose and body geometry.
Four images went in: her, a jacket, a tank top, a pair of boots. Each one had a stated job.
Read the prompt again — six things are named as forbidden before a single garment is described. That ratio is the technique.
prompt Place the dog from the second image into the setting of image 1, right next to the woman, use the same style of lighting, composition and background. Do not change anything else.
The dog came from a second photograph with different light. Matching it was stated in the prompt rather than hoped for.
Every reference needs a stated job — subject, setting, style. Unlabelled references get averaged instead of combined.
OpenAI's published examples, from the GPT Image 2.5 prompting guide · cited, not ours
Check your image before you spend a credit
Not every photo is a valid image to image input, and finding out the hard way costs a credit. Drop a file here and get five answers instantly. Nothing is uploaded — the check runs in your browser. Run it as often as you like; the checker never touches the model, so you are never charged for a file that was never going to work.
Five checksinstant, in your browser, no upload
- Format
JPG, PNG and WebP are accepted
HEIC isn't — and it's what your iPhone saves by default. Settings › Camera › Formats › Most Compatible, or export this one as JPEG.
- File size
25 MB ceiling
Most phone photos are nowhere near it; scans and PSD exports often are
- Dimensions
Read before you spend
Very small sources can't carry detail into a larger output
- Aspect ratio
Output tops out at 3:1
A panorama wider than that has to be cropped somewhere — you'd rather choose where
- Transparency
Whether the file already carries an alpha channel
Most people have no idea, and it changes what happens when you edit
Free, instant, unlimited, and it costs us nothing to give you — which is the only kind of thing that should ever be described as unlimited.
A drawn checkerboard is not transparency
You ask an image to image tool for a transparent background. You get an opaque file with grey-and-white tiles painted onto it — a picture of transparency instead of transparency.
Seeing a faint grid woven into your image quality instead — in clouds, foliage or behind text? Different problem: that is the model's checkerboard artifact, and it is covered on the GPT Image 2 comparison.
Transparent + JPEG fails silentlyHow to ask for it properly
Turn on transparent background and set the output format to PNG or WebP. JPEG has no alpha channel at all, so that combination fails silently — no error, just an opaque file. We block the pairing in the editor; if you're calling the API yourself, this one will cost you an afternoon.One more thing that helps, from a user who worked it out the hard way: ask for an alpha channel by name, not just a transparent background. "I specifically ask it for a thing with an alpha channel and I haven't had it fail on me yet."
Native on the model · not exposed on this routeWhere the switch is, and which route it takes
Transparent backgrounds are native on both 2.5 models, and the panel asks for one on the Background chip — a Starter unlock. It sends the request by the direct API rather than the gateway we use by default, whose endpoint record for these two models lists
autoandopaqueonly and refuses the third; the chip decides that for you. On the free tier the workaround is still a good one: ask for a flat single-colour backdrop with no shadow and key it out — one step, and clean on a hard-edged object.
Three separate reports, one launch weekYou are not imagining this
"Half the time I ask for something on a transparent background it generates an opaque checkerboard background instead.""when you ask it for a transparent background and it just gives you white and grey tiles as the background""it's a bug. I had this issue, I point it out and then it makes clean .png — just don't wait until it realises the issue, you have to point it out."
Tell it what each reference is for
Feed an image to image model three references without saying what they're for and it averages them. You get something that looks a bit like all three and right like none of them. This is the most common multi-reference failure, and OpenAI's prompting guide is explicit about the fix: number your inputs and state each one's job, then say how they combine.
Up to five references per editUpload two to five, then pick a role for each
Subject · Style · Outfit · Background · Prop. A dropdown appears under each image, and we assemble the sentence:
Image 1 is the subject. Image 2 is the style reference — apply its palette and brushwork only. Image 3 is the background. Place the subject from image 1 into image 3, keeping its lighting.- One dropdown per image
The same three images, with and without roles
No roles assigned: the subject's face partly took on the style reference's features, and the background bled into the foreground.Roles assigned: the subject stayed the subject; only palette and brushwork transferred.Same three files, same model, same run settings. The only variable is whether each image was given a job.
- Priced per reference, on the button before you click
Worth knowing before you upload five
Each reference image is billed as image input, separately from the output. One is negligible. Five is not, and no other image to image site we checked breaks this out — which is fine right up until you are the one building a budget.
A GPT Image 2.5 receipt under every image to image edit
ChatGPT doesn't tell you which model edited your image. We do.Every figure below comes straight out of OpenAI's response — model ID, quality tier, pixel size, output format and elapsed time. Nothing gets quietly downgraded to a cheaper tier when we're busy.
Five fieldsgpt-image-2.5-sunburst · high · 1024×1024 · PNG · 9.1s
- Why this is here
People who have been burned by an image to image tool tend to suspect the next one of serving a cheaper model quietly. Fair suspicion. The receipt is how you check instead of trusting us.
- What happens to your photo
Used for your edits and nothing else — never for training, never resold, never published anywhere. Anonymous uploads are deleted within 24 hours. Signed-in files stay until you delete them, and Pro makes generations private by default.
- Who is serving it
OpenAI's official API. gptimage25.top is an independent interface — we are not OpenAI.
GPT Image 2.5 image to image requirements, and what it still gets wrong
Every limit and every known failure, in one table. Other image to image pages list what their tool can do; this one lists the file sizes and formats that get refused, the four things this model still gets wrong, and what each one looks like when it happens — so you find out here rather than after a credit is gone.
UploadJPG · PNG · WebP · ≤25 MB · up to 5 references · output up to 3:1
- Quality
low·medium·high·xhigh·maxDraft / Standard / HD / Ultra / Max in the editor
- Size
7 presets plus validated custom, up to 3840×2160
Above 2560×1440 OpenAI marks output experimental
- Background
auto·opaque·transparentTransparent needs PNG or WebP
- Output format
png·jpeg·webpJPEG returns faster
- Count
1 – 8
4 and 8 on paid plans
- Partial images
0, hard-coded
Each preview frame costs 100 extra output tokens and we render no progress animation
Four traps that cost people real money
None of these produces an error message. All four produce a bill.
Transparent plus JPEG fails silently
No error, no warning
Just an opaque file. JPEG has no alpha channel to write into.
Stacking edits on results
Compounds
Every pass re-encodes the last one and the damage compounds. If you're using another tool, go back to your original between attempts. Here it is automatic.
Not declaring what to preserve
The most common one
The model will interpret anything you didn't protect. One user's description of the result is hard to improve on: *"it generates a new picture with slightly changed camera settings."*
Reference images aren't free
Billed as image input
Each one is billed on top of the output. Five references on a Max-tier edit is a genuinely different price from one reference on a Standard edit.
What this still doesn't do well
From OpenAI's own limitations list: precise text placement can miss; keeping a character consistent across separate generations can drift; placing elements exactly inside a structured layout remains hard.From the community, within a day of launch: the fine checkerboard artifact was not fixed. It shows up in clouds, foliage, hair and behind dense text — and one user pinned down the part that matters here: "The more it's edited the more obvious it becomes."Hands. Still frequently short a finger. Check them.Moderation got stricter. The most-upvoted observation in the launch discussion was that the filter had tightened. If you sell swimwear, lingerie or fitness apparel you've probably had an ordinary product shot refused. Pro runs on OpenAI's more permissive moderation setting, which means fewer false refusals — not an unfiltered model. Anyone offering you the second thing is lying.
Checked against OpenAI's image prompting guide on September 9, 2026.
What one GPT Image 2.5 image to image edit costs
A Standard 1K image to image edit is 6 credits, plus the edit surcharge. Draft is 3, HD is 24, Ultra is 42, Max is 96. Doubling the resolution doubles the credits.And here is the part nobody tells you: your uploaded photo isn't free to look at. It is billed as image input, on top of the image the model produces. One reference is a rounding error. Five references on a high tier is not.
Per edit1024×1024, before the surcharge
- Draft
3 credits
- Standard — the default
6 credits
- HD
24 credits
- Ultra
42 credits
- Max
96 credits
You are not charged when
A run fails, times out, gets blocked by moderation, or changes more than you asked for. Credits come off when an image comes back, and refunds are automatic.
What people actually bring here
Five jobs people actually bring to an image to image tool, and the settings each one needs. The gap between a cutout you can drop into a listing and one you have to redo is usually a format and a tick box rather than better wording — so each row names the mode, the output format and the Keep-Intact boxes that job wants.
Product shots on a clean field
One product, one even backdrop, keyed out in a step. On Starter and above, set Background to
Transparentwith PNG and skip the keying — and put "no checkerboard" in the prompt, because that is the failure the model reaches for when it cannot deliver the real thing.Product photos and background swaps
Same bottle, forty scenes. Tick product geometry and lighting, describe the new background and its light direction, and the label survives.
Sketch → finished image
Your rough drawing is the structure, not the style. Say so, and it works — including the words you scribbled on it.
Remove an object or a person
Take the bystander out, keep the light and the shadow line honest. Tick lighting and composition, and name what is leaving rather than what should remain.
Starting from words instead?
No photo yet? Then this is the wrong tool and that's the right one.
Every one of these runs on the same editor. The mode and the preserve list are the whole difference.
Frequently asked questions
How is GPT Image 2.5 image to image different from text to image?
Is GPT Image 2.5 image to image free here? Do I need an account?
Why does the AI change things I didn't ask about?
Does quality degrade if I keep editing the same image?
Can it keep the same face across edits?
I asked for a transparent background and got grey squares.
Transparent with PNG or WebP, which asks the API for alpha rather than leaving the model to infer it from your sentence. And forbid the pattern by name in the prompt — "no solid backdrop, checkerboard, scenery, or shadow" — because the model has seen the grey-and-white grid captioned transparent in an enormous number of design screenshots.Then check the alpha channel rather than the preview. Against a white page, an opaque white background and a real one are indistinguishable.What image formats and sizes can I upload?
How many reference images can I use, and do they cost more?
Will you keep my uploaded photo?
Can I use the edited images commercially?
My product photo got blocked by moderation.
What still doesn't work well?
Bring the image you're afraid to touch
The one you've already spent an hour on. Tick what's already right, describe the one thing that isn't, and see what image to image gives you back — it's re-rendered from your original either way, so there's nothing to lose but thirty seconds.And if it comes back wrong, your original is still sitting there. It always is.
- First edit free
- Free preflight check
- Nothing overwritten
- Model-ID receipt
- Failed runs cost nothing
SIGN UP FREE · 1 IMAGE · 1KSIGN UP FREE · 1 IMAGE ON GPT IMAGE 2.5 · 1K · STANDARD · NO CARD
gptimage25.top is an independent product with no affiliation with OpenAI. This is not an official OpenAI website. GPT Image, ChatGPT and OpenAI are trademarks of OpenAI. Last updated September 9, 2026.
Written and maintained by Andy SwiftPublished Last updated Tool version 2026.09.1