Image to Image: Change One Thing, Keep Everything Else

Upload up to four images and describe the edit in a sentence — new background, new style, new label copy, or the subject cut out onto transparency. Only what you named changes; the rest of the frame stays exactly where it was. Pick a model, and if an edit will not hold, rerun the same instruction on another. Up to 4K.

Image to Image

Fast default — better than GPT Image 2 at about half the latency. 1K–4K, transparent backgrounds.

Add images — JPG, PNG, or WEBP

Up to 4 images, 1.25 MB each. The first image sets the composition; the rest act as references.

0/2000

0 credits available

Sample image generated in the image workspaceSample — Image to Image

Your image appears here

Hit Generate and the result shows up here as soon as it is ready, with a download link.

Change One Thing and Leave the Rest of the Frame Alone

Most editing tools make you choose between redrawing the whole image and masking by hand. Image to image is the third option: upload what you have, say what should change, and that is what changes — the background swaps, the subject gets restyled, the label gets new copy, and the parts you did not mention stay where they were. You are not tied to one model either. Pick one from the selector, and if an edit does not hold, run the same instruction on another. Up to four reference images, output up to 4K, transparent PNG on demand.

Demo — Before and After
Source image uploaded before an image to image edit
Input
Result returned by the image to image edit
Output

How to Edit an Image With a Prompt

1

Upload one to four images

JPG, PNG, or WEBP, 1.25 MB each and 4 MB in total. The first image sets the composition; anything after it acts as a reference — a style to borrow, a product to insert, a palette to match.

2

Describe only what should change

Name the edit, not the whole picture: replace the background with a grey studio sweep, restyle the portrait as a pencil drawing, change the label to read something else. What you leave unmentioned is what stays untouched.

3

Set the output and generate

Leave the aspect ratio on Auto to match the source, or pick another to reframe. Choose your resolution and whether the background should be opaque or transparent, then generate — the result arrives with a download link.

Six Edits Worth Trying First

EditWhat to writeAlso say
Background swapReplace the background with a seamless light grey studio sweepkeep the subject and its shadow unchanged
Subject extractionExtract the product onto a transparent backgroundno backdrop, no scenery, no shadow
RestyleRender the same scene as a flat vector illustrationsame composition and colour palette
Text changeChange the label copy to read "COLD BREW 500ML"keep the typeface, size, and placement
Canvas extendExtend the scene to 16:9continue the wall and the floor naturally on both sides
Reference blendPlace the product from image 2 onto the table in image 1match the light direction and the perspective

Getting the Edit to Hold

One change per generation

Stacking four edits into one prompt is how details drift. Make the background swap, download, re-upload, then change the text. Each pass is cheap and each result is predictable.

Say what must not change

Adding 'keep the subject, framing, and lighting exactly as they are' measurably tightens the result. Models honour explicit constraints better than they infer them.

Order your reference images

The first upload drives the composition. Put the image you are editing first and the references after it, then refer to them by position — 'the jacket from image 2' — so there is no ambiguity.

Auto ratio keeps the source framing

Leave Aspect ratio on Auto when the edit should not reframe anything. Pick an explicit ratio only when you actually want the canvas extended or cropped.

Switch models when an edit will not hold

Models differ most on precise edits — text, a specific product, a multi-step refinement. If one keeps drifting, pick another from the selector and rerun the same instruction. Your uploads and your prompt stay put; the ratio and resolution reset to the new model's defaults, so reset them yourself if you need an exact comparison.

Nothing to Upload Yet? Start From Text.

Image to image needs a source. If you are starting from a blank page, generate the base first in the AI Image Generator, then bring the result back here to refine it. When the finished frame should move, send it to Image to Video — the composition, colours, and rendered text carry into the clip.

FAQ

What does image to image mean?+
You supply an existing image plus a written instruction, and the model returns an edited version of that image. Unlike text to image, the source picture is the starting point — composition, subject, and colours carry over unless you ask for them to change.
How many images can I upload at once?+
Up to four, at 1.25 MB each and 4 MB combined, in JPG, PNG, or WEBP. The first image sets the composition and the others act as references you can point at in the prompt.
Will it change parts of the image I did not mention?+
Generally no — preserving untouched regions is exactly what these models are tuned for. To be safe on precise work, state explicitly what should stay the same and make one change per generation rather than stacking several into one prompt.
Which models can I edit with?+
Every model currently available appears in the Model selector above the prompt box, each with a short note on what it is best at. The list grows over time, and your uploads and settings carry across all of them, so you can rerun an edit on a new model the day it is added.
Can I change the text inside an image?+
Yes. Quote the exact replacement wording and say to keep the typeface, size, and placement. On the models built for text rendering, labels, headlines, and price tags come back legible and spelled as written.
Can I remove the background from a photo?+
Yes. Set Background to Transparent and ask to extract the subject with no backdrop or shadow. You get a cut-out PNG without masking anything by hand.
What resolution does the edited image come back at?+
Whatever you select, independently of the source file's size — up to 4K. Leave the aspect ratio on Auto to keep the source framing, or pick another ratio to reframe or extend the canvas.
Does editing cost more than generating from text?+
No. Credits are charged per output image and depend mainly on the resolution you pick. The exact cost appears on the Generate button before you submit.
Are uploaded images checked before editing?+
Yes. Every uploaded image and every prompt passes through input moderation, and restricted sexual or NSFW content is blocked before generation starts.

Upload It and Describe the Change

Take the shot that is almost right — the one with the wrong background or the old price on the label — and fix it in a sentence. One upload, one instruction, and the file you needed.

View Pricing and Credits