Image to Image AI
Start from the picture you already have. Describe what should change and leave the rest alone — a background swap, a new style, a fixed detail, a wider frame.
- Edit by describing the change
- Keep the subject you photographed
- Restyle without redrawing
- Straight into upscaling or video
How image to image editing works
The input image is the anchor. Everything you write is read as an instruction about it, which is why short, concrete instructions beat long descriptions of the finished picture.
- 01
Upload the picture you are starting from
A photo, a render, a screenshot, or something generated earlier. Resolution matters less than framing — whatever is outside the frame cannot be edited back in unless you are extending it deliberately.
- 02
Say what changes
'Replace the background with a plain studio grey', 'make it look like an overcast morning', 'remove the cable on the left'. Instructions about the difference are followed far more reliably than descriptions of the whole result.
- 03
Choose how far the model may go
Editing strength decides how much of the original survives. Low keeps the photograph and adjusts it; high effectively re-imagines the picture using yours as a guide. Most disappointing edits are one setting away from being right.
The edits people actually need
Background replacement
Put a product or a person somewhere else without a cut-out, a mask, or a studio booking. The subject's edges and lighting are handled as part of the edit.
Restyling
Take a photograph into illustration, a render into film stock, a daylight shot into evening — while the composition and the subject stay recognisably the same.
Cleanup
Remove the object, reflection, or passer-by that ruins an otherwise usable frame, without cloning by hand.
Extending the frame
Widen a shot that was composed too tight — the usual fix when a vertical photo has to become a wide banner.
Consistent sets
Run the same instruction across a batch so a catalog, a cast, or a product line comes out with one look rather than twelve near-misses.
Finishing in place
Upscale, relight, or animate the edited result without exporting and re-uploading it into another tool.
Where editing beats generating
Product photography
One real shot of the product, then as many backgrounds and seasons as the campaign needs — with the product still the product.
Headshots
A consistent background and lighting treatment across a team whose photos were taken in five different rooms.
Real estate and interiors
Overcast to golden hour, clutter removed, a room presented at its best without a reshoot.
Reformatting
Extending a hero image to the ratios a campaign needs instead of cropping the subject out of its own picture.
Style unification
Assets collected from several sources brought to one look before they appear in the same deck or feed.
Rescuing a frame
The shot with the right expression and the wrong detail, fixed rather than discarded.
Why editing keeps what generating loses
Text-to-image invents everything, which is exactly what you want when nothing exists yet and exactly what you do not want when the subject is real. A product has a shape customers recognise; a person has a face they know. Asking a model to imagine either from a description gets you something plausible and wrong. Starting from the photograph keeps the part that has to be true and lets the model work on the part that does not.
The single most useful control is editing strength, and it is the one most people leave alone. At low strength the model behaves like a retoucher: the photograph survives and the instruction is applied to it. At high strength it behaves like an illustrator who glanced at your picture. Neither is wrong, but they answer different requests — and when an edit comes back either 'nothing changed' or 'that is not my product any more', the setting is almost always the cause.
Instructions phrased as differences also outperform instructions phrased as descriptions. 'Replace the background with a plain grey studio wall' is a change; 'a professional product photo on a plain grey studio wall' is a fresh brief that invites the model to reconsider the whole frame, including the parts you were happy with. The more of the original you want kept, the more your prompt should sound like an edit note.
Finally, edits compose. Background, then colour treatment, then a crop extension as three passes is usually better than one prompt asking for all three, because each pass starts from a result you have already approved. It costs a few more credits and removes most of the guesswork.
Image to image FAQ
- Will the subject still look like itself?
- At low and moderate editing strength, yes — that is the point of starting from your image. Push the strength high and the model is free to reinterpret the subject, so keep it low when identity or product shape matters.
- Can I edit only part of the picture?
- Describe the region in the instruction — 'the background', 'the shirt', 'the object on the left'. Models follow spatial instructions reasonably well, and a lower editing strength further limits how far the change spreads.
- What file types can I upload?
- Standard web image formats. Very large files are resized before generation, so exporting a reasonable size first is faster than uploading a full-resolution original.
- Can I use it on photos of people?
- Yes, for people who have agreed to it. Editing someone's likeness without their agreement is against the terms of service, and there is a takedown route on the support page for anyone who finds themselves in a published result.
- Is it free?
- Yes, within your free credits. Edits are priced like any other image run — by model and resolution — and paid plans add more credits and watermark-free exports.
Where to go next
Upload one picture and change one thing
Free credits on sign-up, every image model in the catalog, and a library that keeps each version you tried.