Image to image: change the scene, keep the person
Upload a photo, describe where it should go, and the model redraws it — same face, new light, new place. Up to 14 references per request.
0/4 reference images · photos are resized on your device before upload
Try one of these
Image to image starts from something real. Instead of inventing a subject from a description, the model works from your photo and changes the parts you ask it to change. That shift from invention to transformation is what makes identity preservation possible at all.
This page covers how much a reference actually constrains, what to upload for the best result, and where the three fidelity tiers differ in practice.
What a reference image actually constrains
A reference is not a template. The model does not paste your photo into a new background; it extracts structure — the proportions of a face, the fall of hair, skin tone, sometimes the cut of a garment — and rebuilds the image around those constraints. What survives depends on the tier and on how much of the frame you change.
Change the background and the lighting, and identity usually holds. Change the hairstyle, the age and the camera angle at once, and even strong models start drifting, because too many of the cues that define a face have been replaced simultaneously. If a result goes wrong, change one variable per attempt.
Faces work best when the reference is frontal, evenly lit and reasonably sharp. A group photo where the subject occupies a tenth of the frame gives the model very little to work with; a clear head-and-shoulders shot gives it a lot.
- Front-facing and evenly lit beats artistic and moody
- One subject per reference — the model will otherwise blend them
- Send two or three angles of the same person when identity is critical
- Avoid heavy filters; the model will preserve the filter as if it were a feature
How the three tiers differ in fidelity
The Fast tier treats references as guidance. It reads the general look — hair colour, build, palette — and then composes freely. That is often exactly what you want for a mood board or a background plate, and it costs a single credit.
Standard and Identity use reference-aware conditioning that is much more literal about facial structure. Identity also renders at up to 4K, which matters if you plan to crop into the face or print. In testing, Identity held freckles and eye colour through a full outfit and location change; Fast produced a plausible sibling rather than the same person.
Preparing a photo without a desktop editor
You do not need to crop or resize anything before uploading. The uploader downscales each photo on your device to a 1024-pixel long edge and re-encodes it as JPEG, which keeps requests small and avoids sending full-resolution originals over the network.
That also means you should not upload a photo whose only usable detail is a tiny region — if the face is already small in the frame, downscaling will not recover it. Crop first in your phone’s photo app if you need to, then upload.
Fidelity by tier, based on live tests
| Tier | Credits | What it preserves | Best for |
|---|---|---|---|
| Fast | 1 | Palette, build, general look | Mood boards, backgrounds, quick direction checks |
| Standard | 2 | Facial structure, hair, skin detail at 2K | Social posts, thumbnails, client previews |
| Identity | 5 | Facial structure and fine detail at up to 4K | Character sheets, print, anything cropped into the face |
3 example prompts are preloaded above — tap one to try it.
Questions, answered
About fidelity, credits and output sizes on this tool.
Other ways to generate
Ready to make something that looks like the person you had in mind?
Start generating