Start from one or more reference images, describe the change, and generate consistent image variations in the same workspace.
Upload references and direct the change in plain language: restyle materials, swap backgrounds, hold characters steady across a whole series. The workflows users praise most, in one panel.
Change the fabric, the lighting, the season, the medium — while the framing and subject stay put. Reference-based restyling means the photo you liked becomes the photo you needed, instead of a lottery ticket on a new generation.
Move a product from white seamless to a sunset beach and the lighting, shadows, and reflections follow automatically — the tedious relight work users say takes the longest in manual editing. E-commerce sets that used to need a studio day take minutes.
Nano Banana's celebrated "character DNA" keeps the same face, outfit, and style across poses and scenes. Build model sheets, story panels, and brand mascots that read as one character — the capability Reddit calls genuinely rare in AI tools.
Day to night, photo to anime, modern to cyberpunk — transform the entire mood of an image while its structure survives. Pair multiple references to fuse subjects, products, and scenes into one coherent composite.
Output from the same engines that drive image to image on this page — Nano Banana, Seedream, GPT Image 2, Qwen.
The engines that hold a subject steady while everything around it changes.
| Input | One or more reference images plus a text instruction. JPG and PNG accepted. |
|---|---|
| Reference limit | Up to 5 images on Nano Banana Pro; one or two on most other engines. The panel shows the limit for the model you selected. |
| Best-suited models | Nano Banana 2 and Nano Banana Pro for fusion and identity, Qwen Image Edit Plus for precise local changes, GPT Image 2 for instruction-heavy edits. |
| Resolution | Up to 2K on most models; 4K on the Ultimate plan. |
| Cost | Per image, by model: 20 to 40 credits for most reference workflows. Shown before you generate. |
| Output | A new image — your source file is never overwritten. Both stay in your history. |
Upload, describe the change, generate — three steps.
Attach one or more reference images. The image to image generator reads composition, subject, and style from what you give it.
Say what to alter and what to preserve. Naming what stays fixed is what separates a clean image to image result from a full regeneration.
Compare against the original, adjust the instruction, and rerun. Every version keeps its reference and prompt in history.
Image to image prompts work differently from text to image: the model already has the picture. Your job is to say what changes and what must not.
Name the things that must survive the edit. This is the part most people skip, and it is the part that decides whether it still looks like your photo.
keep the same face, pose, and clothing
The single transformation you want. One per generation.
replace the background with a sunlit olive grove
Style, light, and finish of the new version.
warm late-afternoon light, soft film grain
Only if the crop or ratio should change.
widen to a 16:9 crop
make it better
keep the subject and pose exactly; relight the scene with soft window light from the left, remove the cluttered shelf behind her, keep the warm colour of her sweater
“Better” is not a direction — the model guesses, and the guess rarely matches yours. Naming what to keep, what to remove, and how to light it gives it three unambiguous instructions.
turn this photo into anime
convert to anime illustration, cel shading with clean linework; keep the hairstyle, glasses, and the red jacket; same three-quarter pose and framing
Style conversions drift on identity unless you anchor them. Listing the two or three features you would use to recognise the person is usually enough to hold it.
Reference-driven generation is one step. The rest of the pipeline is in the same account.
Multi-image fusion and how far character consistency actually holds.
Precise local edits when you want one region changed and nothing else.
Restyling a photo without losing the subject underneath.
Swapping a scene while keeping the person lit correctly.
Five-image fusion, style consistency, and face preservation.
Where image to image sits in the wider generation workflow.
Image to image generation starts free with 80 daily credits. Paid plans add premium engines, higher resolution, and faster queues.
Perfect for getting started
Billed annually · $240/year
Billed annually · $380/year
Generation guided by reference images: you upload one or more images, describe the change, and the model produces a new image that keeps what you specified — subject, composition, style — while changing the rest.
It depends on the model — Nano Banana models fuse multiple references, and GPT Image 2 accepts up to 16. The panel shows the limit for the selected model.
Nano Banana 2 and Nano Banana Pro are the community picks for character consistency and face preservation across scenes and edits.
Yes — for surgical local edits like removing an object or changing one color, use Qwen Image Edit Plus, which preserves everything you did not ask to change.
Yes — background swaps, scene placement, and restyling with automatic lighting adjustment are core e-commerce workflows here.
Per-image credit costs depend on the model — from 20 credits for Nano Banana upward — and are always shown before you generate. Free accounts get 80 daily credits.
Uploads are used to fulfil your generation and stored securely. Everything you create is private by default — no other user can see your generations.
Image to image regenerates a new picture guided by your reference — composition and subject carry over, but the pixels are new. An AI image editor changes a specific region and leaves the rest of the original file untouched. Use image to image for restyling, the editor for retouching.
It depends on the model. Nano Banana Pro accepts up to five for fusion work; most other engines take one or two. The panel shows the limit for whichever image to image model you have selected.
Largely, on the models built for it — but identity preservation is a spectrum, not a switch. Asking for an "identical" face tends to produce stiff results; describing the person and letting expression vary usually looks more like them.
One-click tools for everything around generation — polish, restyle, and share.
