Blog
Recreate the style of a photo with AI
Attach the photograph as a style reference and the model takes its palette, light and surface onto your own subject. One kitchen photo's look moved onto a pair of boots, plus when to decode it into words instead.
To recreate a photograph's style, you do not describe it — you attach it. Upload the photo, badge it as a style reference, and write your own subject plainly. The model is then told in words to take that picture's palette, mood and rendering and ignore its subject. 8 credits at 1K. If you want the style as words rather than a file, decode the photo for 3 credits instead.
One photograph, a different subject
The source was a warm kitchen shot: oak counter, morning window light from the right, everything falling soft behind. The prompt for the new picture mentioned none of that — it said only "a pair of worn leather work boots standing on a floor, laces loose, photographic".
What crossed over: the warm oak, the low directional window light with its bright pool on the floor, the amber-and-brown palette, the shallow focus falling into a soft interior, even the sense of late morning. None of it was written down. What did not cross: the mug, obviously — because the Style role tells the model to ignore the source's subject.
The sentence that does it
Reference 1 is a style reference only — take its palette, mood and rendering quality, and ignore its subject and composition.
That is what badging an upload "Style" writes into your prompt. It matters because the model has no typed reference slots: every attached image lands in a flat array and the model guesses each one's job from the words. An unlabelled reference leaks its subject into your picture; a labelled one does not.
Two routes, and when to take each
| Attach the photo | Decode the photo | |
|---|---|---|
| Cost | Nothing extra — 8 credits at 1K, same as any generation | 3 credits |
| What you get | The look, carried directly | The look as words: palette hexes, a four-layer style DNA, seven style tokens, a prompt |
| Carries material | Yes — wet stone, worn leather, coarse linen | Names it, which is weaker |
| Editable | No. It is one file, take it or leave it | Yes. Keep half, change the rest |
| Portable to other tools | Only where references are accepted | Anywhere that takes text |
Use both together when the style really matters: decode the photo, paste its prompt, and attach the photo as well. The words give you control and the file gives you the surface.
The limit worth stating
A style reference carries palette, light and surface reliably. It does not carry another tool's hand. When we crossed a Midjourney engraving over this way, the palette landed exactly and the line quality did not — Midjourney's lines thicken where a form turns away from the light, and the rebuild laid down regular contours instead. Expect a cousin, not a copy, and judge the cousin on its own terms.
One honest note about this page: the source above is itself a generated picture, made earlier in this series, because everything shown on aikizi is work its owner published. The technique is identical with a photograph you took yourself.
FAQ
How do I recreate the style of a photo with AI?
Upload the photo as a reference and badge it Style, then write your own subject plainly. That role tells the model to take the palette, mood and rendering and ignore the source's subject and composition. 8 credits at 1K on Nano Banana Pro.
Should I describe the style in words instead?
Words are weaker for surfaces and light. You can write "warm morning light on oak" and get a version of it; you cannot write that particular warmth falling off at that rate. Decode the photo for 3 credits if you want the style as text you can edit.
Why did my reference bring its subject along?
Because it was not labelled. With no typed reference slots, the model guesses each picture's job from your wording. Badging it Style is what explicitly rules the subject out.
Can I combine an uploaded reference with a written style?
Yes, and for important work it is the best of both: decode the photo, paste the prompt it writes, and attach the photo too. The words give control, the file gives the surface.
Will it match the original exactly?
Palette, light and material carry well; another tool's specific line or rendering hand does not. Expect a close cousin rather than a copy.
Take a photograph whose light you like, attach it in the Studio and badge it Style — one generation tells you how much of it travels. What a decode gives you instead is in the decode guide.