Guides

Midjourney to aikizi: decode a picture, rebuild its look on Nano Banana Pro

A Midjourney V8 portrait decoded for 3 credits, its own prompt run back through Nano Banana Pro with the original as a style reference. What held, what drifted, and where Midjourney stands as of September 2026.

September 20, 20266 min readAIKIZI

Part 5 of 11 · ← previous · all parts · next → Character consistency

Decode a Midjourney picture for 3 credits and you get its prompt, its palette with hex codes and its style DNA. Feed that prompt back into Nano Banana Pro with the Midjourney image attached as a style reference, and you get something close — but not the same. We did it, and the interesting part is exactly what survived the crossing and what did not.

Where Midjourney is, as of September 2026

Worth getting right, because a lot of what is written about it is a year out of date.

V8.2 is the default model, since 24 July 2026. It is an image model — an update aimed at aesthetics, image quality and Personalization. It is not, despite how it is often described, an edit model.
The Edit Model for V8 is a separate thing, in public testing since 27 August 2026, with a quality update on 29 August. That is the piece that edits an existing image from written instructions, and that takes up to four reference images in one prompt, replacing both Omni Reference and Character Reference.
So --cref is gone from V8. Character Reference is in Midjourney's legacy documentation now: compatible with version 6, replaced by Omni Reference in V7 and by the Edit Model in V8. --oref is likewise unsupported on V8.1 and V8.2.
--sref and --sw still work on V8.1 and V8.2. --sw runs 0 to 1000 and defaults to 100. Inpainting and outpainting in the Editor now run through the Edit Model, and Personalization and moodboards are supported.
alpha.midjourney.com is Midjourney's own in-progress web interface, shipped separately from the main site. In their words on 16 September 2026, the goal is to "make the site easier and more intuitive without taking away any of the control you have today", some things will still be broken, and they still recommend the main site over Alpha on mobile.

The crossing, step by step

Upload the Midjourney image to the Studio and decode it. 3 credits, about fifteen seconds, seven tabs of reading.
Open STYLE and VISUAL. STYLE carries the seven two-word tokens and the layered style DNA; VISUAL carries the palette with hex codes.
Copy the generation prompt the decode wrote. It is already in the shape Nano Banana Pro reads well.
Start a generation, paste the prompt, and attach the Midjourney image itself as a reference with the Style role.
Compare the two side by side and note what drifted. That list is the thing you actually learn.

What we ran

The source is a published Midjourney portrait from the gallery — an engraved android head, white line on black, one spot of blue in the irises. Its decode names the style "Neo-Futurist Cybernetic Engraving", the palette "high-contrast monochromatic with cerulean highlights", and the base layer of its style DNA "intricate burin engraving".

The Midjourney original: an engraved android portrait, white scratched lines on black following the form of the face, a silver bob, white sparkle across the cheeks, blue only in the irises, downcast expression — The source, made in Midjourney. Its decode on aikizi, 3 credits.

The aikizi regeneration: a symmetrical frontal cyborg portrait, regular topographic contour rings across the face, neon blue circuit traces, bright blue irises, a rigid circuit-panel collar — The decode's own prompt run on Nano Banana Pro, 1K, with the Midjourney image attached as a style reference. 8 credits. Its decode.

What held and what drifted

	Midjourney original	Nano Banana Pro rebuild	Verdict
Palette	Black, white, one blue	Black, white, one blue	Held exactly
Medium	Engraved line on black	Engraved line on black	Held
Background	Pure void	Pure void	Held
Where the line goes	Lines follow the form; weight varies with the light	Regular contour rings laid over the face like a map	Drifted — the biggest loss
The glow	White sparkle; blue only in the irises	Neon blue circuit traces across the skin	Drifted
The hair	Soft silver bob, asymmetric, flowing	Slick dark hair, tight to the head	Drifted
Expression	Downcast, melancholy	Frontal, symmetrical, direct	Drifted

Why it drifted, which is the useful part

The decode was not wrong. "Fine-line woodcut engraving", "bioluminescent circuitry", "monochromatic palette with selective saturation" are all true of the original. But those words name a taxonomy, and what makes the Midjourney picture good is a hand: line weight that thickens where the form turns away from the light. No vocabulary carries that. Nano Banana Pro read "engraving" plus "circuitry" and produced a technically correct answer — contour lines, neon traces — which is what the sentence says and not what the picture looked like.

The lesson for anyone crossing over: the decode gives you the palette, the subject, the medium and the mood, and those transfer cleanly. It does not give you the other tool's hand. Expect a cousin, not a copy, and decide whether the cousin is good on its own terms.

Two things to do differently next time

Cut the words that invite a literal reading. "Bioluminescent circuitry" bought us neon. Removing it and keeping "white specular sparkle" would have been closer.
Say where the line goes. Add "line weight follows the form, heavier in shadow, opening to white in the light" — describing the hand rather than the style name.

Neither is a Midjourney parameter, and this page will not give you one: style codes are paid work and never appear on a free page here.

FAQ

Can I turn a Midjourney image into a prompt?

Yes. Decoding an image on aikizi costs 3 credits and takes about fifteen seconds, and every decode ends in a generation prompt you can run anywhere, plus a palette with hex codes and a layered style DNA.

Does --cref still work in Midjourney V8.2?

No. Character Reference sits in Midjourney's legacy documentation: it worked on version 6, was replaced by Omni Reference in V7, and by the Edit Model in V8. --oref is also unsupported on V8.1 and V8.2. As of September 2026, --sref and --sw do still work.

What is Midjourney V8.2?

The default Midjourney model since 24 July 2026, an image-quality and Personalization update. The Edit Model that edits an existing image from written instructions, and accepts up to four reference images in one prompt, is a separate release that entered public testing on 27 August 2026.

Will Nano Banana Pro reproduce my Midjourney style exactly?

No, and our test shows where the line falls. Palette, medium, subject and background carried over intact; line character, the colour of the glow, the hair and the expression all drifted. You get a cousin, not a copy.

What is alpha.midjourney.com?

Midjourney's own in-progress web interface, shipped separately from the main site. As of 16 September 2026 they describe it as making the site easier without removing control, warn that some things will still be broken, and recommend the main site over Alpha on mobile.

If you have a Midjourney picture you like, decoding it is the cheapest useful thing you can do on this site: drop it into the Studio, or read how a decode works first. To keep one character across many pictures rather than one style, part 6 is the tool.