Midjourney has revealed its Version 4, or --v 4, in prompts. Long story short: it's BIG.

I've run a series of lab tests trying to reveal a tiny fraction of how it sees the world. Here are the results of my research.

MRI scan of robot samurai --v 4

Japanese fox god of winter death and rebirth --v 4

intricate raven god —v 4

EMOTIONAL RESPONSE

The --v 4 blows your mind. It exceeded my expectations almost every single time. It knows what's "beautiful" for humans and sends it back to us tenfold. It has excellent attention to detail and knows composition perfectly.

anime girl wearing intricate bird mask

dimensional astrolabe

Slavic goddess of winter and floral decay

digital glitch Thom Yorke

infographics of all know multiverses. 1980s Japanese graphic design

anime girl wearing intricate bird mask

colossal ancient god of deers. Photograph by Josef Sudek

And—it seems to become a feature for Midjourney—even when the results are off or glitch-y, they are off beautifully. I encourage you to experiment!

layout of dissected bird on lab table. Scientific photograph with notes and marks

colossal Sphynx cat looming over rainy London by Josef Sudek

huge cat on lab table. Top view. 1900 scientific photograph with notes and marks

IMAGE PROMPTS

As a portrait photographer, I am overly fascinated by the limits of Image Prompts. I was very impressed by it in V3, but Version 4 takes Image Prompts to the moon. Thus they will be my focus today.

Meet Francis, a friend whom I photographed in Paris some years ago.

And that's what happened when I first "fed" that portrait to --v 3.

large format film monochrome photograph by Richard Avedon with shallow depth of field --iw 0.9 --ar 4:5

no text prompt --stylize 625

wearing cyberpunk mask visor --ar 4:5 --iw 0.5 --stylize 1625

no text prompt --stylize 5000

Sidenote: Francis taught me another important lesson: the more pronounced a person's features are, the better Midjourney "captures" and "reinterprets" them. Bottom line: choose subjects with strong, pronounced characters, and images with clearly visible and well lit features. Also, Francis is ideal for our test.

Today, I asked --v 4 to reimagine the portrait of Francis with a simple addition as someone. And here are the results:

young dramatic actor

knight

fashion designer by Paolo Roversi

cyberpunk preacher

vigilante superhero from 1920s

beetles collection

Summer Of Love child

Wong Kar Wai movie character

guerilla Rock Star by Vivienne Westwood. Photograph by Anton Corbijn

political star by Annie Leibovitz

Japanese folk theatre mask

by Anton Corbijn

fashion icon by Tim Walker

Chernobog

first man to evolve to artificial intellegence. Digital glitch, computer art, VHS distortions

digital glitch god

Lord of Thieves by Rembrandt

Venetian merchant by Masaccio

cruel world ruler

cyberpunk mugshot

ASCII art

eternal interdimensional being beyond Time and Space

opium smoker by René Magritte

Japanese poster graphics

As you can see, V4 is crazy good here. It truly catches the character and carefully transitions the style. And it seems to successfully put together even the craziest combinations! (see below)

Even though V4 knows most artists' styles from my experiments, some confuse AI. In some cases, it seems to only apply a part of the style: like light, color, and composition, but not the actual brushstrokes; or use the artist's style as a backdrop or even some scene elements.

by Franz Kline

by Edward Hopper

Midlibrary by Zdzislaw Beksinski

I then tried some of my other portrait work, including more detailed, wide-frame scenes.

[ORIGINAL PHOTO]

by Tim Walker --v 4

by Tim Walker --v 4

by Tim Walker --v 4

[ORIGINAL PHOTO]

by Tim Walker --v 4

by Tim Walker --v 4

[ORIGINAL PHOTO]

as 19th century photograph by George Melies --v 4

by Georges Méliès --v 4

by Georges Méliès --v 4

[ORIGINAL PHOTO]

in mirror dimesion --v 4

in mirror dimesion --v 4

[ORIGINAL PHOTO]

by Gregory Crewdson --v 4

[ORIGINAL PHOTO]

as collage from pieces of pomegranates by Sergei Parajanov --v 4

[ORIGINAL PHOTO]

--v 3

mediating cybermonk in glitching virtual temple. Ghost In The Shell style --v 4

I see a lot of people use Midjourney for previsualization in their photographic projects. I think its a very cool idea! In my days *coughs* we used Photoshop to put digital collages together, and some more intricate ones could consume hours to prepare. Let's feed one of those old-school collages to Midjourney and compare the result with text-only prompt generation.

[ORIGINAL PHOTOSHOP MOCKUP]

[IMAGE PROMPT]
futuristic samurai fights mecha Godzilla in wasteland --v 4

[TEXT PROMPT]
wandering futuristic samurai with colossal plasma sword in post-apocalyptic wasteland fighting gigantic mecha Godzilla --v 4

[TEXT PROMPT]
wandering futuristic samurai with colossal plasma sword in post-apocalyptic wasteland fighting gigantic mecha Godzilla --v 4

MULTI-IMAGE PROMPTS

And that wasn't enough, here comes the ground-breaking part. You can "feed" Midjourney two images and see how it magically blends them together. Yes, you could do it in --v 3, too. But V4 takes it to a whole new level.

To see how it works, I took a few iconic movie scenes and asked Midjourney to mix them with the portrait of Francis.

The Shining

Blade Runner

Get Out

The Revenant

2046

Predestination

Of course, I tried blending Francis' portrait with some other images from both photographs and even Midjourney generations.

[ORIGINAL PHOTO]

[ORIGINAL MIDJOURNEY GENERATION]

Bonus track: try sending Midjourney its own generations, changing artist's style on each step.

[ORIGINAL PHOTO]

by Tim Walker --v 4

by Hayao Miyazaki --v 4

by Tim Walker --v 4

by René Magritte --v 4

by Ib Kamara --v 4

Mind. Blown. Thank you, Midjourney!

‍-- Yours, Andrei