New feature for Midjourney gives prompts for any image

[15:06 Fri,7.April 2023 by Thomas Richter]

Midjourney, the currently the best image AI has just released a remarkable new feature, namely the ability to get text descriptions for any image and thus let you guess how the AI sees an image. Descriptions are given for both the image subject and the style of the image. But what makes this new "Describe" feature so interesting for AI artists?

And the results of the different prompts generated (from

HungryMinded):

On the one hand, this way theoretically any imaginable image can be "recreated" via Midjourney (see examples below), i.e. the description of a representation image generated by Midjourney also works as a prompt for regenerating a more or less similar image.

Prompts are no longer secrets

.
Some "AI artists" (aka prompt engineers) have so far freely shared the prompts of their images, others have made a secret of it or even tried to make a business out of it - as prompt artists, so to speak. They believe that their image prompts are a precious knowledge, distilled by long series of experiments, which makes it possible to elicit very distinctive images from a certain image ki - comparable to the experience of a visual artist.

This should be over now, as the new "Describe" function is also suitable for reverse engineering and often reveals the prompts with which an image - if it was generated by AI - was generated. However, this does not work 1:1 - there is some fuzziness, so that e.g. even an image generated via midjourney does not output the exact original prompt via Describe, but only similar ones - sometimes also the newly generated image is completely different.

Variation

Midjourney provides four slightly varying descriptions per image, which can be used by click to generate images based on them. You can see so well how the different parameters of the prompt variations affect the images in each case. Since the image is available in the form of a text prompt thanks to Describe, individual display parameters can also be changed in a targeted manner, for example to create very specific variations.

Variationen auf ein Bild - von

HungryMinded:

Zum anderen können jetzt per "Describe" - ähnlich wie bei

Stable Diffusions Reimagine / unCLIP einfach Variationen eines beliebigen Bildes erzeugt werden - sei es, um schönere Versionen zu finden oder auch das Copyright zu umgehen und geschützte Bilder einfach - genügend ähnlich, aber nicht zu exakt - nachzuahmen.

Beispiel von Ethan Mollick - "neue Sachlickeit", "daguerrian"?

Einblick in den Latent Space

Eine weitere Frucht der Erkenntnis ist der Einblick, den "Describe" in Innenleben von Midjourney, d.h. den multidimensionalen

Latent Space, offenbart, also was Midjourney in einem Bild sieht. Dabei sind manche Ausgaben oft überraschend - wer hätte das folgenden Bild, ein Filmstill aus dem populären Hollywood Alpen-Musicalfilm "Sound of Music" etwa mit "made of insects" und "grandparentcore" assoziiert? Das

Beispiel stammt von Ethan Mollick, das vollständige von Describe erzeugte Prompt lautet: "a woman&s open arms, in the style of bold chromaticity, made of insects, mountainous vistas, grandparentcore, joyful and optimistic, unearthly, movie still".

Sound of Music

Learn better prompting

.
Using the generated descriptions, you can also well improve your own prompting skills - you can see which parameters - according to Midjourney himself - define an image in all its dimensions. It is interesting that the Describe results often also mention artist names to describe a style of representation or an image mood. And one gets to know new parameters - concerning image style, object descriptions or artists - that one might not have had on the screen as relevant (or existing) before.

more infos at bei www.midjourney.com

deutsche Version dieser Seite: Neue Funktion für Bild-KI Midjourney: Bild rein - Prompts raus