[12:24 Sun,19.March 2023 by Thomas Richter]
Midjourney, one of the great image-generating AIs has taken another big step with the just released new (alpha) version 5 - unimpressed by the complaints of the artists: the produced images now have an even higher photorealistic image quality than before and also the otherwise often disturbing small image errors are even rarer than before. Even - drum roll - the annoying problems of all image AIs in displaying human hands, which sometimes had more, sometimes less fingers, is now solved in Midjourney V5.
human hands - now finally correctly rendered by the AI in Midjourney V5
This progress reflects the current rapid pace of development in the field of AI: it is only about 1 year ago that DALL-E2 entered the scene and was the first image AI to demonstrate what such an algorithm was capable of both in terms of image quality and image composition as well as the "creative" implementation of complex text prompts. Although the images often still showed numerous display errors and blurred areas, they gave an idea of the surprising quality that AI-generated images can have.
Early DALL-E 2 demo image
Since then, there has been a race between the big image AIs DALLE-E2, Midjourney, and Stable Diffusion, which continue to tweak their models to produce better images all the time. So far, the development has been particularly diverse with Stable Diffusion due to its open nature, which has been able to generate with the right model, photorealistic portraits for some time, on a www.slashcam.de/news/single/Pro-Bild-nur-15-Sekunden--Stable-Diffusion-laeuft-a-17782.html
(smartphone running locally) and thanks to ControlNet now allows extremely precise control over image composition. Here is our new introductory article Stable Diffusion - Tips and Tricks - Part 1 - Setup and Getting Started.
Stable Diffusion plus ControlNet
But Midjourney has now come forward again with version 5 - and the other image AIs have to follow suit. Midjourney now shines especially with photorealistic images of people, which are displayed in extreme detail and are indistinguishable from portraits of "real" people.
Midjourney V5 Portrait
Even famous personalities can now be portrayed very realistically. Emphasis has been placed on realistic skin textures including blemishes and imperfections, which make faces appear even more natural. Portraits of a person from different perspectives are also possible.
But also other objects are rendered with enormous detail.
The typical AI misrepresentations of hands or teeth, which were often used to identify AI images, have been virtually eliminated. Also the display of fonts is better now, but still not perfect and the displayed texts are still not meaningful - but this could change for example by interacting with a multimodal AI like ChatGPT, which can handle both images and texts.
Also, images with many nine aspect ratios can now be generated. Errors still occur, however, so sometimes objects or people - especially in the background - are deformed. But also these problems will probably soon be a thing of the past. Midjourney now also understands text instructions in natural language better, i.e. it no longer has to string cryptic keywords together. The goal, according to David Holz, founder of Midjourney, is to make it more beginner-friendly. The current (alpha) version 5 is still a long way from the final version, he says, and there will be many significant changes.
The possibility to create even more photorealistic and (AI-)error-free images now, together with the improved representation of known personalities, naturally increases the already existing risk of photo fakes.
What does Midjourney cost?
Unfortunately, Midjourney does not offer free image generation - unlike other services, you have to subscribe for at least a month at the price of per month in the Basic plan, which includes 200 image generations or for the Standard plan, which offers 15 hours of fast generation time or unlimited in the relaxed mode which takes a bit longer.
The evolution of Midjourney
But it is to be expected that the open source Stable Diffusion will soon follow with similar improvements and so self-generated images with an even higher quality than before will be freely available, and so thanks to the open source nature of Stable Diffusion soon photo-realistic images can be generated on any home PC (or even smartphone).
Aerial Photography via Image AI
more infos at bei www.midjourney.com
deutsche Version dieser Seite: Bild-KI Midjourney V5 liefert fotorealistische Bilder - und endlich auch korrekte Hände