
On Wednesday, Midjourney announced model 5 of its industrial AI picture synthesis service, which may produce photorealistic pictures at a high quality degree that some AI artwork followers are calling creepy and “too perfect.” Midjourney v5 is obtainable now as an alpha check for purchasers who subscribe to the Midjourney service, which is obtainable via Discord.
“MJ v5 at the moment feels to me like lastly getting glasses after ignoring dangerous eyesight for somewhat bit too lengthy,” mentioned Julie Wieland, a graphic designer who often shares her Midjourney creations on Twitter. “All of a sudden you see every little thing in 4k, it feels weirdly overwhelming but additionally superb.”
Wieland shared a few of her Midjourney v5 generations with Ars Technica (seen under in a gallery and in the primary picture above), they usually definitely present a development in picture element since Midjourney first arrived in March 2022. Model 3 debuted in August, and version 4 debuted in November. Every iteration added extra element to the generated outcomes, as our experiments present:

Ars Technica
Midjourney works equally to picture synthesizers like Stable Diffusion and DALL-E in that it generates pictures primarily based on textual content descriptions referred to as “prompts” utilizing an AI mannequin educated on millions of works of human-made artwork. Just lately, Midjourney was on the coronary heart of a copyright controversy concerning a comic book e-book that used earlier variations of the service.
-
An AI-generated “artificial {photograph}” of a lady via a window generated utilizing Midjourney v5 by Julie Wieland.
Julie Wieland -
An AI-generated “artificial {photograph}” of a cheeseburger generated utilizing Midjourney v5 by Julie Wieland.
Julie Wieland -
An AI-generated “artificial {photograph}” of a boy and flowers generated utilizing Midjourney v5 by Julie Wieland.
Julie Wieland -
An AI-generated “artificial {photograph}” of a clown generated utilizing Midjourney v5 by Julie Wieland.
Julie Wieland -
An AI-generated “artificial {photograph}” of a lady generated utilizing Midjourney v5 by Julie Wieland.
Julie Wieland -
An upscaled model of a Midjourney v5 output with the immediate “a muscular barbarian with weapons beside a CRT tv set, cinematic, 8K, studio lighting.”
After experimenting with v5 for a day, Wieland famous enhancements that embody “extremely real looking” pores and skin textures and facial options; extra real looking or cinematic lighting; higher reflections, glares, and shadows; extra expressive angles or overviews of a scene, and “eyes which might be nearly excellent and never wonky anymore.”
And, after all, the arms.
Only a heads-up – Midjourney’s AI can now do arms appropriately. Be additional crucial of any political imagery (particularly images) you see on-line that’s making an attempt to incite a response. pic.twitter.com/ebEagrQAQq
— Del Walker (@TheCartelDel) March 16, 2023
Over the previous yr, the concept AI artwork turbines cannot render arms appropriately has develop into one thing of a cultural trope. Notably, Midjourney v5 can generate real looking human arms pretty nicely. “Fingers are right more often than not, with 5 fingers as an alternative of 7-10 on one hand,” mentioned Wieland.
Within the service’s Discord launch notes, Midjourney additionally famous that v5 now responds with a “a lot wider stylistic vary” than model 4, whereas additionally being extra delicate to prompting, producing much less undesirable textual content, and providing a 2x enhance in picture decision.
If there is a visible draw back to the Midjourney improve for AI artwork followers, it maybe comes from pictures that may be so real looking and “excellent” that the mannequin’s precision takes away among the thrill of repeatedly producing AI imagery to discover a appropriate consequence, what one may name a “slot machine impact.” Though one Twitter consumer named Philipp Lenssen noted, “If in case you have a particular picture topic in thoughts, it is nonetheless a bit like lottery. However with greater profitable probabilities than v4.”