- Midjourney founder David Holz announces that v5 will be released this week.
- Before that, there will be a final rating session: all paying customers can participate (see below).
- Midjourney has chosen Google Cloud as its partner for training and running its own AI models.
- For training, Midjourney is using Google's v4 TPUs with Deepmind's JAX, with inference running on "large GPU clusters".
Midjourney offers the best compromise between accessibility and results of any image generator on the market. The upcoming version 5 is ready to take generative AI one step further. The first alpha images are very promising.
The leap from Midjourney v3 to v4 was huge in terms of quality: the image AI generated more appealing and detailed images that better matched the input prompts. Midjourney is currently testing the latest algorithm, v5, which will be released soon.
Development of v5 is still ongoing: Midjourney is collecting human feedback by evaluating images generated by v5. The images currently circulating on the Internet are the result of this process.
Since the results of this feedback influence the model development, Midjourney says that these images are not representative of the v5 version to be released. The ratings and resulting data quality are “very serious” to the final version, says Midjourney founder David Holz.
Midjourney already relied on a feedback process for v4, which the startup says further improved the quality of the final v4 release.
Comparing the images of the last #midjourneyV4 rating party with what V4 actually turned into, I wonder how #midjourneyV5 will surprise us this time … 👀#Midjourney #AiArtCommunity pic.twitter.com/xmqF0JXanH
— Trist (@tristwolff) March 9, 2023
One Midjourney developer likens v5 to the “Ship of Theseus”: every component has been replaced at some point, but none has been developed from scratch, he said.
Midjourney v5 could be more detailed and realistic – and handle hands better
The v5 test images show a higher level of detail and accuracy, especially in photorealistic images. According to Midjourney, v5-generated images generally offer a higher default resolution (1024 x 1024) and can display a wider range of styles. The first photorealistic images show a higher level of detail.
Spent the morning rating 2,000 images. Here are a few of the ones I saved, organized by theme in this thread.
The food images surprised me the most TBH
– Nick St. Pierre (@nickfloats) March 9, 2023
Midjourney v5 Alpha Image. pic.twitter.com/3oFRuACX1J
— Smoke-away (@SmokeAwayyy) March 10, 2023
MidJourney V5 is going to blow some freakin' minds. These are wild! Last year, you couldn't convince me that 3 out of 4 of these weren't actual photographs! pic.twitter.com/0Hd0qBW7tz
— Matt Wolfe (@mreflow) March 9, 2023
A big problem with Midjourney and similar tools is the accurate representation of hands and fingers. There are usually too many or too few fingers on the hand, which are also strangely bent or twisted. This is so common that AI art critics have made a crippled hand the symbol of their dissent. What do the hands in Midjourney v5 look like?
They finally fixed AI hands?!
Yup, all these photos are fake.
Mega improvements in Midjourney v5: pic.twitter.com/IfO4Y5EBTl
— Lorenzo Green 〰️ (@mrgreen) March 10, 2023
1/ MidJourney V5 is coming out soon. I helped rate some of the output as they're calibrating it. Another step change, impressive. Look at the hands! pic.twitter.com/vTb9FJVGkn
— Tuur Demeester (@TuurDemeester) March 9, 2023
THE HANDS!!! LOOK AT THE DARN HANDS!!!!!!!#midjourneyv5 https://t.co/RzZraNhQm8 pic.twitter.com/UsfWtnMmyy
— DataChazGPT 🤯 (not a bot) (@DataChaz) March 9, 2023
In the preview material, the defects on the hands and feet are not completely removed. In some pictures, you can still see the characteristic mutated extremities. But there seems to be significant progress, and possibly the final v5 will be bug free.
Paying customers of Midjourney can join the rating party here and see more v5 images. It could go fast until the alpha release: The time between the v4 rating party and the alpha release was about two weeks.
Midjourney founder Holz is a long-term believer in generative AI. He thinks that one day whole game worlds will be generated by much improved Midjourney-like systems.