Home Creators Posts Import Register Favorites Logout

Content

Prompt: detective looking for clues, candid shot, polaroid, outdoors, cinematic lighting, --ar 5:2 --style raw --v 6.0

So MJ v6 is out and I feel that I couldn't track all the differences of this new version. It feels new, but also a bit old and consequently frustrating. But new versions have their own magic, so I am enjoying my experiments. Also, sometimes versions change so fast (--creative, 4a, 4b, 4c, 5.1. 5.2, raw...) that it's not time for a verdict — from me, at least. 

Instead of a few influencers, I don't think extremely mind-blowing (you know this kind of post, right?) yet, but I can see good improvements. So I think the best way of dealing with it is to offer some scattered thoughts:

1) New Prompting? Yes and No

MJ team states that "prompting with V6 is significantly different than V5. You will need to 'relearn' how to prompt." In a sense, it's true, but v4 started that with simpler prompts. After a few weeks, it was responding well to longer, trending-at-artstation-prompts like before. So... I am not sure. My tests with old prompts (short and long) were not 100% surprising. But yes, you need to re-learn a bit (more about that later)

Prompt: photorealistic D.Va wearing Iron Man armor, Asian 19yo girl behind her blurred megacity scene, dynamic pose, digital painting, ray-traced, octane render, 8k, trending on artstation --ar 9:12 --v 6.0 [ --testp --upbeta --creative ]

Prompt: Don Draper, painting by Kenton Nelson

2) Never Out of Style

At the same time, you need to be clear with your words. I think the best example is my old "oobleck" prompt! Compare MJ's old versions with v6:

Prompt: guy made of oobleck v4, v5, v5.1.

Version 4 and 5.1 were more opinionated and "creative" than v 5.0. With 5.0, we had a real guy covered but not made of oobleck. Now let's see with v6:

Prompt: guy made of oobleck --v 6.0 --seed 887778012

V6 Raw tends to generate a realistic picture: it's more probable you will find a man covered on oobleck than made of it. (It doesn't mean that it won't generate an oobleck man). Styles closer to 0 bring the man composed of this weird substance, while high stylize will give you an artistic/artsy image of a person (not necessarily a guy) with a cool bodypaint or something like that.

3) Be Clear not "Vibey"

You really need someone covered on oobleck? Make it clear!

Prompt: guy covered on oobleck --v 6.0 --seed 887778012

Also, "Vibey" words should be avoided, according to the MJ team. The next prompts were generated with the same seed and style raw. The first one is "two best friends at Disney", the second one is "two best friends having fun at Disney", and the third is "two best friends laughing and holding ice creams at Disney." Notice how images 1 and 2 are similar. "Having fun" doesn't say much so MJ needs more from you.

Also, make clear who those friends are. Sometimes MJ generates different friends, but in this case, most of my images had very similar young girls (No boys or old ladies, for example).

I am not saying that MJ will always follow your clear prompt, but I think this is good enough for a first try:

Prompt: Two friends laughing and holding ice creams at Disney, one is a black woman and the other is a middle age latino guy --style raw --v 6.0 --seed 2574713582

4) Text Generation

And of course, this is the biggest change in this version. And, unless you want to recreate the Book of Kells, it works very well for a first, alpha interaction! (See my last post). Put the words you want between "quotes," specify their function or place in your image and MJ will generate something closer to it. 

Tip: As I told you, --style raw or --s 0 improve your chances! 

Prompt: double exposure of woman with clouds, the word "PEACE" written , blue sky, birds, kites, photography, clean design --s 0 --v 6.0

Sometimes it's a bit literal like this:

Prompt: a painting of mona lisa with the text "New Features" written --v 6.0 --s 100

Prompt: cover of "THE HULK" comic book, art by Jack Kirby depicting the Hulk walking alone in a desert, vintage colors --ar 2:3 --v 6.0

==

So what is your first impression? 

let us know!

Jos

Files

Comments

Finn Wraae

I've thrown a bunch of my "old" prompts after v6 without changing anything, and it's working fine - giving me gorgeous new images with a new interpretation of my wording.

Gaffer Wiles

There was something that I picked up with the V6 (Alpha) announcement. I call it Alpha, because that is exactly what it is, an alpha release with a lot of potential but so very many bugs and things still to fix. What I picked up was this statement: "V6 is our third model trained from scratch on our AI superclusters. It's been in the works for 9 months." This means that they trained it on a lot of new data and this might mean that it might not recognize or be able to emulate some artists. And of course we have no idea if any of that data has been poisoned by "Nightshade". Perhaps there should be tests done on MidLibrary and the database updated to see if any of the artists or distinctive styles differ from the 5.2 renders. I am undecided if V6.0 is such a quantum leap forward. Its understanding of natural language is certainly significantly better. I was able to use ChatGPT to describe an image in detail and then cut-paste that description into MidJourney in its entirety and I was (about half the time) able to get something a lot closer to the original image. The prompts were 150 words long (excluding "the" "a" "an" etc. However, the upscalers (Subtle and Creative) are hugely expensive. I blew almost half my monthly GPU fast hours of Subtle and Creative upscales. Secondly they are not very good adding in artifacts and details (even in the Subtle version that messed up the images badly.) Running images through Real-ESRGAN gives far better results. (yes, believe me, those old AI upscalers worked better and it even fixes the terrible rendering of eyes and faces!) If you have Topaz AI or GigaPixel AI or even ON1 Photo Enlarge I am sure the results will be better. V6.0 is SSSSLLLLOOOOWWWW! It memory use on the GPU superclusters is significantly more than 5.2. It shows! A 25-word 5.2 prompt on "Fast" runs on average 32s, whereas a 100-word prompt for 6.0 takes 60 seconds. This is of course V6.0 Alpha and is not the official default MJ model, so take what I say with a pinch of salt. Don't waste your GPU fast hours on the upscaler. It isn't worth it. Using your old prompts on the 6.0 model is not a true test of the models capabilities. You need to push those prompts to the limits of the technology. In other words feed ChatGPT 4 an images generated by 5.2 and ask it to describe it in detail. Then add the extra details into the original 5.2 prompt and feed it to 6.0 and only then will you see a difference, and ytou will be able to truly see what 6.0's strengths are. Happy prompting.