Navigating MidJourney: V4

As the designated artist round these here parts (and isn’t that a strange role for me to have fallen into. Also, proof that if you live long enough, you change profoundly) I’ve been playing with MidJourney for fun and profit for a few months now. Time flies. As does fruit if you propel it fast enough.

I’d said at the beginning of this that the AI tools were only going to get better. About a week ago now, MJ leveled up. Also, there are quite a few other options now, if you are looking to dabble in AI art, but I’ve been using this one as my sandbox. It’s fun, it’s becoming familiar, and there’s often new things to experiment with. I thought I’d get bored, but no. No I am not.

This post is going to dig into the nitty gritty of MJ in particular. If you are just starting out, I recommend reading the help pages over there. If you are getting serious about it, you should pick up Jack Wylder’s book on Ai art prompts, he’s far more an expert than I ever will be.

Version 4 is still in the infancy stages – you can’t change aspect ratios, the images are low-resolution – but it’s already a marked leap forward from what was possible before. Sure, the AI can’t do hands. Neither can most human artists. This is what post-processing in Affinity Photo is for, or Procreate (I don’t use nor recommend the big subscription software any longer). Regardless, a lot of what you can get is eminently usable in the same place you’d have used stock art before, for book covers, blog headers, promotion graphics, and so on.

The prompt I used for the exemplar image that I’ve chosen is:

cutaway view, experiment on translucent liquid sphere terrarium aquaponic pine bonsai, hyperdetailed –v 4

Note that V4 modifier.

This is the initial result:

This is only 1024px square. Fine for most (but not all) web uses, not much more. You’d definitely have to render multiple elements and stitch them with photocomposition to make a cover, for instance.

There are ways to work around what V4 gives you – and I’m guessing these won’t be needed for long, but it’s still interesting to see how the machine learns and can be used.

Next, I chose the ‘Beta Upscale Redo’ button from the presented options. I knew it would make subtle changes to my image, not as much as ‘Make Variations’ but I might lose details I really liked. However…

The result is now 2048px square, and I happen to know that at this size you can use it for an ebook and a print cover, not to mention actual prints. I’ve been happily playing with improbable art for some time, since the afternoon where I fired up my wide-format inkjet printer and tried to make a mess out of MJ images… and failed. I could, and did, print up to 13″x19″ without pixelization or excess loss of details. They look amazing, and I have at least one framed and hanging on my wall.

But! What happens if you take this result, which offers you even more options, and push it further?

I hit the ‘Remaster’ button first.
This is the result of the Remaster render. Remaster is interesting, in that it will give you a very realistic image. Including removing ‘fantastic’ elements of the image. It’s somewhat less creative, in that aspect, and usually gives a ‘smoother’ result. Resolution on this is 2048px square.

Finally, I tried out “Detailed Upscale Redo” to see what I’d get. Immensely more detail, leading to a somewhat cluttered image, and the return of the characteristic MidJourney noise. The resolution here, interestingly, is 1536px square, a reduction from it’s parent image.

I could, by they way, keep on iterating with this image. I know from experience that I’m likely to see more distortion the longer I do this, however. Usually a few iterations is good. Too many, and you start to get frankly Lovecraftian. If I were anthromorphizing MJ (which I totally do, all the time) I’d say it gets confused about what I want, and starts adding or subtracting. Because this particular AI is not collaging from images it’s been trained on, but rather rebuilding techniques to create images based on my prompts, it’s going to have trouble with things like, say, anatomy. However, V4 has made a giant leap forward in terms of stuff like, say, horses. Or Battle Unicorns.

It’s also gotten better at scenes, like the SciFi coffee shop I used as a header image.

28 comments

  1. I’ve been liking V4, but I’m finding that my remasters on V4 are often a step down. Not sure why. But the images are so good that the bad hands are even more disappointing. I may have to invest in editing software.

    1. If you want to go with Freeware there is Gimp. But I do like Affinity Photo and it’s up to V2, just been released and is on sale. As for hands and quality, I think it’s going to get better. It’s turning the art world on its ear as it is.

      1. Thanks for the tips. I’ll look at both. I’ve got some great stuff that just needs the hands cleaned up or the swords shortened.

  2. I’ve barely got my pinkie toe wet yet in MidJourney, and there’s little I can do in post-processing (don’t have the art skills), but I’m finding it absolutely fascinating. Not for book covers — I like my artist — though I might reconsider if I needed a bunch for, say, a group of shorts.

    But for blog posts and the like… wowsers! It’s a great tool for making metaphors visible.

    1. I’ve been using it to do character sketches as well, which is interesting as I like the element of random I get. I rarely have a set mental image of a character, not how my mind works, so using MJ gives me a lot of imagery to describe.

  3. *giggle giggle* Sorry, the first image is “Thomas Kinkade does bonsai!” It’s the pastels in the rocks that give off the Kinkade feel. OK, now to read the rest of the article.

      1. Given apparently Frank Lloyd Wright’s houses supposedly fit very well in Thomas Kinkade landscapes, I findyself wondering what happens when that gets thrown into the mix?

  4. I’ve been playing with V-4 this morning and I love it. Yeah, the hands suck, but there are ways to cure or hide them. But I love the improvements in this latest version although, as mentioned above, the remasters aren’t always as good as they were with V3

  5. I see that last tree globe and its oddities, and find myself thinking that could be an interesting premise for a scifi setting; that it was essentially built by AI and as a result has some… strangeness… to it.

    Might fold a bit of that into the background of my world, actually.

  6. I wasn’t completely sold on V4 for a while, because I really liked the “like a painting” look I could get with v3, along with artist prompts that made things sufficiently abstract that hands and guns were no issue (coughBill Sienkiewiczcough). But I’m coming around on V4.

    I’m still highly annoyed at some of the idiot restrictions they have. I get that they don’t want people using it for nothing but porn, but still, the word restrictions are downright puritannical (how, one wonders, can you do proper horror images without “corpse”?), and the “no nudes” policy is directly contrary to thousands of years of art that was not porn. Yeah, they have to set rules and keep people from doing what people do, but still, it’s obnoxious.

      1. Sienkiewicz is not for everybody, but is an absolute master artist, across many media. And I don’t think he would have done the long runs on various titles that he did if he hated what he was doing. I can absolutely understand not liking his work, especially his paintings, but I generally love it myself.

Comments are closed.