Visiting The Fae
These first few posts are going to be about fairies and mushrooms. It’s not my usual thing, really, but I’m doing some clipart type images of mythical creatures in my signature ink & watercolor style. We’ll get into adding my style later. First we have to have a decent image to apply that style to.
I came up with this prompt, just on a whim
a group of fairies dancing in a circle around a mushroom. the mushroom is on fire.
I started with this a while ago, and I believe Midjourney v6 has advanced quite a ways since i started (the worst part about trying to document process on MJ is that it moves so quickly). A lot of people say that this construction works well for MJ 6, and I’d agree. It can take subsequent sentences about the things you’ve already asked for, as long as there’s no ambiguity. “The mushroom is on fire” is better than “it is on fire” because, while it’s probably clear to humans, the model will have trouble figuring out if “it” is the mushroom or the fire or the circle or maybe even the fairies.
After some false starts, I ended up with this prompt:
fantasy-themed clipart of a group of fairies dancing in a circle around a mushroom. the mushroom is on fire. 2d flat vector art on white background –style raw –stylize 800
Even though I’ve been at it for months, I’m still trying to get the hang of style vs stylize. I know academically what they’re supposed to do, but I keep finding the results unpredictable. So I did a quick exercise to isolate some of the differences
I’ll start with a new run of my prompt so i can copy the seed value. We’ll start with no arguments. I run with stylize medium, which sets --s to 100, so we’re still involving the model’s visual sensibilities, but not too much.
fantasy-themed clipart of a group of fairies dancing in a circle around a mushroom. the mushroom is on fire. 2d flat vector art on white background
My original image has seed 623181606 so we’ll specify that from here on out.
Here we get general prompt adherence, and enough of the model’s style to arrange things decently inside the image.
Here’s what I get with --style raw to take some aesthetic decisions away from the model. Taking some of those choices away, you get images closer to the prompt, but maybe at the expense of the images being a touch less coherent, making a little less sense visually.
fantasy-themed clipart of a group of fairies dancing in a circle around a mushroom. the mushroom is on fire. 2d flat vector art on white background –style raw –seed 623181606
Now we’ll play with the interaction between these two values. Here is --style raw with a lower --style 50 argument, equivalent to setting “stylize low”. Weirdly I think this is noticably poorer prompt adherence, with the right elements but just kind of thrown all over the canvas.
fantasy-themed clipart of a group of fairies dancing in a circle around a mushroom. the mushroom is on fire. 2d flat vector art on white background –s 50 –style raw –seed 623181606
if we go all the way to --s 0 we get quite a lot of prompt adherence. This is fine with me, as we can “pretty it up” later.
fantasy-themed clipart of a group of fairies dancing in a circle around a mushroom. the mushroom is on fire. 2d flat vector art on white background –s 0 –style raw –seed 623181606
For illustration, we’ll go --s 250, equivalent to stylize high. None of the shading and depth that MJ wants to give it by default, but otherwise nicely arranged.
fantasy-themed clipart of a group of fairies dancing in a circle around a mushroom. the mushroom is on fire. 2d flat vector art on white background –s 250 –style raw –seed 623181606
Just to prove our point about involving the model, we’ll also look at these without raw mode.
Here’s --s 50 without --style raw. Notice the overall cohesive circular arrangement, along with some odd details like mushrooms growing out of mushrooms or the mushroom cap being upside down. But the image still has some symmetry to it and you get a nicely composed image, even if what’s in the image is odd.
fantasy-themed clipart of a group of fairies dancing in a circle around a mushroom. the mushroom is on fire. 2d flat vector art on white background –s 50 –seed 623181606
And here’s --s 250 without --style raw. We have a lot more detail in the characters, but we still aren’t completely sure it knows what a mushroom looks like.
fantasy-themed clipart of a group of fairies dancing in a circle around a mushroom. the mushroom is on fire. 2d flat vector art on white background –s 250 –seed 623181606
In general these are the use cases I find for these paramters:
Add --style raw along with a low --s value when the model doesn’t seem to be listening to my prompt. It will take less liberty and give me something closer to the prompt text, even though I’ll have to dress up the presentation later. This is fine by me because I tend to add styles last.
High --s values are great for early in your process, when I’m still trying to see what’s possible. The more iterations I do, the lower I set the --s param so that I’m more in charge of what’s changing.