Transcription
Good day, guys!
Today I'm going to show you how you can control the poses of characters you generate in MidJourney. It doesn't matter what your character looks like or what type of post you want; by the end of this video, you'll have the necessary skills to control all types of character poses.
Now, I have three characters here ready, and I'll be using them to perform different poses. I just want to show everyone that you could do a lot of different poses all over MidJourney, and you have some degree of control.
For starters, I'm going to generate a pose of this lovely young lady. The first lesson in generating the right pose is that the prompting needs to be on point. I have a cheat sheet right here for camera framing and distance, the camera position—bird side view, worm side view—and of course the subject orientation—frontal view, back view, angle view, and side view.
Let’s say you want your character to be, for example, sitting down. You have to also specify what angle will the viewer see her. For instance, let's try to generate a high angle view of her sitting down. I'm just going to copy this prompt, paste it here, and edit it. Since I want a high angle shot like this one, what I'm going to do is type here, "high angle shot."
Let's just say "illustration," because we're creating an illustration, not a photograph. So you can delete the rest of the descriptors because we'll be using the character reference parameter technique, where we'll be able to copy her likeness, including her clothes and her face. I could just leave out “29-year-old female African-American astronomer with curly hair," I'll probably just leave the hair out, and then I'm going to put here "sitting down."
So remember, we have the composition here and then we have the pose. All right, let's generate it. Sorry, I forgot to use the character reference. Hold on a bit. So I'm going to use the character reference parameter here—CF, copy her. The link is here; you could also just drag and drop, by the way. If you don't understand what CF means here, go check out this video; it explains everything you need to know.
All right, so we got this. As you can see, the angle here is high angle. The two images at the bottom didn't follow the prompt instructions very well, but the ones at the top did. So as you can see, let me show you another example.
I generated a low angle shot for the Mexican character, and as you can see, it is sort of a low angle shot, but not quite. Maybe this one is the closest thing we've got. So we're already seeing resistance, even though we're still in the composition stage. Welcome to AI!
I tried to generate another version of him, but this time I added more alternative keywords. As you can see here, there are a few alternative keywords that we could use in the prompt, like upward shot, below shot, and under shot, which I used here and also here. Unfortunately, MidJourney is still a bit too stubborn. I think this one is good; this one is just off—too off.
This is a high angle shot, but this one is good; this one is good. So basically, you have to do trial and error. Eventually, it will generate the right image; you just have to do a little bit of experimentation. But the principle that you need to follow is to just use the right PRP keywords. If one prompt keyword doesn't work, try multiple.
Let's try the other camera angles! Oh, before I forget, you can get your copy of this template which I use; it's very useful if you're trying to generate different compositions, all the alternative keywords that you may want, and you could use this in your prompt. The link is in the description down below, so just download it now.
I tried the 3/4 view following this one, an angled view. I used a few different keywords here—angled shot—and then this one, 3/4 view shot. So this one got it right; this one a little bit. These ones are errors, so half the time you get it right, half the time you don't. But it's okay because you could just generate as many as you want, right?
This time, let's try to generate something that's a bit difficult, something that's more challenging. We're going to generate a back view, and you'll find out why it's difficult. Now, some of you might think that generating a character's back view is easy, but that's only true if you're not using a specific character reference.
I'll show you how to generate a consistent back view for any character you create, ensuring it's not just a random image that actually matches your character perfectly. We're going to use this guy; I'm just going to copy this one. By the way, I'm using Niji version for this character. Niji is the anime model for MidJourney.
For this one, I'm going to type "back view." So the other keywords are rear view, back shot, behind view, and a back view shot of a warrior with a long beard. I'm just going to write here rear view with red long hair. I'm just going to leave that there, wearing knight's armor, and I'm going to delete the rest of the descriptors here because we won't be needing this one. I'm just going to put here "standing well, facing away from the viewer."
So, yeah, I put in a lot of keywords there, so we should be able to generate it right, right? Oh, and don't forget to put the character reference here; otherwise, it won't work. We need to keep the characters consistent, so we're going to put this. All right, and let's find out.
As you can see, guys, it didn't work! Why do you think it didn't work? The reason is AI is really bad at extrapolating what it doesn't see from the reference. Since we're using this guy as a reference, the AI has to imagine what his backside looks like, but his backside isn't seen here. However, there are ways we could like force the AI or give it enough reference so it could generate the back view.
So that's where we go into the advanced methods. Let's go now. As you can see here, it wasn't able to generate the back view properly, but it did generate the character properly. So this is the character reference I used for this one—pretty cool, right? A photo-realistic character reference—but I was able to generate a comic style art.
But anyway, as you can see here, this looks like a jacket but inverted. It's an inverted jacket. So why is that? The AI confuses the front and the back, that's why. Using these things, you have to be a bit creative, and this is what I'm going to teach you right now. By the way, I actually pre-generated the images so teaching you would be easier.
Starting with this one, I like this pose, so I upscaled it. Here, as you can see, there is a way to just fix this area. The AI was able to generate him, so all I need to do is fix the clothing. You could do that with very region; it's loading very slow. That's what happens when you generate an old image—it's very hard to use in painting again.
But anyway, just to explain what you can do here—you could edit this part. You could regenerate only this area; you could select an area you want, you just paint all over it, and then MidJourney will generate a new version based on the new prompt that you will indicate. In this case, this is the new prompt I typed, and it was able to generate a new backside.
So that's one strategy to overcome difficult poses, but the problem is this doesn't always work. Some characters or art styles—MidJourney just can't generate them properly, even when using in-painting, like the image I've shown you earlier. There's no way you can just fix this with in-painting, right? Because it's facing the wrong direction.
So here's an even better solution: you will create a character sheet out of your character. This is especially important for the back view because you don’t really care so much about the facial consistency too much, right, 'cause this is back view. That's what I did here; I created a character sheet.
Let me show you the character real quick. So this is the character, the original character, right? And so I wanted to generate a back view of him, but the problem is it's too difficult when you're using a character reference. So what I did is I created a character sheet. Take a look at the prompt I used here: of course, the descriptor, the subject, and the description.
I just added this magic prompt right here—different poses, multiple angles, character sheet, white background—and you need to specify an aspect ratio that's a bit wide. That's all you need to add. That's what I was able to do here; I was able to generate different views of him. Here are other examples.
I’ve done a bit of experimentation here, trying to change the prompt a little bit, but overall, it's just the same. I just tried to add more descriptions here. Anyway, I wanted to generate something that's true to my art style, so I could just use any of these as a reference. That's how I generated this one, which is a back view version of this guy.
When you're creating comics, that's all the consistency you need for back view. So I'm going to show you how I use this character sheet to generate something like this and this. So as you can see the links here—me show you, that's one, that's two, that's three, and so on and so forth.
So basically what I did was I added an image prompt, right? However, it's not just that; I also use it on the CRF parameter right here. As you can see, I used the added image prompt here and also used it as a character reference here. You could just screenshot this one and put it here, and then you got this.
You could use this as a link to just prompt imagine, and this is the link it will shorten once generated. Then you just type in the prompt like this one; after CF, which is the character reference parameter, you just add more links—the same link, this one you just copy-paste—and then the other links for this and this and this and this and this, you know, and so on and so forth.
Having multiple references does help, but I'll be honest with you; you probably need just two or three—then that will be enough. In this case, I just tried to overcompensate just to make sure that it really generates the back view.
Now, what I've shown you is an advanced process because generating back view is usually the hardest type of pose, so I'm going to show you something else, a bit different. For this one, I wanted my character to be floating in the air as if he were being carried by the wind, like some Buddha or something. You're going to have to use Google Images to find similar poses to what you want your character to perform.
So check out the image prompt I used here. See? It was able to copy the pose very well. If there's one main takeaway from this video, it's to use an image prompt as a pose reference—it's by far the most effective way to get a pose right.
However, other variables can affect your results, and sometimes adding an image prompt isn't enough and may actually ruin it. In the next section, I'll show you how a mix-and-match approach can help you achieve optimal results. The character reference pose heavily influences the pose of your new generation, but that's not all. Even the style reference, this one, also influences the pose of what you're about to generate.
See, AI is not perfect; it has those types of biases, right? You have to be a bit careful with the type of reference that you use. As much as possible, try to copy a similar pose. Let me show you something else, like for this guy. As you can see here, he's running, right? But I'll show you the character reference of him—he's just standing there.
He's just standing there. When I generated the first images of him, it was hard to get him to a running pose even though I prompted it very well—"running fast," "sprinting." What I did and what fixed the problem—oh, by the way, I'm using an image prompt here of a guy running, and that didn't help.
So what I did was I used the style reference here—this one—and this helps influence the final generation, making him look like this, right? So now he's really in a running pose. So again, even style reference affects the final generation. The character reference affects it more, but usually, the character reference is just a pose of standing.
Now, I want to give you a bit of a warning. Be careful when using image prompts. You know why? Because it could massively change the appearance and art style of your character. Let me show you this one. The original art style was supposedly this, and it got like this.
Look at the style reference; the style reference is supposed to be the style that we're copying. Yeah, this guy! You see this style and this style? See the disparity between the styles? The reason for that is because of the image prompt I used—this one. You can see I used a weird image, and this rough outline transferred into my composition.
As you can see, there's a rough outline; it's the same here. So how did I fix it? I used a different image prompt—this one. When I used this prompt, it immediately reverted back to my original art style, which is this. See? No more rough outlines here.
Of course, you could change up the character reference—this one and this one—the links here. If your character is a little bit deviating from its original look because of the image reference that you used, another thing as you can see from this image—it wasn't able to generate his whole body completely.
Let's say you want the whole body to be in the image, right? But this one, a few parts are excluded due to the canvas size. So what you can do is simple: you upscale the image that you want. For example, if I want this one, I upscale it, and I got this.
What I'm going to do is I'm going to extend it here using either the zoom-out tool—this one—or the pan tools, as using this. What I did with this one is I actually used the pan tool—this one—which moved the canvas to this side, and you got this. If ever you have parts of the body that are excluded, you can just extend the canvas using the zoom or the pan tools.
Now, I'm going to show you another difficult pose, which is the spit take. This gave me a lot of trouble, believe me! So you're welcome. Right here, I'm generating Leonardo DiCaprio as the character. Yeah, I got Leo in my story. By the way, for those of you who don't know, a spit take looks like this: something takes you by surprise and you accidentally spit what you were drinking—like, yeah, you know what I'm talking about.
So the problem is it's so difficult to do that with MidJourney because maybe it doesn't have a lot of training data on spit takes. But here's how I succeeded in doing that. So this is the character reference that I used, and this is the style reference; so this is for like a colored manga that I am creating.
I didn't use an image prompt, and here's why. An image prompt is extremely important because sometimes with MidJourney, words just aren't enough. So I added an image prompt—this one. Well, as you can see, it doesn't look like a spit take. I don't know; what do you think? What does it look like?
It's doing, and then I use a weird image, which got me this type of result, which I didn't like, by the way. That's because I used this as the image prompt. So again, I keep repeating myself because it's so important to use the right image prompt. And if the image prompt you use doesn't work well, you can just try another one. It's always a trial-and-error thing until I got to a point where I generated this.
This was as close as I got, and this is the image prompt that I used. It took two wrong image prompts for me to get to the right one, so expect something like this when generating different poses—especially the difficult ones.
Another lesson that I want to instill in you is to make sure that you use the proper aspect ratio. Why? Let me show you here. Earlier I showed you that we could generate a character, and if parts of his body were excluded from the canvas, you could just use the zoom-out and pan tools to make your composition complete.
But the problem is there are times when, because of the aspect ratio that you've set, the image really won't generate right. Like, it just doesn't generate the type of pose that you want. Now what you can do is just change the aspect ratio and see what happens.
For example, with this one, I wanted him to perform a pose like this—like throwing something. I didn't even want to follow this exact pose, but I just wanted him to look like he’s throwing something. However, as you can see, he doesn't look like he's doing that. He looks like he's dancing.
I tried it again, even changing the prompt a little bit, trying to add new image prompts—this one. So now we have two, and still, it's the same thing; he looks like he's dancing. So what fixed it? Well, the moment I removed this aspect ratio, so it defaulted back to the square aspect ratio, look what we got. Now, still a little bit like it's dancing, but it's already close.
In here, I finally got this one. This is a throwing pose—check this out!
Yeah, see?
All right, so we've generated a bunch of poses today, but now I'm going to generate five more difficult poses, and I'm going to do it live in real-time. Some of you have probably seen it live.
So I just finished filming live, and I'm proud to say that I was able to accomplish all five poses successfully and within one hour.
Let me show you.
First is the sitting cross-legged. It was so easy; the first generation I made was already able to accomplish it.
The second one is the falling from height. This one was the most difficult; it required in-painting the face because it was so difficult to really generate the falling down pose just from using the character's reference image.
The third is the flying kick. This one was easier than I expected as well; it just took a little bit of careful planning with the aspect ratio just to make sure that there's enough room in the canvas for her to perform a flying kick like this, so it needed to be a landscape format.
The fourth is the handstand—handstand with one arm. This was easier than I thought as well; I thought I would have a really hard time with this.
Lastly, is the yoga tree pose. This one took a bit of creativity; I was expecting this to be extremely difficult. It was a little bit difficult, but I was able to do it relatively quicker than I imagined.
So I hope this proves that you have some control over character poses in MidJourney.
Before the video ends, I want to mention that if you want to precisely manipulate characters, Stable Diffusion might be a better AI generator for you. Its Control Net feature is much more accurate.
If you feel you have questions, feel free to reach out to me in my Discord group; the links are down below.
See you in the next video!
My only medicine! Yeah, everything I do, I'm just being genuine! Yeah, I'm SI.