There are some unique challenges with generating multiple characters with AI. There's a reason why almost all images focus on a single character. The more focused and detailed the image, the better the AI is at working with the prompts and patterns. It can do a portrait of a face very well. It loses its fidelity when you do upper body, and especially whole body. It struggles with multiple characters and especially in long shot / out of focus scenes.
By circumstance, AI can be good at hitting the right notes for QS images because our work tends to only use upper body (or just head shots), so we can get away with more detailed generations. However, it struggles a lot with multi characters. AI doesn't distinguish between characters A, B, C, etc. like a human can, so it tends to create clones, and as the scene becomes bigger, each character loses quality and distinctiveness.
This is what you get when you put in "cheerleader team in quicksand", the way Duncan put it.
cheerleader team in quicksand.png
It's not a simple matter of typing in a line and getting amazing results. You don't simply ask the computer to make you that scene.
Here's the same prompt combined with the other prompts and tags I've been using to get my textures and setting.
cheer_team.png
Suffice to say, while the overall "artwork" is very nice, the lack of detail is appalling. It's certainly no Acidtester cheerleader bus crash drawing. It's impossible to decide how deep each individual character sinks, what they are doing, their expressions, etc. The AI just lops them into the same kind of thing. Cheerleaders gonna cheer, teams gonna team. That's probably how it reads the prompt. And while it does lend well to the imagination, the lack of care and detail make it grating to look at.
So we have to change the approach. Making multi-character scenes in good detail means creating each individual character, then splicing the images together and running it back through the AI.
As a benchmark, here's a solo cheerleader.
cheer1.png
The solo one is relatively easy after developing my method and prompts. I can do dozens of these with almost no tweaks. This doesn't make them easy to make for someone who doesn't know how to work with AI, but this is a baseline standard.
cheer2.png
With 2 cheerleaders, the difficulty became exponentially harder. I generated two individual characters with defined characteristics (notably hair colour) and spliced them together in Photoshop, then ran it through the AI again. Things begin to get frustrating as the AI doesn't recognise that there are necessarily two different characters in the scene. I had to do more inpainting to prompt the AI to create a background, but it was easy enough to keep the distinct hair colours, and the mud ripples were excellent. However, you can see the AI cloning the faces to be uncannily similar.
cheer3.png
With 3, it becomes headache inducing. Firstly, generating three good characters with the same style, right poses and right textures is already taxing. Splicing them together means blending in the mud and background or drawing a new one. Running it through the AI means that now you have three separate characters, each with things that can go wrong, plus an entire background that can go wrong. This one took the better part of an hour to get to look like this. While the central character is distinct, you can see the two others begin to look the same. One benefit is that since the AI likes to clone similar features, it's easy to create a uniform look for... uniforms. I'm not sure I want to push more than three characters, as the rewards are often not worth the effort of forcing the AI to get the quality given the extra scale.
As an aside, the problem isn't too dissimilar to my problems when I did art. Making one character was already a ton of work. Drawing two or more was intensely extra effort. This was why I didn't really jump into making Quicky Sanders work with AI, since most of my QS artwork featured multiple unique characters in unique dynamic poses, which the AI struggles with. It certainly can be done, but the process is like bashing rocks together.
You do not have the required permissions to view the files attached to this post.