So, just to satisfy some people's curiosity, I will explain this.
I'll start with a reminder that I usually do the Lora to use them myself. Often I make them to make ONE image that I have in mind and never touch them again. When I do a Lora, I do it with a particular idea in mind, so I usually adjust it for that purpose.
For example, when I have a character that I am going to generate a lot in all sorts of styles and poses and costumes, I train it using fan art of multiple styles so it is as flexible as possible and doesn't lead to a particular style.
If I want to make images that look like screenshots of a particular series, then I use screenshots only and don't put a single non-screenshot image in there to do the training.
That being said, WHY sometimes do I put two characters in the same lora.
The answer is simple. As you all may know, the more Lora you put into a single image, the more "messy" it gets. So an image that uses one Lora is going to be less messy than an image with two character Loras on it.
Of course, it's not that you can't work with two different Loras; you can; it's just that when you generate, if you only use one character Lora, you have more room for more and fewer issues.
Long story short, I put two characters in the same Lora when I personally intend to generate an image of them together or think I'm going to generate both together often.
On characters trained with fanart, this effect is not very strong, so I do not worry much about it much because the effect is not going to be that problematic.
But when it's characters with a strong style embedded in them, I prefer to put them on the same Lora, because on top of the image getting "messy," the style is going to double down and come out extra strong.
Even when you use a regional prompter, the styles are going to add to each other, turning into an excessive overall style weight.
This can be "fixed" by lowering the str of the lora of each individual character a bit (which can affect the likeness depending on the character) or impainting or other methods, but it is simply more comfortable and simple to have both on the same lora.
To make it clear, see this:
The image is exactly the same prompt, model, and regional prompter, but the image on the left was made using two different characters Loras. The image on the right was made using a Lora with both characters in it.
Since the Loras have a strong "flat anime" style, when you put TWO of them in the same image, it may look a bit TOO flat, or at least it looks that way to me. While in 2-in-1, the image of them together looks more natural.
So...that's all the mystery. I put them together when I think I'm going to generate them together and don't want to deal with issues.
Then again, it's not that you can't deal with these issues if they are in different Loras, but since I'm doing the Lora for my use and I know beforehand the issues I'm going to face, why complicate my life having to play with str levels and inpainting when I can simply train the Lora in a way I won't have to deal with it?
IMAGES: Image post by neclordx | Civitai