Ill explain the new process im doing, because the old process for the characters is bad and not modular at all lol. The new process still will get better consistency between renders and y

I use comfy UI, have a very large node structure for each character, and the pluging Lora Manager to manage my loras for a graph. 

Each character has their own workspace and the loras associated with those characters. I typically use 3-4 loras per character, because after that i either get dimising returns or comfyUI generates some nightmare blob of pixels.

To keep some consistency between characters, I ill feed in the previous image generated. For example, if im working on tranformation 2, ill use transformation 1 as the latent image  with a input for an advance k sampler  then set the start_at_step anywhere from 2-10 and test.

After that the image will go through a node called FaceDetailer and then save out the final image.

The last note is i use illustrious for my image. I use to use pony but i like illustrious more now lol

I hope this is helpful!