Nice to see more people taking an interest! I'll drop some knowledge I've gathered in case somebody may benefit.
The best practice we've discovered so far is that it is better to recreate the original image from scratch using txt2img and once you have a good naked, move on to inpainting for clothing, faces, etc. Make sure the head is inside the frame, preferably with enough room to add hats later on.
Using the original images from the game unfortunately won't get us very far.
copy-paste the generation info for the character into a txt file if you stumble upon something good.
To maintain consistency I would use inpainting instead of img2img because you can draw masks on the picture to limit what areas the AI can change. You can invert the masks from the ui to paint everything that isn't masked instead.
Remember to utilize the seed and prompts used for the original generation during inpainting! Even if you're only changing the eyes or mouth, etc, make sure most if not all of the original prompts are in there, and from there you change only the bits you want. Otherwise you may get drastically different looking results.
parentheses (()) also are really good if you want to emphasize a prompt more. It made outfit generation a lot easier for me.
Denoiser is a slider that regulates how creative the AI is.
change the cfg if your images are oversaturated or too sharp, look weird. The lower it is, the softer the drawing.
step size - anywhere from 20 to 50 is good enough. Sometimes you may wanna experiment with it.
For negative prompts, always have "Low Quality + Bad Anatomy" by default. The exception is when you are making monster girls and stuff. Remove Bad Anatomy.