I'm also using automatic1111. On the starfire ones I used sleeplessmix, eulera-20 steps, cfg7 at 512x768. Then I sent it to inpaint, not using face restore, and inpainted the face. I set it to inpaint only masked area as well. Usually for any other areas I want to touch up, I'll do the same as I did for the face. Denoise 0.5 or maybe slightly higher for the inpainting, and then I sent that result to img2img. From there, I scale it from 512x768 to 1024x1536, doubling it. Usually that gives me really nice results. Liiike this supergirl for example was made that way, only I used comicsvision for this one. Also I wanna say that there are several models I've used that can give good results, although as you can probably tell the face style doesn't quite match. I generally prefer the different faces though, and the body style stays the same anyway so its not a huge sacrifice imo. Anyway, enough ranting, the different models I like to use are comicsvision, comicbabes can be decent sometimes, sleeplessmix, occidentalmix, and I've even had ok results with the M4RV3LSDUNGEONS model as well.Are you using automatc111 too?
For producing that I used anything 4.5 with default euler-20,cfg-7, and 512x800. Then cropped the results to 512x768, because the bottom of the legs wasn't cool.
I based my prompts/negative prompts on the Starfire example under the sunsetriders lora from in civit.ai + badhands embedding.
After that, I just 2-3 times used img2img with comic-chicks model, with the same prompt, but with 1024x1536 size, some small noise (0.2-0.4) and facefix to improve the result - worked better than realsergan or the facefix/highres in text2img for improving.
What are your approaches?
Oh one more thing, I realized I forgot to add the prompts. So they can be seen if you open the image with notepad. That will show whatever prompt was used on the last edit to the image. For mine, it will mostly be the same because one of the last steps I use is the Img2Img thing.
Last edited: