>>53474156NTA, but in my experience, inpainting works best in small increments. This Coco gen I made earlier had some weird things I wanted to get rid of, and a missing thumb. I inpainted the things I wanted to get rid of one by one. I didn't modify my prompt from the original. For getting rid of them, I used Fill, Whole Picture, 4 blur, .65 denoise, and then the same sampler/steps/cfg as the txt2img gen. For the thumb, I just opened the image in some image editing software, crudely drew in where the thumb should be, and then put that image back into the inpaint box. Then I masked over the thumb, and inpainted with the same settings except Original instead of Fill, and a lower denoise like .45 just to blend it in. This gif shows the gen from start to finish.
For faces I usually use Only Masked instead of Original with 32-128 pixels of padding. This usually results in the face looking a little sharper than the rest of the image, but it looks better.
Inpainting to make bigger changes like inserting objects/people uses similar technique, but you would modify your prompt to just be what you want inserted.
Inpainting feels pretty critical for making good gens into great gens, so I'm surprised there aren't better guides.
