>>20055989
Make sure you clean up the image first, I can tell it still had the speech bubbles which won't help. CFG was 7,5, steps were 45 but some of the ones I've posted here were 30, sampler was Euler a. Strength varies depending on picture but I never go above 0.8. Keep SDXL resolutions in mind as well, 1024x1024, 1152x896, 1216x832, etcetera. Make sure the picture you're throwing it fits into those proportions. I only used a LORA of the character at 0.5 with a suitable prompt:
seras victoria, short hair, blonde hair, blue eyes, military uniform, emblem, belt, yellow skirt, pencil skirt, (all of those came from the LORA), huge breasts, (bursting breasts:1.4), looking at viewer, confused, arms behind back, parted lips, wide hips, no bra,
Keep in mind the checkpoint's page has a suggested set of prompts and negatives, make sure those are included. The face in yours looks bad and that's normal, it happens with mine all the time, you'll have to inpaint over the face (also other parts of the image that might seem off or missing details) with the same prompt at medium strength (0.5 or so) and it should look better. Another thing you can do is photoshop things from the original back into the realified image and inpaint over them, here's an example of a middle of the road image I had.
>>20056076
Even the best models (including closed source ones) are still kinda bad with consistency and details. I think we'll get there eventually though.
