John6666

/

big-love-xl4-sdxl

like 4

Hmm... This may not be the only cause, but for Pony and its derivative models, output often becomes corrupted unless clip_skip=2 is specified.

import torch
from diffusers import StableDiffusionXLPipeline
model_id = "John6666/big-love-xl4-sdxl"

pipe = StableDiffusionXLPipeline.from_pretrained(
    model_id,
    torch_dtype=torch.float16,
    device_map="auto",
)

prompt =  "score_9, score_8_up, score_7_up, source_anime, rating_safe, 1girl, short white hair, blue eyes, black cropped hoodie, full body, park, smile"

image = pipe(
    prompt=prompt,
    negative_prompt=None,
    num_inference_steps=28,
    guidance_scale=6.,
    clip_skip=2, # required by Pony V6 XL and its derivative models
).images[0]

image.save("pony.png")

Ya, tried that, still the same

If that's the case, it might be an issue with SDXL's single-pass characteristics. It's not suited for expressing human details. If you have enough VRAM, using FLUX.1 would be better, but...
It seems the only options are face restoration or ADetailer (which recognizes the face and auto-impaints only the face). https://huggingface.co/spaces/r3gm/DiffuseCraft

However, there's no standard implementation for Diffusers regarding Detailer. (DIY is assumed)😅

Edit:
Additionally, while not as reliable as ADetailer, there is also a method that applies AI upscaling after generation to supplement details.
https://huggingface.co/spaces/gokaygokay/PonyRealism

Thanks for your input. I ended up trying a few face and realism loras and got good results.