Stable Diffusion 照片修改（不使用webui）

mkdir sd_script
cd sd_script
poetry new .

# 我习惯把venv放在工程目录
poetry config virtualenvs.in-project true --local
poetry env use python3.10

# 一顿乱装
poetry add diffusers transformers accelerate scipy safetensors omegaconf
# 参考：https://pytorch.org/get-started/locally/
poetry source add --priority=supplemental pytorch https://download.pytorch.org/whl/rocm5.4.2
poetry add --source pytorch torch torchvision torchaudio

# 进入venv
source .venv/bin/activate.fish

from diffusers import StableDiffusionImg2ImgPipeline, DPMSolverMultistepScheduler
from PIL import Image

pipe = StableDiffusionImg2ImgPipeline.from_single_file(
    # https:civitai.com/models/7240
    "./model/model/MeinaMix.v13.safetensors"
)
pipe.scheduler = DPMSolverMultistepScheduler.from_config(
    pipe.scheduler.config, use_karras_sigmas=True
)

# rocm不配拥有名字
pipe.to("cuda")

# 明明不是R18 ...
if pipe.safety_checker is not None:
    pipe.safety_checker = None

# 天灵灵，地灵灵，神灵灵，祖宗灵灵。
prompt = "masterpiece, best quality, ultra detail,(2girl), upper body,close up, smile, happy, open eye, with glasses"
nprompt = "NSFW,(worst quality, low quality:1.4), (bad_prompt_version2:1), muscular, greyscale, monochrome, lineart, 2koma, 3koma, 4koma, manga, 3D, 3Dcubism, pablo picasso, disney, marvel, mutanted breasts, mutanted nipple, cropped, lowres, bad anatomy, bad hands, text, error, missing fingers, extra digit, fewer digits, cropped, worst quality, low quality, normal quality, jpeg artifacts, signature, watermark, username, blurry,  artist name, lowres, trademark, watermark, title, text, deformed, bad anatomy, disfigured, mutated, extra limbs, ugly, missing limb, floating limbs, disconnected limbs, out of frame, mutated hands and fingers, poorly drawn hands, malformed hands, poorly drawn face, poorly drawn asymmetrical eyes, (blurry:1.4), duplicate, (worst quality, low quality:1.4)"

init_image = Image.open('./resources/xxxxx.jpg').convert("RGB")
init_image.thumbnail((512, 512))

image = pipe(prompt=prompt, 
             negative_prompt=nprompt, 
             image=init_image, 
             strength=0.45, 
             guidance_scale=7.5,
             num_inference_steps=100
             ).images[0]
image.save("output.png")