Stable Diffusion

# Stable Diffusion XL 이미지 생성 예제
from diffusers import StableDiffusionXLPipeline
import torch

# SDXL 파이프라인 로드 (약 6.5GB VRAM 필요)
pipe = StableDiffusionXLPipeline.from_pretrained(
    "stabilityai/stable-diffusion-xl-base-1.0",
    torch_dtype=torch.float16,
    use_safetensors=True,
    variant="fp16"
)
pipe.to("cuda")

# 메모리 최적화 (VRAM이 부족한 경우)
pipe.enable_model_cpu_offload()
pipe.enable_vae_slicing()

# 이미지 생성
prompt = "A majestic mountain landscape at sunset, photorealistic, 8k uhd"
negative_prompt = "blurry, low quality, distorted, artifacts"

image = pipe(
    prompt=prompt,
    negative_prompt=negative_prompt,
    num_inference_steps=30,  # 품질과 속도 트레이드오프
    guidance_scale=7.5,      # CFG scale: 높을수록 프롬프트에 충실
    width=1024,
    height=1024,
    generator=torch.Generator("cuda").manual_seed(42)
).images[0]

image.save("output.png")


# Image-to-Image 변환
from diffusers import StableDiffusionXLImg2ImgPipeline
from PIL import Image

img2img_pipe = StableDiffusionXLImg2ImgPipeline.from_pretrained(
    "stabilityai/stable-diffusion-xl-base-1.0",
    torch_dtype=torch.float16,
    use_safetensors=True
).to("cuda")

init_image = Image.open("input.png").resize((1024, 1024))

result = img2img_pipe(
    prompt="Convert to anime style, vibrant colors",
    image=init_image,
    strength=0.7,  # 0-1: 높을수록 변화가 큼
    num_inference_steps=30
).images[0]


# ControlNet으로 포즈 제어
from diffusers import ControlNetModel, StableDiffusionXLControlNetPipeline
from controlnet_aux import OpenposeDetector

controlnet = ControlNetModel.from_pretrained(
    "thibaud/controlnet-openpose-sdxl-1.0",
    torch_dtype=torch.float16
)

pipe = StableDiffusionXLControlNetPipeline.from_pretrained(
    "stabilityai/stable-diffusion-xl-base-1.0",
    controlnet=controlnet,
    torch_dtype=torch.float16
).to("cuda")

# 포즈 추출 및 조건부 생성
pose_detector = OpenposeDetector.from_pretrained("lllyasviel/ControlNet")
pose_image = pose_detector(Image.open("reference_pose.png"))

image = pipe(
    prompt="Professional dancer in elegant dress",
    image=pose_image,
    controlnet_conditioning_scale=0.8
).images[0]