AI 置景 + 合成精修工程化：从单图生成到多工具协同的完整管线

import torch
from diffusers import StableDiffusionControlNetPipeline, ControlNetModel
from diffusers.utils import load_image
from controlnet_aux import MidasDetector

controlnet = ControlNetModel.from_pretrained(
    "lllyasviel/control_v11f1p_sd15_depth",
    torch_dtype=torch.float16)

pipe = StableDiffusionControlNetPipeline.from_pretrained(
    "runwayml/stable-diffusion-v1-5",
    controlnet=controlnet,
    torch_dtype=torch.float16).to("cuda")

midas = MidasDetector.from_pretrained("lllyasviel/Annotators")

def build_scene(ref_image: str, prompt: str, seed: int = 42) -> tuple:
    """从参考图提取深度，生成透视一致的背景"""
    src = load_image(ref_image)
    depth = midas(src)                     # 提取深度图
    generator = torch.Generator("cuda").manual_seed(seed)
    out = pipe(
        prompt=prompt,
        image=depth,
        num_inference_steps=30,
        guidance_scale=7.5,
        controlnet_conditioning_scale=0.9,  # 深度约束强度
        generator=generator).images[0]
    return out, depth

import cv2
import numpy as np

def align_perspective(subject: np.ndarray, bg: np.ndarray,
                      src_pts: np.ndarray, dst_pts: np.ndarray) -> np.ndarray:
    """把主体透视变换到背景的相机视角"""
    h, w = bg.shape[:2]
    M = cv2.getPerspectiveTransform(src_pts, dst_pts)
    warped = cv2.warpPerspective(subject, M, (w, h))
    return warped

def composite(subject: np.ndarray, bg: np.ndarray,
              mask: np.ndarray, blur: int = 3) -> np.ndarray:
    """用羽化蒙版做无缝合成"""
    mask = cv2.GaussianBlur(mask, (blur * 2 + 1,) * 2, 0)
    mask = mask.astype(np.float32) / 255.0
    mask = np.stack([mask] * 3, axis=-1)
    result = subject * mask + bg * (1 - mask)
    return result.astype(np.uint8)

def relight(subject: np.ndarray, light_dir: tuple = (-1, -1),
            intensity: float = 0.3) -> np.ndarray:
    """根据光照方向给主体叠加方向性光影"""
    h, w = subject.shape[:2]
    yy, xx = np.mgrid[0:h, 0:w]
    lx, ly = light_dir
    grad = (xx * lx + yy * ly).astype(np.float32)
    grad = (grad - grad.min()) / (grad.max() - grad.min() + 1e-6)
    shade = 1.0 - intensity * (1.0 - grad)
    shade = np.stack([shade] * 3, axis=-1)
    return np.clip(subject.astype(np.float32) * shade, 0, 255).astype(np.uint8)

def frequency_separate(img: np.ndarray, sigma: int = 15) -> tuple:
    """频率分离：低频管光影，高频管质感"""
    low = cv2.GaussianBlur(img, (0, 0), sigma)
    high = cv2.subtract(img.astype(np.int16),
                        low.astype(np.int16)) + 128
    return low, high.astype(np.uint8)

def pipeline(subject_path: str, ref_scene: str,
             prompt: str, light_dir: tuple) -> dict:
    subject = cv2.imread(subject_path)
    bg, depth = build_scene(ref_scene, prompt)
    bg_np = np.array(bg)

    mask = extract_mask(subject)          # 主体蒙版
    aligned = align_perspective(subject, bg_np, src_pts, dst_pts)
    composed = composite(aligned, bg_np, mask)
    final = relight(composed, light_dir=light_dir, intensity=0.3)

    low, high = frequency_separate(final)
    return {"final": final, "low": low, "high": high,
            "depth": depth}