georgefen/Face-Landmark-ControlNet

1import cv22import numpy as np3import torch4 5from einops import rearrange6from .api import MiDaSInference7 8 9class MidasDetector:10    def __init__(self):11        self.model = MiDaSInference(model_type="dpt_hybrid").cuda()12 13    def __call__(self, input_image, a=np.pi * 2.0, bg_th=0.1):14        assert input_image.ndim == 315        image_depth = input_image16        with torch.no_grad():17            image_depth = torch.from_numpy(image_depth).float().cuda()18            image_depth = image_depth / 127.5 - 1.019            image_depth = rearrange(image_depth, 'h w c -> 1 c h w')20            depth = self.model(image_depth)[0]21 22            depth_pt = depth.clone()23            depth_pt -= torch.min(depth_pt)24            depth_pt /= torch.max(depth_pt)25            depth_pt = depth_pt.cpu().numpy()26            depth_image = (depth_pt * 255.0).clip(0, 255).astype(np.uint8)27 28            depth_np = depth.cpu().numpy()29            x = cv2.Sobel(depth_np, cv2.CV_32F, 1, 0, ksize=3)30            y = cv2.Sobel(depth_np, cv2.CV_32F, 0, 1, ksize=3)31            z = np.ones_like(x) * a32            x[depth_pt < bg_th] = 033            y[depth_pt < bg_th] = 034            normal = np.stack([x, y, z], axis=2)35            normal /= np.sum(normal ** 2.0, axis=2, keepdims=True) ** 0.536            normal_image = (normal * 127.5 + 127.5).clip(0, 255).astype(np.uint8)37 38            return depth_image, normal_image39