import time import torch import onnx import cv2 import onnxruntime import numpy as np from tqdm import tqdm from onnx import numpy_helper from skimage import transform as trans arcface_dst = np.array( [[38.2946, 51.6963], [73.5318, 51.5014], [56.0252, 71.7366], [41.5493, 92.3655], [70.7299, 92.2041]], dtype=np.float32) def estimate_norm(lmk, image_size=112, mode='arcface'): assert lmk.shape == (5, 2) assert image_size % 112 == 0 or image_size % 128 == 0 if image_size % 112 == 0: ratio = float(image_size) / 112.0 diff_x = 0 else: ratio = float(image_size) / 128.0 diff_x = 8.0 * ratio dst = arcface_dst * ratio dst[:, 0] += diff_x tform = trans.SimilarityTransform() tform.estimate(lmk, dst) M = tform.params[0:2, :] return M def norm_crop2(img, landmark, image_size=112, mode='arcface'): M = estimate_norm(landmark, image_size, mode) warped = cv2.warpAffine(img, M, (image_size, image_size), borderValue=0.0) return warped, M class Inswapper(): def __init__(self, model_file=None, batch_size=32, providers=['CPUExecutionProvider']): self.model_file = model_file self.batch_size = batch_size model = onnx.load(self.model_file) graph = model.graph self.emap = numpy_helper.to_array(graph.initializer[-1]) self.input_mean = 0.0 self.input_std = 255.0 self.session_options = onnxruntime.SessionOptions() self.session = onnxruntime.InferenceSession(self.model_file, sess_options=self.session_options, providers=providers) inputs = self.session.get_inputs() self.input_names = [inp.name for inp in inputs] outputs = self.session.get_outputs() self.output_names = [out.name for out in outputs] assert len(self.output_names) == 1 self.output_shape = outputs[0].shape input_cfg = inputs[0] input_shape = input_cfg.shape self.input_shape = input_shape self.input_size = tuple(input_shape[2:4][::-1]) def forward(self, imgs, latents): batch_preds = [] for img, latent in zip(imgs, latents): img = (img - self.input_mean) / self.input_std pred = self.session.run(self.output_names, {self.input_names[0]: img, self.input_names[1]: latent})[0] batch_preds.append(pred) return batch_preds def get(self, imgs, target_faces, source_faces): batch_preds = [] batch_aimgs = [] batch_ms = [] for img, target_face, source_face in zip(imgs, target_faces, source_faces): if isinstance(img, str): img = cv2.imread(img) aimg, M = norm_crop2(img, target_face.kps, self.input_size[0]) blob = cv2.dnn.blobFromImage(aimg, 1.0 / self.input_std, self.input_size, (self.input_mean, self.input_mean, self.input_mean), swapRB=True) latent = source_face.normed_embedding.reshape((1, -1)) latent = np.dot(latent, self.emap) latent /= np.linalg.norm(latent) pred = self.session.run(self.output_names, {self.input_names[0]: blob, self.input_names[1]: latent})[0] pred = pred.transpose((0, 2, 3, 1))[0] pred = np.clip(255 * pred, 0, 255).astype(np.uint8)[:, :, ::-1] batch_preds.append(pred) batch_aimgs.append(aimg) batch_ms.append(M) return batch_preds, batch_aimgs, batch_ms def batch_forward(self, img_list, target_f_list, source_f_list): num_samples = len(img_list) num_batches = (num_samples + self.batch_size - 1) // self.batch_size preds = [] aimgs = [] ms = [] for i in tqdm(range(num_batches), desc="Swapping face by batch"): start_idx = i * self.batch_size end_idx = min((i + 1) * self.batch_size, num_samples) batch_img = img_list[start_idx:end_idx] batch_target_f = target_f_list[start_idx:end_idx] batch_source_f = source_f_list[start_idx:end_idx] batch_pred, batch_aimg, batch_m = self.get(batch_img, batch_target_f, batch_source_f) preds.extend(batch_pred) aimgs.extend(batch_aimg) ms.extend(batch_m) return preds, aimgs, ms def laplacian_blending(A, B, m, num_levels=4): assert A.shape == B.shape assert B.shape == m.shape height = m.shape[0] width = m.shape[1] size_list = np.array([4, 8, 16, 32, 64, 128, 256, 512, 1024, 2048, 4096]) size = size_list[np.where(size_list > max(height, width))][0] GA = np.zeros((size, size, 3), dtype=np.float32) GA[:height, :width, :] = A GB = np.zeros((size, size, 3), dtype=np.float32) GB[:height, :width, :] = B GM = np.zeros((size, size, 3), dtype=np.float32) GM[:height, :width, :] = m gpA = [GA] gpB = [GB] gpM = [GM] for i in range(num_levels): GA = cv2.pyrDown(GA) GB = cv2.pyrDown(GB) GM = cv2.pyrDown(GM) gpA.append(np.float32(GA)) gpB.append(np.float32(GB)) gpM.append(np.float32(GM)) lpA = [gpA[num_levels-1]] lpB = [gpB[num_levels-1]] gpMr = [gpM[num_levels-1]] for i in range(num_levels-1,0,-1): LA = np.subtract(gpA[i-1], cv2.pyrUp(gpA[i])) LB = np.subtract(gpB[i-1], cv2.pyrUp(gpB[i])) lpA.append(LA) lpB.append(LB) gpMr.append(gpM[i-1]) LS = [] for la,lb,gm in zip(lpA,lpB,gpMr): ls = la * gm + lb * (1.0 - gm) LS.append(ls) ls_ = LS[0] for i in range(1,num_levels): ls_ = cv2.pyrUp(ls_) ls_ = cv2.add(ls_, LS[i]) ls_ = np.clip(ls_[:height, :width, :], 0, 255) return ls_ def paste_to_whole(bgr_fake, aimg, M, whole_img, laplacian_blend=True, crop_mask=(0,0,0,0)): IM = cv2.invertAffineTransform(M) img_white = np.full((aimg.shape[0], aimg.shape[1]), 255, dtype=np.float32) top = int(crop_mask[0]) bottom = int(crop_mask[1]) if top + bottom < aimg.shape[1]: if top > 0: img_white[:top, :] = 0 if bottom > 0: img_white[-bottom:, :] = 0 left = int(crop_mask[2]) right = int(crop_mask[3]) if left + right < aimg.shape[0]: if left > 0: img_white[:, :left] = 0 if right > 0: img_white[:, -right:] = 0 bgr_fake = cv2.warpAffine( bgr_fake, IM, (whole_img.shape[1], whole_img.shape[0]), borderValue=0.0 ) img_white = cv2.warpAffine( img_white, IM, (whole_img.shape[1], whole_img.shape[0]), borderValue=0.0 ) img_white[img_white > 20] = 255 img_mask = img_white mask_h_inds, mask_w_inds = np.where(img_mask == 255) mask_h = np.max(mask_h_inds) - np.min(mask_h_inds) mask_w = np.max(mask_w_inds) - np.min(mask_w_inds) mask_size = int(np.sqrt(mask_h * mask_w)) k = max(mask_size // 10, 10) img_mask = cv2.erode(img_mask, np.ones((k, k), np.uint8), iterations=1) k = max(mask_size // 20, 5) kernel_size = (k, k) blur_size = tuple(2 * i + 1 for i in kernel_size) img_mask = cv2.GaussianBlur(img_mask, blur_size, 0) / 255 img_mask = np.tile(np.expand_dims(img_mask, axis=-1), (1, 1, 3)) if laplacian_blend: bgr_fake = laplacian_blending(bgr_fake.astype("float32").clip(0,255), whole_img.astype("float32").clip(0,255), img_mask.clip(0,1)) bgr_fake = bgr_fake.astype("float32") fake_merged = img_mask * bgr_fake + (1 - img_mask) * whole_img.astype(np.float32) return fake_merged.astype("uint8")