from __future__ import annotations from pathlib import Path from typing import Iterable, Mapping, Sequence import cv2 import numpy as np import os Operation = Mapping[str, object] class SimilarityMLS: def __init__(self, grid_size=50, alpha=1.0): self.grid_size = grid_size self.alpha = alpha def _calculate_weights(self, grid_pts, ctrl_pts): n_grid = grid_pts.shape[0] n_ctrl = ctrl_pts.shape[0] grid_pts_exp = np.tile(grid_pts[:, np.newaxis, :], (1, n_ctrl, 1)) ctrl_pts_exp = np.tile(ctrl_pts[np.newaxis, :, :], (n_grid, 1, 1)) d2 = np.sum((grid_pts_exp - ctrl_pts_exp) ** 2, axis=2) + 1e-8 return 1.0 / (d2 ** self.alpha) def warp(self, img, src_pts, dst_pts): h, w = img.shape[:2] grid_x = np.linspace(0, w, w // self.grid_size + 1) grid_y = np.linspace(0, h, h // self.grid_size + 1) grid_x, grid_y = np.meshgrid(grid_x, grid_y) grid_pts = np.vstack([grid_x.ravel(), grid_y.ravel()]).T n_grid = grid_pts.shape[0] weights = self._calculate_weights(grid_pts, src_pts) total_weights = np.sum(weights, axis=1, keepdims=True) p_star = (weights @ src_pts) / total_weights q_star = (weights @ dst_pts) / total_weights p_hat = src_pts[np.newaxis, :, :] - p_star[:, np.newaxis, :] q_hat = dst_pts[np.newaxis, :, :] - q_star[:, np.newaxis, :] mu = np.sum(weights[:, :, np.newaxis] * p_hat ** 2, axis=(1, 2), keepdims=True) v_hat = grid_pts - p_star p_hat_perp = np.stack([-p_hat[:, :, 1], p_hat[:, :, 0]], axis=2) vp_dot_phat = np.sum(v_hat[:, np.newaxis, :] * p_hat, axis=2) vp_dot_phat_perp = np.sum(v_hat[:, np.newaxis, :] * p_hat_perp, axis=2) T1 = vp_dot_phat[:, :, np.newaxis] * q_hat T2 = vp_dot_phat_perp[:, :, np.newaxis] * np.stack([-q_hat[:, :, 1], q_hat[:, :, 0]], axis=2) weighted_sum = np.sum(weights[:, :, np.newaxis] * (T1 + T2), axis=1) result_grid = q_star + weighted_sum / mu.squeeze(-1) map_x = cv2.resize(result_grid[:, 0].reshape(grid_x.shape).astype(np.float32), (w, h)) map_y = cv2.resize(result_grid[:, 1].reshape(grid_y.shape).astype(np.float32), (w, h)) return cv2.remap(img, map_x, map_y, interpolation=cv2.INTER_LINEAR) class FaceEditor: def __init__(self): try: import mediapipe as mp except ImportError as exc: raise ImportError("LLW face retouching requires mediapipe. Install the mirrorppr data extra or run the provided environment setup.") from exc self.mp_face_mesh = mp.solutions.face_mesh self.face_mesh = self.mp_face_mesh.FaceMesh( static_image_mode=True, max_num_faces=1, refine_landmarks=True, min_detection_confidence=0.5 ) def _get_landmarks(self, image): h, w = image.shape[:2] results = self.face_mesh.process(cv2.cvtColor(image, cv2.COLOR_BGR2RGB)) if not results.multi_face_landmarks: return None landmarks = results.multi_face_landmarks[0].landmark points = np.array([(int(l.x * w), int(l.y * h)) for l in landmarks]) return points def _create_roi_mask(self, h, w, points, regions_indices, blur_ratio): """ regions_indices: List[List[int]], 例如 [[左眼索引...], [右眼索引...]] """ mask = np.zeros((h, w), dtype=np.uint8) for indices in regions_indices: if not indices: continue roi_points = points[indices].astype(np.int32) hull = cv2.convexHull(roi_points) cv2.fillConvexPoly(mask, hull, 255) blur_k = int(min(h, w) * blur_ratio) if blur_k % 2 == 0: blur_k += 1 mask = cv2.GaussianBlur(mask, (blur_k, blur_k), 0) return mask.astype(np.float32) / 255.0 def _get_default_params(self, op_type): """ 获取操作的默认参数，方便在 process_batch 中合并 """ defaults = { 'grid_size': 30, 'alpha': 1.0, 'blur_ratio': 0.05 } return defaults def _get_operation_config(self, op_type, landmarks, strength, params): """ 根据操作类型分发配置逻辑 """ if op_type == 'eye_resize': return self._config_eye_resize(landmarks, strength, params) elif op_type == 'eye_distance': return self._config_eye_distance(landmarks, strength, params) elif op_type == 'nose_length': return self._config_nose_length(landmarks, strength, params) elif op_type == 'nose_alar': return self._config_nose_alar(landmarks, strength, params) elif op_type == 'mouth_position': return self._config_mouth_position(landmarks, strength, params) elif op_type == 'lip_thickness': return self._config_lip_thickness(landmarks, strength, params) elif op_type == 'mouth_resize': return self._config_mouth_resize(landmarks, strength, params) elif op_type == 'nose_bridge': return self._config_nose_bridge(landmarks, strength, params) else: raise ValueError(f"Unknown operation: {op_type}") def _config_nose_length(self, landmarks, strength, params): """ 鼻子变短/变长配置 Strength < 0: 变短 (Shorten) - 鼻头上移，人中变长 Strength > 0: 变长 (Lengthen) - 鼻头下移，人中变短 """ moving_indices = [ 1, 4, 19, 279, 49, 2, 98, 327, 456, 236, 94 ] anchor_idx = [ 168, 6, 197, 195, 33, 133, 362, 263, 185,40,39,37,0,267,269,270,409, 116, 123, 345, 352 ] mask_idx_nose_long = [47, 128, 142, 164, 165, 167, 168, 193, 203, 244, 277, 357, 371, 391, 393, 417, 423, 464] mask_groups = [mask_idx_nose_long] src_pts = [] dst_pts = [] nose_vec_ref = landmarks[1] - landmarks[168] nose_length = np.linalg.norm(nose_vec_ref) if strength >= 0: factor = params.get('max_nose_len_widen_ratio', 0.15) else: factor = params.get('max_nose_len_shorten_ratio', 0.15) move_dist = nose_length * (abs(strength) / 100.0) * factor vec_axis = (landmarks[1] - landmarks[168]).astype(np.float32) norm_axis = np.linalg.norm(vec_axis) if norm_axis > 0: vec_axis /= norm_axis if strength >= 0: final_vec = vec_axis * move_dist else: final_vec = -vec_axis * move_dist for idx in moving_indices: pt = landmarks[idx] src_pts.append(pt) dst_pts.append(pt + final_vec) for idx in anchor_idx: pt = landmarks[idx] src_pts.append(pt) dst_pts.append(pt) return np.array(src_pts, dtype=np.float32), np.array(dst_pts, dtype=np.float32), mask_groups def _config_nose_alar(self, landmarks, strength, params): """ 鼻翼变窄/变宽配置 Strength < 0: 变窄 (Narrow) - 鼻翼向内收 Strength > 0: 变宽 (Widen) """ left_alar_indices = [49, 102, 64, 218, 129] right_alar_indices = [279, 331, 294, 438, 358] anchor_idx = [ 1, 2, 94, 19, 168, 6, 197, 195, 4, 0, 37, 267, 205, 50, 123, 116, 425, 280, 352, 345 ] mask_idx_alar = [ 195, 4, 279, 425, 331, 294, 327, 2, 94, 98, 64, 102, 205, 49 ] mask_groups = [mask_idx_alar] src_pts = [] dst_pts = [] alar_width = np.linalg.norm(landmarks[331] - landmarks[102]) if strength >= 0: factor = params.get('max_alar_widen_ratio', 0.25) scale = 1.0 + (strength / 100.0) * factor else: factor = params.get('max_alar_narrow_ratio', 0.25) scale = 1.0 - (abs(strength) / 100.0) * factor vec_center_line = (landmarks[2] - landmarks[168]).astype(np.float32) norm_cl = np.linalg.norm(vec_center_line) if norm_cl > 0: vec_center_line /= norm_cl def get_projection_point(pt, line_start, line_vec): vec_ap = (pt - line_start).astype(np.float32) proj_len = np.dot(vec_ap, line_vec) return line_start + line_vec * proj_len all_moving_indices = left_alar_indices + right_alar_indices for idx in all_moving_indices: pt = landmarks[idx] src_pts.append(pt) proj_pt = get_projection_point(pt, landmarks[168], vec_center_line) vec_radial = pt - proj_pt dst_pts.append(proj_pt + vec_radial * scale) for idx in anchor_idx: pt = landmarks[idx] src_pts.append(pt) dst_pts.append(pt) return np.array(src_pts, dtype=np.float32), np.array(dst_pts, dtype=np.float32), mask_groups def _config_eye_resize(self, landmarks, strength, params): """ 眼睛放大/缩小的具体配置 """ left_eye_idx = [249, 263, 362, 373, 374, 380, 381, 382, 384, 385, 386, 387, 388, 390, 398, 466] right_eye_idx = [7, 33, 133, 144, 145, 153, 154, 155, 157, 158, 159, 160, 161, 163, 173, 246] anchor_idx = [1, 168, 197, 6, 195, 4, 226, 446, 152, 10, 50, 280] mask_idx_right =[22, 23, 24, 25, 26, 27, 28, 29, 30, 56, 110, 112, 130, 190, 243, 247] mask_idx_left = [252, 253, 254, 255, 256, 257, 258, 259, 260, 286, 339, 341, 359, 414, 463, 467] mask_groups = [mask_idx_right, mask_idx_left] src_pts = [] dst_pts = [] if strength >= 0: factor = params.get('max_enlarge', 0.25) scale = 1.0 + (strength / 100.0) * factor else: factor = params.get('max_shrink', 0.20) scale = 1.0 + (strength / 100.0) * factor def process_region(indices): pts = landmarks[indices] center = np.mean(pts, axis=0) for pt in pts: src_pts.append(pt) vec = pt - center dst_pts.append(center + vec * scale) process_region(left_eye_idx) process_region(right_eye_idx) for idx in anchor_idx: pt = landmarks[idx] src_pts.append(pt) dst_pts.append(pt) return np.array(src_pts, dtype=np.float32), np.array(dst_pts, dtype=np.float32), mask_groups def _config_eye_distance(self, landmarks, strength, params): """ 眼距调整 (修正版：只平移，不缩放) Strength > 0: 眼距变宽 Strength < 0: 眼距变窄 """ right_moving_idx = [22, 23, 24, 25, 26, 27, 28, 29, 30, 56, 110, 112, 130, 190, 243, 247] left_moving_idx = [252, 253, 254, 255, 256, 257, 258, 259, 260, 286, 339, 341, 359, 414, 463, 467] anchor_idx = [ 1, 2, 98, 327, 10, 152, 234, 454, 13, 14, 78, 308 ] mask_idx_right = [31, 113, 189, 221, 222, 223, 224, 225, 226, 228, 229, 230, 231, 232, 233, 244] mask_idx_left=[261, 342, 413, 441, 442, 443, 444, 445, 446, 448, 449, 450, 451, 452, 453, 464] mask_groups = [mask_idx_right, mask_idx_left] src_pts = [] dst_pts = [] center_idx = 168 center_pt = landmarks[center_idx] if strength >= 0: factor = params.get('max_dist_widen', 0.15) scale = 1.0 + (strength / 100.0) * factor else: factor = params.get('max_dist_narrow', 0.15) scale = 1.0 - (abs(strength) / 100.0) * factor r_pts = landmarks[right_moving_idx] r_centroid = np.mean(r_pts, axis=0) r_vec = r_centroid - center_pt r_centroid_new = center_pt + r_vec * scale r_translation = r_centroid_new - r_centroid for idx in right_moving_idx: pt = landmarks[idx] src_pts.append(pt) dst_pts.append(pt + r_translation) l_pts = landmarks[left_moving_idx] l_centroid = np.mean(l_pts, axis=0) l_vec = l_centroid - center_pt l_centroid_new = center_pt + l_vec * scale l_translation = l_centroid_new - l_centroid for idx in left_moving_idx: pt = landmarks[idx] src_pts.append(pt) dst_pts.append(pt + l_translation) for idx in anchor_idx: pt = landmarks[idx] src_pts.append(pt) dst_pts.append(pt) src_pts.append(center_pt) dst_pts.append(center_pt) return np.array(src_pts, dtype=np.float32), np.array(dst_pts, dtype=np.float32), mask_groups def _config_mouth_position(self, landmarks, strength, params): """ 嘴巴上下移动配置 (基于解剖学距离限制) Strength < 0: 上移 (最大幅度参照 人中长度) Strength > 0: 下移 (最大幅度参照 下唇窝距离) """ lips_indices = [0, 13, 14, 17, 37, 39, 40, 61, 78, 80, 81, 82, 84, 87, 88, 91, 95, 146, 178, 181, 185, 191, 267, 269, 270, 291, 308, 310, 311, 312, 314, 317, 318, 321, 324, 375, 402, 405, 409, 415] anchor_idx = [ 2, 98, 327, 94, 19, 1, 152, 377, 148, 365, 136, 234, 454, 58, 288, 361, 132, 93, 323, 45, 275 ] mask_idx_mouth_zone = [18, 43, 57, 83, 92, 106, 164, 165, 167, 182, 186, 273, 287, 313, 322, 335, 391, 393, 406, 410] mask_groups = [mask_idx_mouth_zone] src_pts = [] dst_pts = [] pt_0 = landmarks[0] pt_164 = landmarks[164] pt_17 = landmarks[17] pt_18 = landmarks[18] dist_up_limit = float(abs(pt_0[1] - pt_164[1])) dist_down_limit = float(abs(pt_17[1] - pt_18[1])) move_y = 0.0 if strength < 0: factor = params.get('limit_mouth_up_factor', 0.8) move_dist = dist_up_limit * (abs(strength) / 100.0) * factor move_y = -move_dist else: factor = params.get('limit_mouth_down_factor', 0.8) move_dist = dist_down_limit * (strength / 100.0) * factor move_y = move_dist translation = np.array([0, move_y], dtype=np.float32) for idx in lips_indices: pt = landmarks[idx] src_pts.append(pt) dst_pts.append(pt + translation) for idx in anchor_idx: pt = landmarks[idx] src_pts.append(pt) dst_pts.append(pt) return np.array(src_pts, dtype=np.float32), np.array(dst_pts, dtype=np.float32), mask_groups def _config_lip_thickness(self, landmarks, strength, params): """ 嘴唇变厚/变薄配置 (严格解剖学版：上下唇独立参照各自厚度) """ upper_lip_outer_moving = [185,40,39,37,0,267,269,270,409] lower_lip_outer_moving = [146, 91, 181, 84, 17, 314, 405, 321, 375] anchor_idx = [191, 80, 81, 82, 13, 312, 311, 310, 415, 14, 87, 88, 95, 178, 317, 318, 324, 402, 61, 291, 2, 98, 327, 18, 200, 234, 454, 58, 288, 361, 132, 93, 323] mask_idx_expanded =[18, 43, 57, 83, 92, 106, 164, 165, 167, 182, 186, 273, 287, 313, 322, 335, 391, 393, 406, 410] mask_groups = [mask_idx_expanded] src_pts = [] dst_pts = [] upper_thickness = np.linalg.norm(landmarks[13] - landmarks[0]) lower_thickness = np.linalg.norm(landmarks[14] - landmarks[17]) thickness = max(upper_thickness,lower_thickness) if strength >= 0: factor = params.get('max_lip_thicken_ratio', 0.4) else: factor = params.get('max_lip_thin_ratio', 0.3) move_dist_upper = thickness * (abs(strength) / 100.0) * factor move_dist_lower = thickness * (abs(strength) / 100.0) * factor vec_up = (landmarks[2] - landmarks[13]).astype(np.float32) norm_up = np.linalg.norm(vec_up) if norm_up > 0: vec_up /= norm_up vec_down = (landmarks[152] - landmarks[14]).astype(np.float32) norm_down = np.linalg.norm(vec_down) if norm_down > 0: vec_down /= norm_down if strength >= 0: final_vec_upper = vec_up * move_dist_upper final_vec_lower = vec_down * move_dist_lower else: final_vec_upper = -vec_up * move_dist_upper final_vec_lower = -vec_down * move_dist_lower for idx in upper_lip_outer_moving: pt = landmarks[idx] src_pts.append(pt) dst_pts.append(pt + final_vec_upper) for idx in lower_lip_outer_moving: pt = landmarks[idx] src_pts.append(pt) dst_pts.append(pt + final_vec_lower) for idx in anchor_idx: pt = landmarks[idx] src_pts.append(pt) dst_pts.append(pt) return np.array(src_pts, dtype=np.float32), np.array(dst_pts, dtype=np.float32), mask_groups def _config_mouth_resize(self, landmarks, strength, params): """ 嘴巴整体缩放配置 Strength > 0: 变大 (Enlarge) Strength < 0: 变小 (Shrink) """ lips_indices = [0, 13, 14, 17, 37, 39, 40, 61, 78, 80, 81, 82, 84, 87, 88, 91, 95, 146, 178, 181, 185, 191, 267, 269, 270, 291, 308, 310, 311, 312, 314, 317, 318, 321, 324, 375, 402, 405, 409, 415] anchor_idx = [ 2, 98, 327, 205, 425, 152, 377, 148, 234, 454, 58, 288, 361, 132, 93, 323, 164 ] mask_idx_expanded = [18, 43, 57, 83, 92, 106, 164, 165, 167, 182, 186, 273, 287, 313, 322, 335, 391, 393, 406, 410] mask_groups = [mask_idx_expanded] src_pts = [] dst_pts = [] mouth_pts = landmarks[lips_indices] center = np.mean(mouth_pts, axis=0) if strength >= 0: factor = params.get('max_mouth_enlarge', 0.25) scale = 1.0 + (strength / 100.0) * factor else: factor = params.get('max_mouth_shrink', 0.25) scale = 1.0 - (abs(strength) / 100.0) * factor for idx in lips_indices: pt = landmarks[idx] src_pts.append(pt) vec = pt - center dst_pts.append(center + vec * scale) for idx in anchor_idx: pt = landmarks[idx] src_pts.append(pt) dst_pts.append(pt) src_pts.append(center) dst_pts.append(center) return np.array(src_pts, dtype=np.float32), np.array(dst_pts, dtype=np.float32), mask_groups def _config_nose_bridge(self, landmarks, strength, params): """ 鼻梁变窄/变宽配置 Strength < 0: 变窄 (Narrow) - 更加立体/精致 Strength > 0: 变宽 (Widen) """ left_bridge_indices = [ 193, 245, 128, 122, 121, 100 ] right_bridge_indices = [ 417, 465, 357, 351, 350, 329 ] anchor_idx = [ 1, 2, 4, 5, 6, 19, 48, 64, 94, 98, 115, 168, 195, 197, 278, 294, 327, 344, 362, 133, 359, 130, 123, 50, 116, 352, 280, 345 ] mask_idx_bridge = [9, 55, 97, 134, 164, 174, 188, 193, 220, 236, 237, 245, 285, 326, 363, 399, 412, 417, 440, 456, 457, 465] mask_groups = [mask_idx_bridge] src_pts = [] dst_pts = [] if strength >= 0: factor = params.get('max_nose_widen_ratio', 0.3) scale = 1.0 + (strength / 100.0) * factor else: factor = params.get('max_nose_narrow_ratio', 0.25) scale = 1.0 - (abs(strength) / 100.0) * factor vec_center_line = (landmarks[2] - landmarks[168]).astype(np.float32) norm_cl = np.linalg.norm(vec_center_line) if norm_cl > 0: vec_center_line /= norm_cl def get_projection_point(pt, line_start, line_vec): vec_ap = (pt - line_start).astype(np.float32) proj_len = np.dot(vec_ap, line_vec) return line_start + line_vec * proj_len all_moving_indices = [3, 44, 45, 51, 122, 193, 196, 248, 274, 275, 281, 351, 417, 419] for idx in all_moving_indices: pt = landmarks[idx] src_pts.append(pt) proj_pt = get_projection_point(pt, landmarks[168], vec_center_line) vec_radial = pt - proj_pt dst_pts.append(proj_pt + vec_radial * scale) for idx in anchor_idx: pt = landmarks[idx] src_pts.append(pt) dst_pts.append(pt) return np.array(src_pts, dtype=np.float32), np.array(dst_pts, dtype=np.float32), mask_groups def apply_deformation(self, image, src_pts, dst_pts, mask_indices, landmarks, params): h, w = image.shape[:2] mls = SimilarityMLS( grid_size=params.get('grid_size', 50), alpha=params.get('alpha', 1.0) ) warped_image = mls.warp(image, dst_pts, src_pts) mask = self._create_roi_mask( h, w, landmarks, mask_indices, blur_ratio=params.get('blur_ratio', 0.08) ) mask = mask[:, :, np.newaxis] result = warped_image * mask + image * (1.0 - mask) return result.astype(np.uint8) def process_and_save(self, image_path, output_path, op_type='eye_resize', strength=50, hyperparams=None): """ 单次操作的便捷入口，内部调用 process_batch """ operation = { 'op_type': op_type, 'strength': strength, 'params': hyperparams if hyperparams else {} } self.process_batch(image_path, output_path, [operation]) def process_batch(self, image_path, output_path, operations, diff_output_path=None): """ 批量执行多个编辑操作 (Pipeline模式) :param operations: 操作列表 :param diff_output_path: (新增) 指定差异图的保存路径，如果为None则不保存或使用默认命名 """ img = cv2.imread(image_path) if img is None: print(f"Error: Could not read image: {image_path}") return arr_source = img.astype(np.float32) current_img = img.copy() for i, op in enumerate(operations): op_type = op.get('op_type') strength = op.get('strength') custom_params = op.get('params', {}) landmarks = self._get_landmarks(current_img) if landmarks is None: break params = self._get_default_params(op_type) if custom_params: params.update(custom_params) try: src_pts, dst_pts, mask_groups = self._get_operation_config( op_type, landmarks, strength, params ) current_img = self.apply_deformation( current_img, src_pts, dst_pts, mask_groups, landmarks, params ) except ValueError as e: continue os.makedirs(os.path.dirname(output_path), exist_ok=True) cv2.imwrite(output_path, current_img) class LLWFaceRetoucher(FaceEditor): """Public wrapper for Landmark-Guided Local Warping face retouching.""" def apply_operations_to_array(self, image_bgr: np.ndarray, operations: Sequence[Operation]) -> np.ndarray: current_img = image_bgr.copy() for op in operations: op_type = str(op.get("op_type") or op.get("operation") or op.get("name")) strength = float(op.get("strength", op.get("value", 100))) custom_params = op.get("params", {}) or {} landmarks = self._get_landmarks(current_img) if landmarks is None: raise RuntimeError("No face landmarks were detected before applying operation: " + op_type) params = self._get_default_params(op_type) params.update(custom_params) src_pts, dst_pts, mask_groups = self._get_operation_config(op_type, landmarks, strength, params) current_img = self.apply_deformation(current_img, src_pts, dst_pts, mask_groups, landmarks, params) return current_img def apply_operations(self, image_path: str | Path, operations: Sequence[Operation], output_path: str | Path | None = None) -> np.ndarray: image = cv2.imread(str(image_path)) if image is None: raise FileNotFoundError(f"Could not read image: {image_path}") result = self.apply_operations_to_array(image, operations) if output_path is not None: output_path = Path(output_path) output_path.parent.mkdir(parents=True, exist_ok=True) cv2.imwrite(str(output_path), result) return result def normalize_operations(operation_specs: Iterable[str]) -> list[dict[str, object]]: operations = [] for spec in operation_specs: if ":" not in spec: raise ValueError(f"Operation must be formatted as name:strength, got {spec!r}") name, value = spec.split(":", 1) operations.append({"op_type": name.strip(), "strength": float(value)}) return operations