AI 跟练的素材生产工具,产物写入 TUICallKit-Vue3/training: - 教练动作帧:AI 设定稿 → 切帧对齐、锁脸、插帧 → 导出小程序帧与 COS 高清帧 - 示范视频:Remotion 渲染四个动作的示范视频 - 背景音乐:程序合成的原创循环音乐(npm run bgm)
66 lines
2.7 KiB
Python
66 lines
2.7 KiB
Python
"""
|
|
锁脸:同一个动作的所有帧都用第一帧(站姿)的脸
|
|
|
|
每一帧是 AI 单独画的,五官、发际线每帧都有细微差别,按十几帧每秒播放时脸会“一闪一闪”。
|
|
这里按模板匹配找到每帧的脸,把整帧水平平移让头对齐第一帧(头不再左右晃),
|
|
再把第一帧的脸(额头刘海到下巴)用羽化椭圆贴上去。头顶轮廓、马尾不动,避免盖住举过头的手臂。
|
|
"""
|
|
import cv2
|
|
import numpy as np
|
|
|
|
BG = np.array([244, 248, 246], np.float32)
|
|
|
|
|
|
def _on_bg(rgba: np.ndarray) -> np.ndarray:
|
|
f = rgba.astype(np.float32)
|
|
return f[..., :3] * (f[..., 3:4] / 255) + BG * (1 - f[..., 3:4] / 255)
|
|
|
|
|
|
def _face(kp: dict):
|
|
"""脸中心(两眼中点与鼻子之间)和眼距"""
|
|
ex = (kp["leye"][0] + kp["reye"][0]) / 2
|
|
ey = (kp["leye"][1] + kp["reye"][1]) / 2
|
|
return ex, ey + 0.6 * (kp["nose"][1] - ey), abs(kp["leye"][0] - kp["reye"][0])
|
|
|
|
|
|
def _shift_x(rgba: np.ndarray, dx: int) -> np.ndarray:
|
|
out = np.zeros_like(rgba)
|
|
if dx > 0:
|
|
out[:, dx:] = rgba[:, :-dx]
|
|
elif dx < 0:
|
|
out[:, :dx] = rgba[:, -dx:]
|
|
else:
|
|
out[:] = rgba
|
|
return out
|
|
|
|
|
|
def lock_faces(frames: list, kps: list):
|
|
"""frames:对齐后的主帧(RGBA);kps:对应关键点(画布坐标)。返回处理后的帧与平移后的关键点"""
|
|
ref = frames[0]
|
|
cx, cy, d = _face(kps[0])
|
|
h = int(round(2.2 * d)) # 模板半边长:盖住整个脸和刘海
|
|
x0, y0 = int(round(cx)) - h, int(round(cy)) - h
|
|
tpl = _on_bg(ref)[y0 : y0 + 2 * h, x0 : x0 + 2 * h]
|
|
src = ref[y0 : y0 + 2 * h, x0 : x0 + 2 * h].astype(np.float32)
|
|
# 羽化椭圆:上到刘海发际线、下到下巴,左右到脸颊
|
|
yy, xx = np.mgrid[0 : 2 * h, 0 : 2 * h].astype(np.float32)
|
|
ex, ey = cx - x0, cy - y0 - 0.3 * d
|
|
ell = (((xx - ex) / (1.35 * d)) ** 2 + ((yy - ey) / (1.75 * d)) ** 2 <= 1).astype(np.float32)
|
|
mask = cv2.GaussianBlur(ell, (0, 0), 0.2 * d)[..., None]
|
|
|
|
out_frames, out_kps = [], []
|
|
for f, kp in zip(frames, kps):
|
|
fx, fy, _ = _face(kp)
|
|
pad = 40
|
|
sx0, sy0 = int(round(fx)) - h - pad, int(round(fy)) - h - pad
|
|
win = _on_bg(f)[sy0 : sy0 + 2 * (h + pad), sx0 : sx0 + 2 * (h + pad)]
|
|
_, _, _, loc = cv2.minMaxLoc(-cv2.matchTemplate(win, tpl, cv2.TM_SQDIFF))
|
|
tx, ty = sx0 + loc[0], sy0 + loc[1]
|
|
dx = x0 - tx
|
|
g = _shift_x(f, dx).astype(np.float32)
|
|
region = g[ty : ty + 2 * h, x0 : x0 + 2 * h]
|
|
region[:] = src * mask + region * (1 - mask)
|
|
out_frames.append(np.clip(g, 0, 255).round().astype(np.uint8))
|
|
out_kps.append({k: (x + dx, y, s) for k, (x, y, s) in kp.items()})
|
|
return out_frames, out_kps
|