Files
long 3cd3553af2 feat: 新增 training-videos 素材流水线
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)
2026-10-31 10:33:25 +08:00

79 lines
3.1 KiB
Python
Raw Permalink Blame History

This file contains ambiguous Unicode characters
This file contains Unicode characters that might be confused with other characters. If you think that this is intentional, you can safely ignore this warning. Use the Escape button to reveal them.
"""
镜像换边:左侧动作帧水平翻转得到右侧动作帧,头部(脸 + 马尾)保留原图方向。
左右两侧出自同一张设定稿(同一次渲染),换边时脸、衣服都不会“闪”;
只翻转身体,所以马尾始终在同一侧,不会左右跳。
"""
import cv2
import numpy as np
ALPHA_ON = 40
def _premul(rgba):
f = rgba.astype(np.float32) / 255
f[..., :3] *= f[..., 3:4]
return f
def _unpremul(f):
a = np.maximum(f[..., 3:4], 1e-6)
out = f.copy()
out[..., :3] = np.where(f[..., 3:4] > 0, f[..., :3] / a, 0)
return (np.clip(out, 0, 1) * 255).round().astype(np.uint8)
def _mirror(arr, xc):
"""以 x = xc 为轴水平翻转(整像素平移,不重采样)"""
w = arr.shape[1]
flipped = arr[:, ::-1]
shift = int(round(2 * xc - (w - 1)))
out = np.zeros_like(arr)
if shift >= 0:
out[:, shift:] = flipped[:, : w - shift]
else:
out[:, : w + shift] = flipped[:, -shift:]
return out
def _head_mask(rgba, kp, y_cut):
"""肩线以上、包含鼻子的连通区域 = 头 + 脖子 + 马尾上段"""
on = rgba[..., 3] > ALPHA_ON
on[int(y_cut) :] = False
_, lab = cv2.connectedComponents(on.astype(np.uint8), connectivity=8)
k = lab[int(round(kp["nose"][1])), int(round(kp["nose"][0]))]
if k == 0:
raise ValueError("鼻子不在人物区域内")
return lab == k
def _hair_mask(rgba, head, sh_y):
"""与头部相连的深色像素 = 头发 + 整条马尾(马尾会垂到肩线以下)"""
rgb = rgba[..., :3].astype(np.float32)
lum = 0.3 * rgb[..., 0] + 0.59 * rgb[..., 1] + 0.11 * rgb[..., 2]
dark = (lum < 72) & (rgba[..., 3] > ALPHA_ON)
dark[int(sh_y + 60) :] = False
_, lab = cv2.connectedComponents(dark.astype(np.uint8), connectivity=8)
ids = np.unique(lab[head & dark])
return np.isin(lab, ids[ids > 0])
def mirror_frame(rgba: np.ndarray, kp: dict) -> np.ndarray:
"""rgba:对齐后的主帧;kp:该帧关键点(需要 nose / lsh / rsh)"""
xc = (kp["lsh"][0] + kp["rsh"][0]) / 2
sh_y = min(kp["lsh"][1], kp["rsh"][1])
y_cut = sh_y - 14
head = _head_mask(rgba, kp, y_cut)
hair = _hair_mask(rgba, head, sh_y)
# 头部(肩线以上)用原图,靠近肩线处渐变到翻转后的身体
region = cv2.dilate((head | _mirror(head, xc)).astype(np.uint8), np.ones((9, 9), np.uint8)) > 0
region[int(y_cut) :] = False
ramp = np.clip((y_cut - np.arange(rgba.shape[0], dtype=np.float32)) / 16, 0, 1)[:, None]
w_head = cv2.GaussianBlur(region.astype(np.float32), (0, 0), 2.5) * ramp
# 原图的马尾保留;翻转过来的马尾(另一侧肩上)连同半透明发梢一起用原图同位置的内容盖掉
own = cv2.dilate(hair.astype(np.uint8), np.ones((7, 7), np.uint8))
other = cv2.dilate(_mirror(hair, xc).astype(np.uint8), np.ones((17, 17), np.uint8))
w_hair = cv2.GaussianBlur(np.maximum(own, other).astype(np.float32), (0, 0), 2)
w = np.maximum(w_head, w_hair)[..., None]
return _unpremul(_premul(_mirror(rgba, xc)) * (1 - w) + _premul(rgba) * w)