""" 插帧与挑帧 插帧:在相邻两帧之间用 RIFE(rife-ncnn-vulkan,v4.6 模型)生成中间帧。 帧已经很密、相邻两帧动作很小,插出来的中间帧和原帧难以区分; 但手从下垂变成握拳这类“外形突变”的地方会插出半透明残影——用中间帧里半透明像素的多少判断, 比两侧原帧多出 10% 以上就不用。 RIFE 不处理透明通道:颜色贴在浅灰底上插(边缘就算有一点偏差也是浅灰,在亮色背景上看不出), 透明度单独插一次,再反解出前景色。 挑帧:按动作幅度均匀地挑出 N 帧(包体有上限,帧数和清晰度要平衡)。 RIFE 首次使用时自动下载到 training-videos/work/models(macOS 版)。 """ import subprocess import tempfile import urllib.request import zipfile from pathlib import Path import numpy as np from PIL import Image ROOT = Path(__file__).resolve().parents[1] MODEL_DIR = ROOT / "work" / "models" RIFE_URL = "https://github.com/nihui/rife-ncnn-vulkan/releases/download/20221029/rife-ncnn-vulkan-20221029-macos.zip" RIFE_DIR = MODEL_DIR / "rife-ncnn-vulkan-20221029-macos" MODEL = "rife-v4.6" BG = np.array([237, 241, 239], np.float32) GHOST_RATIO = 1.10 def _rife() -> Path: exe = RIFE_DIR / "rife-ncnn-vulkan" if not exe.exists(): MODEL_DIR.mkdir(parents=True, exist_ok=True) zip_path = MODEL_DIR / "rife.zip" print("下载 RIFE …") urllib.request.urlretrieve(RIFE_URL, zip_path) with zipfile.ZipFile(zip_path) as z: z.extractall(MODEL_DIR) exe.chmod(0o755) return exe def _run(a: Path, b: Path, out: Path): subprocess.run([str(_rife()), "-0", str(a), "-1", str(b), "-o", str(out), "-m", str(RIFE_DIR / MODEL), "-s", "0.5"], check=True, capture_output=True) def interpolate(a: np.ndarray, b: np.ndarray) -> np.ndarray: """两张 RGBA 主帧 → 正中间的一帧""" with tempfile.TemporaryDirectory() as tmp: t = Path(tmp) for name, f in (("a", a), ("b", b)): x = f.astype(np.float32) al = x[..., 3:4] / 255 Image.fromarray((x[..., :3] * al + BG * (1 - al)).round().astype(np.uint8)).save(t / f"{name}_rgb.png") Image.fromarray(np.dstack([f[..., 3]] * 3)).save(t / f"{name}_a.png") _run(t / "a_rgb.png", t / "b_rgb.png", t / "mid_rgb.png") _run(t / "a_a.png", t / "b_a.png", t / "mid_a.png") rgb = np.array(Image.open(t / "mid_rgb.png").convert("RGB")).astype(np.float32) al = np.array(Image.open(t / "mid_a.png").convert("L")).astype(np.float32) / 255 al = np.where(al < 0.04, 0, al) fg = np.clip((rgb - (1 - al[..., None]) * BG) / np.maximum(al, 1e-3)[..., None], 0, 255) return np.dstack([fg, al * 255]).round().astype(np.uint8) def _semi(f: np.ndarray) -> int: al = f[..., 3] / 255.0 return int(((al > 0.25) & (al < 0.75)).sum()) def densify(frames: list, kps: list, amounts: list, max_gap: float): """幅度差大于 max_gap 的相邻两帧之间插一帧(有残影的跳过);返回新的帧、关键点、幅度与插入数""" out_f, out_k, out_a = [frames[0]], [kps[0]], [amounts[0]] added = 0 for i in range(len(frames) - 1): a, b = frames[i], frames[i + 1] if amounts[i + 1] - amounts[i] > max_gap: mid = interpolate(a, b) if _semi(mid) <= GHOST_RATIO * (_semi(a) + _semi(b)) / 2: out_f.append(mid) out_k.append({k: tuple((np.array(kps[i][k]) + np.array(kps[i + 1][k])) / 2) for k in kps[i]}) out_a.append(round((amounts[i] + amounts[i + 1]) / 2, 3)) added += 1 out_f.append(b) out_k.append(kps[i + 1]) out_a.append(amounts[i + 1]) return out_f, out_k, out_a, added def thin(amounts: list, n: int) -> list: """按动作幅度均匀挑 n 帧(保留首尾),返回下标""" if len(amounts) <= n: return list(range(len(amounts))) picked = [0] for k in range(1, n - 1): target = k / (n - 1) # 在剩余可选范围内找幅度最接近目标的帧,保证下标递增、给后面留够位置 lo, hi = picked[-1] + 1, len(amounts) - (n - 1 - k) picked.append(min(range(lo, hi), key=lambda i: abs(amounts[i] - target))) picked.append(len(amounts) - 1) return picked