Files
admin 7753939a3b feat: ai-manju skill 初始提交
AI 漫剧流水线技能完整落地:文件驱动、可校验、单镜可重做。

七阶段:基线资产 → 策划叙事 → 分镜中枢 → 配音先行 → 出图/出视频 → 缝合 → 审片归档。

- scripts/: manju.py 编排、submit_t2i/submit_i2v 提交、cleanup_comfyui 清理
- assets/comfyui-manju-cleanup/: ComfyUI 清理扩展,归属校验 fail-closed
- references/: models.md 模型节点图全表、workflow/director/storyboard/pitfalls
- .gitignore: 排除 __pycache__ 与运行产物
2026-11-14 17:49:39 +08:00

272 lines
11 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.
#!/usr/bin/env python3
"""manju 缝合器:已采用片段 + 配音 + 字幕 + HUD 标注 + 落版 -> 成片。
规格 720x1280 / 30fps / H.264+AAC。字幕取 30_shots.json 的 line 原文,
按标点折 1–3 行;HUD 文字全片只标一次(一横一竖);落版 zoompan 缓推。
去水印用 crop 底部带 + 居中填充,不用 drawbox。
用法: python3 tools/render.py projects/EPxx/30_shots.json
"""
import argparse
import json
import os
import re
import subprocess
import sys
def _find_root():
"""工作空间根:MANJU_ROOT > cwd 向上找 config/pipeline.json > 脚本上级。"""
env = os.environ.get("MANJU_ROOT")
if env:
return os.path.abspath(os.path.expanduser(env))
d = os.getcwd()
while True:
if os.path.isfile(os.path.join(d, "config", "pipeline.json")):
return d
nd = os.path.dirname(d)
if nd == d:
break
d = nd
return os.path.dirname(os.path.dirname(os.path.abspath(__file__)))
ROOT = _find_root()
CFG = json.load(open(os.path.join(ROOT, "config", "pipeline.json"), encoding="utf-8"))
OUT = CFG["output"]
def ff():
import imageio_ffmpeg
return imageio_ffmpeg.get_ffmpeg_exe()
def run(args, quiet=True):
cmd = [ff(), "-hide_banner", "-loglevel", "error", "-y"] + args
r = subprocess.run(cmd, capture_output=True, text=True)
if r.returncode != 0:
print("FFMPEG FAILED:\n" + " ".join(cmd) + "\n" + r.stderr[-3000:], file=sys.stderr)
raise SystemExit(1)
return r
def probe(path):
r = subprocess.run([ff(), "-hide_banner", "-i", str(path)],
stdout=subprocess.DEVNULL, stderr=subprocess.PIPE, text=True)
m = re.search(r"Duration:\s*(\d+):(\d+):([\d.]+)", r.stderr)
dur = (int(m.group(1)) * 3600 + int(m.group(2)) * 60 + float(m.group(3))) if m else None
return dur, r.stderr
def ts(t):
h = int(t // 3600)
m = int(t % 3600 // 60)
s = t % 60
return "%d:%02d:%05.2f" % (h, m, s)
PUNCT = set(",。;!?、,.;!?")
def wrap_line(text, maxlen=11):
"""按标点折 1–3 行;标点永不落单,永远附着在前一行末尾。"""
parts = [p for p in re.split(r"([,。;!?、,.;!?])", text) if p]
lines, cur = [], ""
for p in parts:
if p in PUNCT:
# 标点直接并入当前行(哪怕略超 maxlen),绝不起新行
cur += p
elif len(cur) + len(p) <= maxlen or not cur:
cur += p
else:
lines.append(cur)
cur = p
if cur:
lines.append(cur)
# 若最后一行只有标点(理论上不会发生),并入上一行
if len(lines) > 1 and all(c in PUNCT for c in lines[-1]):
lines[-2] += lines.pop()
if len(lines) > 3:
# 极端长句:均分三段,切点尽量落在标点后
n = len(text)
cuts = [i + 1 for i, c in enumerate(text) if c in PUNCT]
best = None
for a in cuts:
for b in cuts:
if a < b < n and max(a, b - a, n - b) < (best[0] if best else 999):
best = (max(a, b - a, n - b), a, b)
if best and best[0] <= maxlen + 3:
_, a, b = best
lines = [text[:a], text[a:b], text[b:]]
else:
step = maxlen
lines = [text[i:i + step] for i in range(0, n, step)][:3]
return "\\N".join(lines)
ASS_HEAD = """[Script Info]
ScriptType: v4.00+
PlayResX: 720
PlayResY: 1280
ScaledBorderAndShadow: yes
WrapStyle: 2
[V4+ Styles]
Format: Name, Fontname, Fontsize, PrimaryColour, SecondaryColour, OutlineColour, BackColour, Bold, Italic, Underline, StrikeOut, ScaleX, ScaleY, Spacing, Angle, BorderStyle, Outline, Shadow, Alignment, MarginL, MarginR, MarginV, Encoding
Style: Default,PingFang SC,46,&H00FFFFFF,&H00FFFFFF,&H000C1420,&H90000000,0,0,0,0,100,100,0,0,1,3,1,2,60,60,150,1
Style: HUD,PingFang SC,26,&H00E8D4A8,&H00E8D4A8,&H000C1420,&H90000000,0,0,0,0,100,100,2,0,1,2,0,7,40,40,60,1
[Events]
Format: Layer, Start, End, Style, Name, MarginL, MarginR, MarginV, Effect, Text
"""
def build_ass(shots, ass_path):
lines = [ASS_HEAD]
t = 0.0
for i, s in enumerate(shots):
d = float(s["duration"])
body = wrap_line(s["line"])
lines.append("Dialogue: 0,%s,%s,Default,,0,0,0,,{\\fad(250,250)}%s\n" %
(ts(t + 0.22), ts(t + d - 0.22), body))
if i == 0: # 横排 HUD:全片只标一次
lines.append("Dialogue: 0,%s,%s,HUD,,0,0,0,,{\\an7\\pos(56,92)\\fad(400,400)}2010 / 机房\n" %
(ts(t + 0.5), ts(t + 2.8)))
if i == len(shots) - 1: # 竖排 HUD:全片只标一次
lines.append("Dialogue: 0,%s,%s,HUD,,0,0,0,,{\\an5\\pos(650,340)\\fad(400,400)}二\\N〇\\N二\\N六\n" %
(ts(t + 0.5), ts(t + 2.8)))
t += d
with open(ass_path, "w", encoding="utf-8") as f:
f.writelines(lines)
return t
def main():
ap = argparse.ArgumentParser()
ap.add_argument("shots")
ap.add_argument("--crop-bottom", type=int, default=0,
help="底部水印带高度;A14B 实测无水印,默认 0(不裁画面)")
ap.add_argument("--outro-seconds", type=float, default=2.0)
ap.add_argument("--name", default=None)
a = ap.parse_args()
shots_path = os.path.abspath(a.shots)
data = json.load(open(shots_path, encoding="utf-8"))
ep = data["episode"]
shots = data["shots"]
vid_dir = os.path.join(ROOT, "work", ep, "videos")
aud_dir = os.path.join(ROOT, "work", ep, "audio")
tmp = os.path.join(ROOT, "work", ep, "render")
os.makedirs(tmp, exist_ok=True)
runs_path = os.path.join(vid_dir, "a14b_runs.json")
runs = json.load(open(runs_path, encoding="utf-8")) if os.path.exists(runs_path) else {}
# 1) 逐段标准化:去水印带 -> 居中填充到 720x1280 -> 30fps -> clone 补足精确时长
norm = []
for i, s in enumerate(shots, 1):
sid = s["shotId"]
rec = runs.get(sid)
if not rec:
print("ERROR: 缺 %s 的 A14B 产物记录" % sid, file=sys.stderr)
return 1
src = os.path.join(ROOT, rec["out"])
if not os.path.exists(src):
print("ERROR: 缺视频 %s" % src, file=sys.stderr)
return 1
d = float(s["duration"])
keep = float(s.get("keepSeconds", d))
dst = os.path.join(tmp, "norm_%s.mp4" % sid)
base = ("crop=iw:ih-%d:0:0,scale=%d:%d:force_original_aspect_ratio=increase,"
"crop=%d:%d,setsar=1,fps=%d" %
(a.crop_bottom, OUT["width"], OUT["height"], OUT["width"], OUT["height"],
OUT["fps"]))
if keep < d - 0.05:
# 局部修正:只采用前 keep 秒(漂移前的稳定段),慢放到目标时长
tail = ("trim=duration=%.3f,setpts=PTS*%.6f,fps=%d,"
"trim=duration=%.3f,setpts=PTS-STARTPTS,format=yuv420p" %
(keep, d / keep, OUT["fps"], d))
note = "采用前 %.2fs 慢放至 %.2fs" % (keep, d)
else:
# 常规:A14B 约 5.06s,clone 尾帧补足精确时长
tail = ("tpad=stop_mode=clone:stop_duration=3,"
"trim=duration=%.3f,setpts=PTS-STARTPTS,format=yuv420p" % d)
note = "全段"
vf = base + "," + tail
run(["-i", src, "-vf", vf, "-an", "-c:v", "libx264", "-preset", "veryfast",
"-crf", "19", "-r", str(OUT["fps"]), dst])
norm.append(dst)
print(" 标准化 %s -> %.2fs(%s)" % (sid, d, note), flush=True)
# 2) 落版:光环 / 光束 / 星点 + zoompan 缓推,无文字
outro_png = os.path.join(ROOT, "assets", "overlay", "outro.png")
outro = os.path.join(tmp, "outro.mp4")
frames = int(a.outro_seconds * OUT["fps"])
zvf = ("zoompan=z='min(zoom+0.0008,1.10)':x='iw/2-(iw/zoom/2)':"
"y='ih/2-(ih/zoom/2)':d=%d:s=%dx%d:fps=%d,format=yuv420p" %
(frames, OUT["width"], OUT["height"], OUT["fps"]))
run(["-loop", "1", "-framerate", str(OUT["fps"]), "-i", outro_png,
"-vf", zvf, "-t", str(a.outro_seconds), "-an", "-c:v", "libx264",
"-preset", "veryfast", "-crf", "19", outro])
print(" 落版 %.2fs" % a.outro_seconds, flush=True)
# 3) 字幕 + HUD(一横一竖,各一次)
ass_path = os.path.join(tmp, "ep01.ass")
main_dur = build_ass(shots, ass_path)
total = main_dur + a.outro_seconds
print(" 字幕 %d 条 / 主片 %.2fs / 总长 %.2fs" % (len(shots), main_dur, total), flush=True)
# 4) 一次性拼接 + 烧字幕(单次编码,避免多次重编码掉画质)
video_only = os.path.join(tmp, "video_final.mp4")
ins = []
for p in norm + [outro]:
ins += ["-i", p]
concat_in = "".join("[%d:v]" % i for i in range(len(norm) + 1))
fc = ("%sconcat=n=%d:v=1:a=0[cat];[cat]ass=%s:fontsdir=/System/Library/Fonts[v]" %
(concat_in, len(norm) + 1, ass_path))
run(ins + ["-filter_complex", fc, "-map", "[v]", "-an", "-c:v", "libx264",
"-preset", "veryfast", "-crf", "19", "-pix_fmt", OUT["pix_fmt"],
"-r", str(OUT["fps"]), video_only])
# 5) 音轨:每段配音补静音到本镜时长,再补落版静音
audio_m4a = os.path.join(tmp, "audio_final.m4a")
ains, parts, labels = [], [], []
for i, s in enumerate(shots):
p = os.path.join(aud_dir, "vo_%02d.mp3" % (i + 1))
if not os.path.exists(p):
print("ERROR: 缺配音 %s" % p, file=sys.stderr)
return 1
ains += ["-i", p]
parts.append("[%d:a]apad,atrim=duration=%.3f,aresample=48000[a%d]" %
(i, float(s["duration"]), i))
labels.append("[a%d]" % i)
afc = ";".join(parts) + ";" + "".join(labels) + \
"concat=n=%d:v=0:a=1,apad,atrim=duration=%.3f[aout]" % (len(shots), total)
run(ains + ["-filter_complex", afc, "-map", "[aout]", "-c:a", "aac",
"-b:a", "192k", "-ar", "48000", audio_m4a])
# 6) 封装
out_name = a.name or "%s_成片_v2.mp4" % ep
out_path = os.path.join(ROOT, "renders", out_name)
os.makedirs(os.path.dirname(out_path), exist_ok=True)
run(["-i", video_only, "-i", audio_m4a, "-c", "copy", "-shortest",
"-movflags", "+faststart", out_path])
dur, info = probe(out_path)
streams = [l.strip() for l in info.splitlines() if "Stream #" in l]
report = {"output": os.path.relpath(out_path, ROOT), "duration": dur,
"mainDuration": round(main_dur, 2), "outroSeconds": a.outro_seconds,
"bytes": os.path.getsize(out_path), "streams": streams,
"shots": [{"shotId": s["shotId"], "duration": s["duration"],
"video": runs[s["shotId"]]["out"]} for s in shots]}
with open(os.path.join(tmp, "render_report.json"), "w", encoding="utf-8") as f:
json.dump(report, f, ensure_ascii=False, indent=2)
print("\n=== 成片 ===", flush=True)
print(out_path, flush=True)
print("时长 %.2fs | %.2f MB" % (dur or 0, os.path.getsize(out_path) / 1048576), flush=True)
for l in streams:
print(" " + l, flush=True)
return 0
if __name__ == "__main__":
sys.exit(main())