feat: 原生音画同出闭环 + 队列自检修复 (v1.2.0)

- 新增 render_native.py: 台词镜缝合(保留原生音轨+烧h3v3字幕+瑕疵遮盖,不叠TTS)
- 新增 fetch_t2i.py: 出图脚本被杀后按pid救回单图
- 修复 queue_h3_native.py --check 将中断/error误判为在队的bug
- 文档: render.md/native-av.md/pitfalls.md/SKILL.md/README 补齐双轨缝合与选机实测
This commit is contained in:
ssw committed 2026-11-19 10:18:38 +08:00
1 parent 7753939a3b
commit d25c9304e1
15 files changed
+1485 -83

No files matched your search

+199
View File
@@ -0,0 +1,199 @@
#!/usr/bin/env python3
"""manju H3AV 排队器:把整集台词镜头一次性推进 ComfyUI 队列(只提交、不等待)。
为什么分成「排队 / 取回」两步:单镜约 90 分钟,一集 8 镜 ≈ 12 小时。
本地进程挂在等待上没有任何收益 —— 会话被回收、终端断掉,任务还在服务端跑,pid 却丢了。
所以:一次排完 → 落盘 prompt_id → 任意时刻 fetch_h3_native.py 取回。
用法:
python3 tools/queue_h3_native.py --ep EP01 # 排全部带 .h3v3.txt 的镜
python3 tools/queue_h3_native.py --ep EP01 --only S01a S02
python3 tools/queue_h3_native.py --ep EP01 --check # 只做丢失自检,不提交
★ 交完必须自检(脚本最后自动做):/queue 里找不到、history 里也没有 = 任务已丢。
queue 与 history 都是内存态,实例重启会静默清空 —— 见 references/native-av.md §6。
"""
import argparse
import hashlib
import json
import os
import re
import sys
import time
import uuid
sys.path.insert(0, os.path.dirname(os.path.abspath(__file__)))
import submit_h3_native as H # noqa: E402
def load_shots(root, ep):
p = os.path.join(root, "projects", ep, "30_shots.json")
if not os.path.isfile(p):
raise SystemExit("找不到分镜表:%s" % p)
d = json.load(open(p, encoding="utf-8"))
return d.get("shots", d) if isinstance(d, dict) else d
def find_prompt(root, ep, sid):
base = os.path.join(root, "projects", ep, "40_prompts", sid)
for ext in (".h3v3.txt", ".h3.txt"):
if os.path.isfile(base + ext):
return base + ext
return None
def find_first(root, ep, sid):
"""优先用分镜表里的 firstFrame,其次 work/<EP>/images/<SHOT>.png,再退回去掉尾字母。"""
cand = [os.path.join(root, "work", ep, "images", sid + ".png"),
os.path.join(root, "work", ep, "images", re.sub(r"[a-z]$", "", sid) + ".png")]
return next((p for p in cand if os.path.isfile(p)), None)
def queue_pids(host):
try:
q = H.getj(host, "/queue", timeout=20)
except Exception as e:
print("读取队列失败(继续):", str(e)[:80], flush=True)
return set()
return {str(it[1]) for it in (q.get("queue_running") or []) + (q.get("queue_pending") or [])}
def selfcheck(host, m, only_new=None):
"""排队/取回前后都能跑的丢失自检。返回 (在队, 已完成, 丢失)。"""
live = queue_pids(host)
q_ok = q_done = lost = 0
for sid in sorted(m):
pid = m[sid].get("prompt_id")
if not pid or (only_new and sid not in only_new):
continue
if pid in live:
q_ok += 1
continue
try:
hh = H.getj(host, "/history/" + pid, timeout=20)
except Exception:
q_ok += 1 # 网络问题不下结论
continue
if pid in hh and hh[pid].get("status", {}).get("status_str") == "success":
q_done += 1
elif pid in hh:
# 有 history 记录但非 success:被中断/报错,不能算"在队"
_st = hh[pid].get("status", {}).get("status_str", "?")
lost += 1
print(" ⛔ %-5s pid=%s 状态=%s(非 success,疑似被中断/失败)→ 需 --force 重排"
% (sid, pid[:8], _st), flush=True)
else:
lost += 1
print(" ⛔ %-5s pid=%s 既不在队列也不在 history → 已随实例重启丢失,需重排"
% (sid, pid[:8]), flush=True)
return q_ok, q_done, lost
def main():
ap = argparse.ArgumentParser()
ap.add_argument("--ep", required=True)
ap.add_argument("--only", nargs="*", default=None)
ap.add_argument("--force", action="store_true", help="已出片/已提交的也重排(换提示词后用)")
ap.add_argument("--check", action="store_true", help="只做丢失自检,不提交")
ap.add_argument("--w", type=int, default=0)
ap.add_argument("--h", type=int, default=0)
ap.add_argument("--steps", type=int, default=0)
a = ap.parse_args()
root = H.find_root()
cfg = H.load_cfg(root)
host = cfg["endpoint"]
fps = int(cfg.get("fps", 24))
W = a.w or int(cfg.get("width", 576))
Hh = a.h or int(cfg.get("height", 1024))
steps = a.steps or int(cfg.get("steps", 20))
outdir = os.path.join(root, "work", a.ep, "videos", "native")
os.makedirs(outdir, exist_ok=True)
mapfile = os.path.join(outdir, "_queue_map.json")
m = {}
if os.path.isfile(mapfile):
try:
m = json.load(open(mapfile, encoding="utf-8"))
except Exception:
m = {}
if a.check:
q_ok, q_done, lost = selfcheck(host, m)
print("\n在队 %d / 已完成 %d / 丢失 %d" % (q_ok, q_done, lost))
return 0 if lost == 0 else 2
weights = H.pick_weights(host) # 一次探测,全批复用
live = queue_pids(host)
new = set()
for s in load_shots(root, a.ep):
sid = s.get("shotId") or s.get("id")
if not sid or (a.only and sid not in a.only):
continue
pf = find_prompt(root, a.ep, sid)
if not pf:
print("[skip] %-5s 没有 .h3v3.txt 提示词" % sid, flush=True)
continue
out = os.path.join(outdir, sid + ".mp4")
if os.path.exists(out) and not a.force:
print("[skip] %-5s 已出片 %.1f MB" % (sid, os.path.getsize(out) / 1e6), flush=True)
continue
prev = m.get(sid, {}).get("prompt_id")
if prev and not a.force:
if prev in live:
print("[skip] %-5s 已在本机队列 pid=%s" % (sid, prev[:8]), flush=True)
continue
try:
hh = H.getj(host, "/history/" + prev, timeout=20)
if prev in hh and hh[prev].get("status", {}).get("status_str") == "success":
print("[skip] %-5s 服务端已成功(用 fetch 取回)" % sid, flush=True)
continue
except Exception:
pass
print("[skip] %-5s 已提交过 pid=%s(如已随重启丢失,加 --force 重排)"
% (sid, prev[:8]), flush=True)
continue
img = find_first(root, a.ep, sid)
if not img:
print("[skip] %-5s 没有首帧图(要纯文生视频用 submit_h3_native.py --t2v)"
% sid, flush=True)
continue
prompt = open(pf, encoding="utf-8").read().strip()
secs = float(s.get("duration") or 7.0)
length = H.snap_len(round(secs * fps))
if length < 124:
print(" ⚠ %-5s %.2fs → %d 帧,低于训练区间 124 帧,画质会掉;"
"建议把这场戏并到相邻镜" % (sid, secs, length), flush=True)
image_name = H.upload(host, img)
seed = int(hashlib.sha256((sid + prompt).encode()).hexdigest()[:8], 16) % (2 ** 31)
graph = H.build(weights, image_name, prompt, W, Hh, length, seed, steps,
True, True, fps, cfg.get("prefix", "video/h3k"))
pid, num = H.submit(host, graph, "manju-h3av-" + uuid.uuid4().hex[:8])
m[sid] = {"prompt_id": pid, "first": os.path.relpath(img, root),
"promptFile": os.path.relpath(pf, root), "seconds": secs,
"length": length, "w": W, "h": Hh, "steps": steps, "seed": seed,
"endpoint": host,
"submitted_at": time.strftime("%Y-%m-%d %H:%M:%S")}
json.dump(m, open(mapfile, "w", encoding="utf-8"), ensure_ascii=False, indent=2)
new.add(sid)
print("[queue] %-5s pid=%s 队列#%s %dx%d %d帧=%.2fs"
% (sid, pid, num, W, Hh, length, length / float(fps)), flush=True)
time.sleep(1.5)
total = sum(v.get("length", 0) for v in m.values()) / float(fps)
print("\n本次新排 %d 镜,累计 %d 镜 / %.1f 秒成片;单镜约 90 分钟,"
"按串行估 ≈ %.1f 小时" % (len(new), len(m), total, len(m) * 1.5), flush=True)
print("映射表 ->", mapfile, flush=True)
print("\n丢失自检(在队=队列有任务;丢失=重启清掉或被中断/error):", flush=True)
q_ok, q_done, lost = selfcheck(host, m)
print("在队 %d / 已完成 %d / 丢失 %d" % (q_ok, q_done, lost), flush=True)
if lost:
print("⚠ 有任务丢失/失败:加 --force 重排;error 状态多为被中断,核对是否有人动了队列", flush=True)
return 0
if __name__ == "__main__":
sys.exit(main())