lukas/lektor-ai
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

Survive a ComfyUI crash, retry clips the model cuts short

Browse Source 
Crash recovery
  ComfyUI's queue and history do not persist across a restart, so a prompt
  submitted before a crash exists nowhere afterwards. The old poll loop waited
  on that dead prompt for ever, and its error branch skipped the backoff, so it
  hammered the closed port four times a second printing "Poll error" on every
  attempt.

  run_workflow() replaces wait_for_completion(). It reports once when the
  backend goes away and once when it returns, polls every 3s in between, and on
  reconnect checks both /history and /queue: if the prompt is in neither, the
  queue did not survive and the clip is submitted again. When it cannot tell, it
  assumes the job is still alive — resubmitting a running job would waste a
  generation.

  It also starts ComfyUI itself, via the script named by [comfy] start_command,
  in a new terminal window. Guarded by a 90s cooldown and 3 attempts per run so
  a backend that dies during startup cannot spawn an endless stream of windows;
  the window pauses on exit so a failed start leaves its error on screen. Falls
  back to a background launch with a log file when there is no display. Set
  [comfy] autostart = no, or pass --no-comfy-autostart, to turn it off.

Clips cut short mid-word
  About one clip in twelve comes back truncated: the model stops before the line
  is finished. Measured on raw ComfyUI output, 6 of 80 recent clips, and 15 of
  472 in one episode. It is invisible in the text but obvious in the audio — a
  finished clip decays to about -65 dB, a cut one stops at about -16 dB.

  Truncated clips are now regenerated with a distant seed, up to
  [tts] truncation_retries extra attempts, and a run that still has bad clips
  ends by printing them as a ready-made --regen list. --check-clips reports the
  same for a job already on disk.

  Worth recording what this is NOT: the trailing-silence trimming added earlier
  was suspected, and cleared. It only cuts silence that runs to end of file, so
  a pause between words never triggers it — verified against six synthetic cases
  including a 1.2s mid-clip pause, and against real clips containing pauses. The
  truncation is present in ComfyUI's own output before this script touches it.

settings.ini
  New: [tts] truncation_retries, [comfy] autostart and start_command, all
  documented in place. The editor renders yes/no keys as a dropdown.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
This commit is contained in:
claude
2026-08-10 14:12:33 +02:00
co-authored by Claude Opus 5
parent 200d221816
commit 7e2cee02e5
3 changed files with 298 additions and 33 deletions
Download Patch File Download Diff File Expand all files Collapse all files
lektor_ai_comfy.py
+273 -30
View File
@@ -11,8 +11,10 @@ Requires ComfyUI with TTS-Audio-Suite running at http://127.0.0.1:8189
import argparse
import configparser
import json
import os
import random
import re
import shlex
import shutil
import subprocess
import sys
@@ -32,6 +34,17 @@ VOICE_PROMPTS_DIR = Path(__file__).resolve().parent / "voices"
SETTINGS_FILE = Path(__file__).resolve().parent / "settings.ini"
def _as_bool(raw) -> bool:
"""yes/no setting. configparser's own getboolean is bypassed by the
per-key cast in load_settings, so accept the same spellings here."""
value = str(raw).strip().lower()
if value in ("1", "true", "yes", "on"):
return True
if value in ("0", "false", "no", "off"):
return False
raise ValueError(f"{raw!r} is not a yes/no value")
# Fallback values, used when settings.ini is absent or omits a key. Keeping them
# here means the script still runs correctly with no settings file at all.
# Each entry is (ini_section, ini_key, type, value).
@@ -40,16 +53,23 @@ _DEFAULTS = {
"speed": ("tts", "speed", float, 1.2),
"temperature": ("tts", "temperature", float, 0.9),
"max_tokens": ("tts", "max_tokens", int, 4096),
"truncation_retries": ("tts", "truncation_retries", int, 2),
"strip_leading_ms": ("tts", "strip_leading_ms", float, 0.0),
"trailing_fade_ms": ("tts", "trailing_fade_ms", float, 80.0),
"orig_volume": ("mix", "orig_volume", float, 0.60),
"tts_volume": ("mix", "tts_volume", float, 1.8),
"output_dir": ("paths", "output_dir", str, "output"),
"comfy_url": ("comfy", "comfy_url", str, "http://127.0.0.1:8189"),
"comfy_autostart": ("comfy", "autostart", _as_bool, True),
"comfy_start_command": ("comfy", "start_command", str, "start_comfyui.sh"),
}
COMFY_URL = _DEFAULTS["comfy_url"][3] # rebound from settings/CLI in main()
# Rebound in main() too; used when the backend stops answering.
COMFY_AUTOSTART = True
COMFY_START_SCRIPT = Path(__file__).resolve().parent / "start_comfyui.sh"
def load_settings(path: Path) -> dict:
"""Read settings.ini, falling back to _DEFAULTS for anything missing.
@@ -416,6 +436,31 @@ FLAT_TAIL_MIN = 5.0 # the flat run must last at least this long
RUNAWAY_SECONDS = 20.0 # only look for tone padding on clips this long
# A clip the model finished decays to near silence; one it cut off mid-word
# stops at full level. Measured across 80 raw outputs the two groups sit about
# 50 dB apart (-65 dB vs -16 dB), so the test is unambiguous.
TRUNCATION_PEAK_DB = -25.0
TRUNCATION_WINDOW = 0.08
def ends_abruptly(path: Path, window: float = TRUNCATION_WINDOW) -> bool:
"""True when a clip stops at full level, i.e. the model cut it mid-word."""
try:
duration = get_audio_duration(path)
except RuntimeError:
return False
if duration <= window:
return False
result = subprocess.run(
["ffmpeg", "-hide_banner", "-ss", f"{duration - window:.3f}",
"-t", f"{window}", "-i", str(path),
"-af", "volumedetect", "-f", "null", "-"],
capture_output=True, text=True,
)
peak = re.search(r"max_volume:\s*(-?[0-9.]+)", result.stderr)
return bool(peak) and float(peak.group(1)) > TRUNCATION_PEAK_DB
def flat_tail_start(path: Path, window: float = 0.5) -> float:
"""Start of a long, unnaturally constant tail, or None."""
result = subprocess.run(
@@ -781,31 +826,146 @@ def submit_workflow(workflow: dict) -> str:
return resp["prompt_id"]
def wait_for_completion(prompt_id: str, poll_interval: float = 0.25,
max_interval: float = 2.0, backoff_after: float = 10.0) -> dict:
"""Block until ComfyUI reports the job finished, then return its history entry.
BACKEND_RETRY_SECONDS = 3.0 # how often to look for ComfyUI while it is down
The poll interval, not the GPU, governs perceived generation speed: a clip
that finishes 100 ms after a poll still waits for the next one. This used to
be a flat 3 s, which quantised every clip to a 3 s boundary — measured across
2075 real clips, 91% of the gaps landed exactly on a multiple of 3 s, and
roughly a third of each run was spent sleeping on already-completed work.
Polling starts tight so the common case (most clips finish in 3-6 s) is
picked up almost immediately, then backs off once a job looks slow so a
stuck or very long generation does not sit at 4 requests/second forever.
# Launching the backend is guarded: a ComfyUI that dies on startup would
# otherwise spawn a new terminal window every few seconds.
COMFY_AUTOSTART_MAX = 3 # attempts per run
COMFY_AUTOSTART_COOLDOWN = 90.0 # seconds between attempts
_autostart = {"attempts": 0, "last": 0.0}
_TERMINALS = (
("gnome-terminal", lambda c: ["gnome-terminal", "--title", "ComfyUI", "--",
"bash", "-lc", c]),
("konsole", lambda c: ["konsole", "-p", "tabtitle=ComfyUI", "-e",
"bash", "-lc", c]),
("xfce4-terminal", lambda c: ["xfce4-terminal", "--title", "ComfyUI",
"-e", f"bash -lc {shlex.quote(c)}"]),
("xterm", lambda c: ["xterm", "-T", "ComfyUI", "-e", "bash", "-lc", c]),
)
def start_comfy_backend(script: Path) -> bool:
"""Launch ComfyUI in its own terminal window. True if one was started."""
now = time.monotonic()
if _autostart["attempts"] >= COMFY_AUTOSTART_MAX:
return False
if _autostart["last"] and now - _autostart["last"] < COMFY_AUTOSTART_COOLDOWN:
return False
if not script.is_file():
print(f" Cannot start ComfyUI automatically: {script} not found.",
file=sys.stderr)
_autostart["attempts"] = COMFY_AUTOSTART_MAX # do not keep complaining
return False
_autostart["attempts"] += 1
_autostart["last"] = now
# Pause on exit so a backend that fails to start leaves its error on screen
# instead of the window vanishing.
inner = (f"{shlex.quote(str(script))}; echo; "
f"echo '[ComfyUI exited — press any key to close]'; read -n 1")
if os.environ.get("DISPLAY") or os.environ.get("WAYLAND_DISPLAY"):
for name, build in _TERMINALS:
if not shutil.which(name):
continue
try:
subprocess.Popen(build(inner), start_new_session=True)
except OSError:
continue
print(f" Starting ComfyUI in a new {name} window "
f"(attempt {_autostart['attempts']} of {COMFY_AUTOSTART_MAX})…",
file=sys.stderr)
return True
log = script.parent / "comfyui_autostart.log"
try:
with open(log, "ab") as handle:
subprocess.Popen([str(script)], stdout=handle, stderr=handle,
start_new_session=True)
except OSError as exc:
print(f" Could not start ComfyUI: {exc}", file=sys.stderr)
return False
print(f" No terminal available — started ComfyUI in the background, "
f"logging to {log}", file=sys.stderr)
return True
def _prompt_known(prompt_id: str) -> bool:
"""True while ComfyUI still has this prompt queued, running or in history.
A restart wipes both, so this is how a job lost to a crash is spotted.
If the answer cannot be determined, say yes: resubmitting a job that is
actually still running would waste a generation.
"""
print(f" Waiting for {prompt_id[:8]}…", end="", flush=True)
start = time.monotonic()
interval = poll_interval
next_dot = 1.0
try:
if prompt_id in _comfy_get(f"/history/{prompt_id}"):
return True
queue = _comfy_get("/queue")
except Exception: # noqa: BLE001
return True
for key in ("queue_running", "queue_pending"):
for item in queue.get(key, []):
if isinstance(item, (list, tuple)) and len(item) > 1 and item[1] == prompt_id:
return True
return False
def run_workflow(workflow: dict, poll_interval: float = 0.25,
max_interval: float = 2.0, backoff_after: float = 10.0) -> dict:
"""Submit a workflow, wait for it, and return its history entry.
Survives ComfyUI dying mid-job. Its queue and history do not persist across
a restart, so a prompt submitted beforehand exists nowhere afterwards and
polling it would wait for ever — the job has to be sent again. While the
backend is down this reports once rather than once per poll, and slows to
BACKEND_RETRY_SECONDS instead of hammering a closed port.
The poll interval, not the GPU, governs perceived speed: a clip that
finishes 100 ms after a poll still waits for the next one. It was a flat
3 s, which quantised every clip to a 3 s boundary — across 2075 real clips
91% of the gaps landed exactly on a multiple of 3 s. So polling starts
tight and only backs off once a job looks slow.
"""
prompt_id, offline = None, False
start, interval, next_dot = time.monotonic(), poll_interval, 1.0
while True:
if prompt_id is None: # first submission, or a resubmit
try:
prompt_id = submit_workflow(workflow)
except Exception as exc: # noqa: BLE001
if not offline:
print(f"\n ComfyUI unreachable ({exc}).", file=sys.stderr)
offline = True
if not (COMFY_AUTOSTART and start_comfy_backend(COMFY_START_SCRIPT)):
print(" Waiting for it to come back — start it and this "
"will carry on.", file=sys.stderr)
time.sleep(BACKEND_RETRY_SECONDS)
continue
start, interval, next_dot = time.monotonic(), poll_interval, 1.0
print(f" Waiting for {prompt_id[:8]}…", end="", flush=True)
try:
history = _comfy_get(f"/history/{prompt_id}")
except Exception as e:
print(f"\n Poll error: {e}", file=sys.stderr)
time.sleep(min(interval, 1.0))
except Exception as exc: # noqa: BLE001
if not offline:
print(f"\n ComfyUI went away ({exc}).", file=sys.stderr)
offline = True
if not (COMFY_AUTOSTART and start_comfy_backend(COMFY_START_SCRIPT)):
print(" Waiting for it to come back — restart it and this "
"will carry on.", file=sys.stderr)
time.sleep(BACKEND_RETRY_SECONDS)
continue
if offline: # it answered again
offline = False
print(" ComfyUI is back.", file=sys.stderr)
if not _prompt_known(prompt_id):
print(" The queue did not survive the restart — resubmitting "
"this clip.", file=sys.stderr)
prompt_id = None
continue
if prompt_id in history:
@@ -843,13 +1003,33 @@ def extract_output_audio(history_entry: dict) -> Path:
def _generate_one(text: str, voice_filename: str, voice_ref_text: str,
seed: int, speed: float, out_dir: Path, label: str,
strip_leading_ms: float, temperature: float = 0.9,
trailing_fade_ms: float = 0, max_new_tokens: int = 4096) -> tuple:
"""Submit one TTS job and return (path, duration_after_speedup)."""
trailing_fade_ms: float = 0, max_new_tokens: int = 4096,
retries: int = 2) -> tuple:
"""Submit one TTS job and return (path, duration_after_speedup, truncated).
About one clip in twelve comes back cut off mid-word — the model stops
generating before the line is finished. It is not predictable from the
text, but it is obvious in the audio, so a clip that ends at full level is
simply asked for again with a different seed.
"""
attempt, truncated = 0, False
while True:
workflow = build_text_workflow(text, voice_filename, voice_ref_text, seed,
temperature, max_new_tokens)
prompt_id = submit_workflow(workflow)
history_entry = wait_for_completion(prompt_id)
raw_clip = extract_output_audio(history_entry)
raw_clip = extract_output_audio(run_workflow(workflow))
truncated = ends_abruptly(raw_clip)
if not truncated or attempt >= retries:
break
attempt += 1
seed += 7919 # a distant seed, not the neighbouring sample
print(f" ! ends mid-word — retrying with seed {seed} "
f"(attempt {attempt + 1} of {retries + 1})")
if truncated:
print(" ! still ends mid-word after "
f"{attempt + 1} attempt{'s' if attempt else ''} — "
"regenerate this one with --regen")
if strip_leading_ms > 0:
stripped = out_dir / f"{label}_s.flac"
@@ -873,14 +1053,14 @@ def _generate_one(text: str, voice_filename: str, voice_ref_text: str,
speed_up_audio(raw_clip, fast_clip, speed, fade_out_ms=trailing_fade_ms)
trimmed.unlink(missing_ok=True)
return fast_clip, get_audio_duration(fast_clip)
return fast_clip, get_audio_duration(fast_clip), truncated
def generate_clips(blocks: list, voice_filename: str, voice_ref_text: str,
base_seed: int, speed: float, clips_dir: Path,
strip_leading_ms: float = 0.0, resume_from: int = 0,
temperature: float = 0.9, trailing_fade_ms: float = 0,
max_new_tokens: int = 4096) -> list:
max_new_tokens: int = 4096, retries: int = 2) -> list:
"""
Generate one audio clip per block via ComfyUI, apply per-clip speed-up,
then calculate start positions with "push forward if overlap" logic.
@@ -893,6 +1073,7 @@ def generate_clips(blocks: list, voice_filename: str, voice_ref_text: str,
clips = []
cursor = 0.0
total = len(blocks)
still_truncated = []
for idx, (t_start, t_end, text) in enumerate(blocks):
clip_n = idx + 1
@@ -914,11 +1095,14 @@ def generate_clips(blocks: list, voice_filename: str, voice_ref_text: str,
print(f" [{clip_n}/{total}] {_fmt_srt_time(t_start)} {display[:60]!r}")
seed = base_seed + idx
fast_clip, duration = _generate_one(
fast_clip, duration, truncated = _generate_one(
text, voice_filename, voice_ref_text,
seed, speed, clips_dir, label, strip_leading_ms, temperature,
trailing_fade_ms=trailing_fade_ms, max_new_tokens=max_new_tokens,
retries=retries,
)
if truncated:
still_truncated.append(clip_n)
# Place at SRT time or after previous clip, whichever is later.
# No trimming — clips play to their natural end; re-align at the next
@@ -931,6 +1115,11 @@ def generate_clips(blocks: list, voice_filename: str, voice_ref_text: str,
delay_str = f" [+{delay:.2f}s behind SRT]" if delay > 0.05 else ""
print(f" → placed at {_fmt_srt_time(actual_start)}, duration {duration:.2f}s{delay_str}")
if still_truncated:
print(f"\n {len(still_truncated)} clip(s) still end mid-word after retries. "
f"To try again:")
print(f" --regen {' '.join(str(n) for n in still_truncated)}")
return clips
@@ -1023,6 +1212,37 @@ def _rebuild_video(args, out_dir: Path):
print(f"\nDone! Output: {final_output}")
def _check_job_clips(args, out_dir: Path):
"""List clips in this video's job folder that end mid-word.
Prints them as a --regen argument list, so fixing them is a copy and paste.
"""
job = job_dir(out_dir, args.video)
clips = sorted(job.glob("clip_*.flac"))
if not clips:
sys.exit(f"No clips found in {job}." + _legacy_hint(out_dir, job))
print(f"\n Checking {len(clips)} clips in {job}…")
bad = []
for clip in clips:
if ends_abruptly(clip):
try:
number = int(clip.stem.split("_")[1])
except (IndexError, ValueError):
continue
bad.append(number)
print(f" {clip.name} ends at full level "
f"({get_audio_duration(clip):.2f}s)")
if not bad:
print(f"\n No truncated clips. All {len(clips)} end cleanly.")
return
print(f"\n {len(bad)} of {len(clips)} clips ({100 * len(bad) / len(clips):.0f}%) "
f"end mid-word. To regenerate them:")
print(f" python3 {Path(__file__).name} \"{args.video}\" --regen "
f"{' '.join(str(n) for n in bad)}")
def _trim_job_clips(args, out_dir: Path):
"""Trim trailing silence from every clip already on disk for this video.
@@ -1100,12 +1320,13 @@ def _regen_clips(args, out_dir: Path):
label = f"clip_{clip_n:04d}"
print(f"\n REGEN [{clip_n}/{total}] {text[:70]!r}")
seed = base_seed + idx
clip_path, duration = _generate_one(
clip_path, duration, truncated = _generate_one(
text, voice_filename, voice_ref_text,
seed, args.speed, clips_dir, label,
args.strip_leading_ms, args.temperature,
trailing_fade_ms=args.trailing_fade_ms,
max_new_tokens=args.max_tokens,
retries=args.truncation_retries,
)
print(f" → saved as {clip_path.name} ({duration:.2f}s)")
@@ -1117,7 +1338,7 @@ def _regen_clips(args, out_dir: Path):
# ---------------------------------------------------------------------------
def main():
global COMFY_URL
global COMFY_URL, COMFY_AUTOSTART, COMFY_START_SCRIPT
# --settings is resolved before the main parser is built, because the file's
# contents become that parser's defaults.
@@ -1188,6 +1409,18 @@ def main():
"(same number as in clips/clip_NNNN.flac and tts_input.srt). "
"Only replaces the selected FLAC files; run --rebuild "
"afterwards to reassemble audio and rebuild the video.")
parser.add_argument("--no-comfy-autostart", action="store_true",
help="Do not launch ComfyUI automatically when it is not "
"answering. By default it is started via the script named "
"by [comfy] start_command in settings.ini.")
parser.add_argument("--truncation-retries", type=int,
default=cfg["truncation_retries"], metavar="N",
help=f"How many extra attempts to make when the model returns a "
f"clip cut off mid-word (default: {cfg['truncation_retries']}; "
f"0 disables retrying).")
parser.add_argument("--check-clips", action="store_true",
help="Report which clips in this video's job folder end mid-word, "
"as a ready-made --regen list, then stop.")
parser.add_argument("--trim-silence", action="store_true",
help="Trim trailing silence from the clips already generated for "
"this video, then stop. Repairs runs where the model failed "
@@ -1204,6 +1437,10 @@ def main():
args = parser.parse_args()
COMFY_URL = args.comfy_url
COMFY_AUTOSTART = cfg["comfy_autostart"] and not args.no_comfy_autostart
start_cmd = Path(cfg["comfy_start_command"]).expanduser()
COMFY_START_SCRIPT = start_cmd if start_cmd.is_absolute() \
else Path(__file__).resolve().parent / start_cmd
if settings_path.is_file():
print(f" Settings: {settings_path}")
@@ -1215,7 +1452,7 @@ def main():
# --regen without an SRT), argparse puts the lone positional into args.srt.
# Detect this and move it to args.video.
needs_srt = not (args.embedded_subs or args.rebuild or args.trim_silence
or bool(args.regen))
or args.check_clips or bool(args.regen))
if args.video is None and args.srt is not None and not needs_srt:
args.video = args.srt
args.srt = None
@@ -1228,6 +1465,11 @@ def main():
out_dir = Path(args.output_dir)
out_dir.mkdir(parents=True, exist_ok=True)
# ── Maintenance: report clips the model cut short, then stop ─────────────
if args.check_clips:
_check_job_clips(args, out_dir)
return
# ── Maintenance: trim silence from existing clips, then stop ─────────────
if args.trim_silence:
_trim_job_clips(args, out_dir)
@@ -1333,7 +1575,8 @@ def main():
resume_from=args.resume_from,
temperature=args.temperature,
trailing_fade_ms=args.trailing_fade_ms,
max_new_tokens=args.max_tokens)
max_new_tokens=args.max_tokens,
retries=args.truncation_retries)
assembled_tts = clips_dir / "tts_assembled.flac"
assemble_timed_audio(clips, str(assembled_tts))
settings.ini
+15
View File
@@ -29,6 +29,11 @@ temperature = 0.6
# mid-word.
max_tokens = 4096
# Roughly one clip in twelve comes back cut off mid-word. Those are detected and
# asked for again with a different seed, up to this many extra attempts. 0 to
# switch retrying off.
truncation_retries = 2
# Trim this many ms from the start of every clip, to remove an inhale artifact.
# Try 80-150 if you hear one. Too high clips the first consonant.
strip_leading_ms = 0
@@ -54,3 +59,13 @@ output_dir = output
[comfy]
# Address of the ComfyUI backend running Higgs Audio V3.
comfy_url = http://127.0.0.1:8189
# Start ComfyUI automatically when it stops answering. A crash loses its queue,
# so the clip in flight is resubmitted once the backend is back. At most 3
# attempts per run, 90s apart, so a backend that dies on startup cannot spawn an
# endless stream of terminal windows. yes/no.
autostart = yes
# The script used to start it. Relative paths are resolved next to
# lektor_ai_comfy.py.
start_command = start_comfyui.sh
ui.py
+8 -1
View File
@@ -222,12 +222,15 @@ SETTING_LABELS = {
"speed": "speed",
"temperature": "speech temperature",
"max_tokens": "max tokens per clip",
"truncation_retries": "retries if cut mid-word",
"strip_leading_ms": "strip leading (ms)",
"trailing_fade_ms": "trailing fade (ms)",
"orig_volume": "sound volume",
"tts_volume": "voiceover volume",
"output_dir": "output folder",
"comfy_url": "ComfyUI URL",
"comfy_autostart": "start ComfyUI if down",
"comfy_start_command": "start script",
}
SECTION_TITLES = {"tts": "Speech", "mix": "Mix", "paths": "Paths", "comfy": "Backend"}
@@ -353,7 +356,11 @@ class SettingsDialog(tk.Toplevel):
font=(F, 10)).grid(row=row, column=0, sticky="w", padx=(6, 14))
var = tk.StringVar(value=current[key])
self.vars[key] = var
if key == "voice" and voices:
if core._DEFAULTS[key][2] is core._as_bool:
w = ttk.Combobox(body, textvariable=var, values=["yes", "no"],
width=26, state="readonly",
style="Dark.TCombobox", font=(F, 10))
elif key == "voice" and voices:
w = ttk.Combobox(body, textvariable=var, values=voices, width=26,
style="Dark.TCombobox", font=(F, 10))
else: