🚫 禁區:LatentSync 對嘴(2026-10-18 用戶否決)
LatentSync 對嘴產出的嘴巴變得很不自然(面具偽影)——用戶明確否決,這條路不行,勿再嘗試。
- 口型對不上根因:把 IndexTTS 音訊硬接回 H3 主檔,畫面嘴型是 H3 原生語音,必錯位。原本嘗試用 LatentSync 對嘴解決,但對嘴後嘴巴不自然,用戶否決 → LatentSync 列為禁區。
- LatentSync 呼叫(Windows PowerShell):D:\AI TOOL\WANTZ\LatentSync50兼容50系显卡数字人工具\extracted\LatentSync50-V16,設 PATH=\wzf;\wzf\Scripts;\wzf\ffmpeg\bin;\wzf\Lib\site-packages\torch\lib;\wzf\Library\bin;C:\Users\Lucky star\AppData\Local\Doubao\User Data\Default\connector_runtime\shims;C:\Users\Lucky star\AppData\Local\Doubao\User Data\Default\connector_runtime\tool_prefix\bin;C:\Users\Lucky star\AppData\Local\Doubao\User Data\Default\connector_runtime\tool_prefix;C:\Users\Lucky star\AppData\Local\Doubao\User Data\Default\connector_runtime\artifacts;C:\Users\Lucky star\AppData\Local\Doubao\User Data\Default\sandbox_envs_dir\envs\bd96021b-db6a-4205-8dd2-151d36fedec9\override_dlcs\bin;C:\Users\Lucky star\AppData\Local\Doubao\User Data\Default\sandbox_envs_dir\envs\bd96021b-db6a-4205-8dd2-151d36fedec9\override_dlcs;C:\Users\Lucky star\AppData\Local\Doubao\User Data\Default\sandbox_envs_dir\envs\bd96021b-db6a-4205-8dd2-151d36fedec9\override\bin;C:\Users\Lucky star\AppData\Local\Doubao\User Data\Default\sandbox_envs_dir\envs\bd96021b-db6a-4205-8dd2-151d36fedec9\override;C:\Users\Lucky star\AppData\Local\Doubao\User Data\sandbox_runtime\bases\d0ace100fdff5b2d247847aca1e3d83b\node_modules\.bin;C:\Users\Lucky star\AppData\Local\Doubao\User Data\sandbox_runtime\bases\d0ace100fdff5b2d247847aca1e3d83b\python\Scripts;C:\Users\Lucky star\AppData\Local\Doubao\User Data\sandbox_runtime\bases\d0ace100fdff5b2d247847aca1e3d83b\python;C:\Users\Lucky star\AppData\Local\Doubao\User Data\sandbox_runtime\bases\d0ace100fdff5b2d247847aca1e3d83b\node;C:\Users\Lucky star\AppData\Local\Doubao\User Data\sandbox_runtime\bases\d0ace100fdff5b2d247847aca1e3d83b\bin;C:\Users\Lucky star\AppData\Local\Doubao\User Data\sandbox_runtime\bases\d0ace100fdff5b2d247847aca1e3d83b;D:\Arkdobau\PYTHON 3.11.09\Scripts\;D:\Arkdobau\PYTHON 3.11.09\;C:\WINDOWS\system32;C:\WINDOWS;C:\WINDOWS\System32\Wbem;C:\WINDOWS\System32\WindowsPowerShell\v1.0\;C:\WINDOWS\System32\OpenSSH\;C:\Program Files\dotnet\;C:\Program Files (x86)\NVIDIA Corporation\PhysX\Common;D:\openclaw\Git\cmd;C:\Program Files\nodejs\;C:\Users\Lucky star\AppData\Local\pnpm;C:\Users\Lucky star\AppData\Local\Microsoft\WindowsApps;C:\Users\Lucky star\AppData\Local\Programs\Ollama;C:\Users\Lucky star\.bun\bin;C:\Users\Lucky star\AppData\Roaming\npm;C:\Users\Lucky star\AppData\Local\Doubao\User Data\sandbox_runtime\bases\d0ace100fdff5b2d247847aca1e3d83b\pwsh、HF_HOME=\hf_cache,wzf\python.exe -m scripts.inference --unet_config_path configs/unet/stage2.yaml --inference_ckpt_path checkpoints/latentsync_unet.pt --guidance_scale 1.5 --video_path <無聲主檔> --audio_path <TTS mp3> --video_out_path <out> --seed 1247 --inference_steps 20(cwd=)。
- 輸出:h264+aac 16kHz mono,時長=音訊長度(故 L02 台詞 9.61s 對嘴後影片 9.6s,不需 atempo 加速);音量 mean≈-26dB 偏低需後製 loudnorm=I=-16:TP=-1.5:LRA=11。
- 對 6 段(0007/FastH3 × L02/L03/L05)各約 35–56s 曾跑通,但輸出嘴巴不自然(面具偽影),用戶於 2026-10-18 明確否決 → 此段(含先前「無面具偽影」判斷)作廢,LatentSync 列為禁區。
- 後製:ffmpeg drawtext(OHIYOYO x=8:y=4 fontsize=26 boxcolor=0xF7C6D0@0.95 boxborderw=8;字幕 y=h-130 fontsize=28 白字黑邊;中文逗號轉 \,)+loudnorm。腳本 tmp\gen_compare_final_ls.py。