Troubleshooting

Cette page n’est pas encore traduite : voici la version anglaise.

Each entry names the symptom, the likely cause and the fix. If you see a message code such as llm-all-failed or qc.loudness, look it up in Messages and error codes.

Setup and engine

ffmpeg not found

The engine looks for ffmpeg and ffprobe on your PATH, then falls back to the static-ffmpeg package.

Install it: brew install ffmpeg.
Or point at specific binaries: VSTUDIO_FFMPEG=/path/to/ffmpeg and VSTUDIO_FFPROBE=/path/to/ffprobe.

The hardware H.264 encoder doesn’t work

VSTUDIO_H264_ENCODER (or persona export.h264_encoder) chooses the encoder: libx264 (default), h264_videotoolbox on macOS, h264_mf on Windows. The engine test-encodes the chosen one once. If it doesn’t work on this machine, everything falls back to libx264 on its own, and a failed hardware encode is re-run with libx264. Nothing to fix unless you want the hardware encoder; then check your ffmpeg build supports it.

Whisper model download is slow, or you’re offline

The first transcription downloads the model into the Hugging Face cache (~/.cache/huggingface/hub, or $HF_HOME/hub): whisper-large-v3-turbo for mlx-whisper on Apple Silicon, large-v3-turbo for faster-whisper elsewhere.

Let the first download finish once on a good connection; later runs reuse it.
To use a model you already have: VSTUDIO_WHISPER_MLX=/path/to/whisper-large-v3-turbo-mlx or VSTUDIO_WHISPER_FW=/path/to/faster-whisper-large-v3-turbo (a size name such as small also works).
Add HF_HUB_OFFLINE=1 so it never touches the network.

“font ‘…’ not found” warning

A font role (cjk, cjk-bold, serif, mono …) is missing from ~/.cache/video-studio/fonts. Anything drawn with it uses a fallback face, which is why the warning is loud.

Run ./install.sh again; it downloads the open-licensed fonts.
Or point the role at your own file in persona.local.yaml under fonts:. One face of a collection works as path/to/Fonts.ttc#N.

“model ‘…’ not found” (MediaPipe)

Face tracking, retouch, cover frame picking and matting use the MediaPipe face landmarker and segmenters. ./install.sh downloads them into ~/.cache/video-studio/models. Run it again if one is missing. Without a model, cover layouts fall back to a centred crop.

The Mac app can’t find the engine

If the app shows Demo mode instead of Ready on this Mac, it couldn’t import the engine and is running a stand-in.

Open Settings → Engine. It shows the engine folder, Python and data folder In use.
Set Engine folder (Reelfold repo) to the folder that contains lib/vstudio, and Python to an interpreter that has requirements.txt installed.
The same can come from the environment: VSTUDIO_ENGINE_PATH for the engine and DESK_PYTHON for Python. The app looks at Settings first, then those variables, then the repo it lives in.

An old cache folder

Everything is cached under $VSTUDIO_CACHE, by default ~/.cache/video-studio. Older builds wrote some of it to ~/.cache/vstudio. That folder is still read but no longer written; delete it once you no longer need it.

AI providers

“Every AI provider failed” (`llm-all-failed`)

The routed provider and all its fallbacks failed. The message lists which ones were tried.

Check what this Mac has: python3 -m vstudio.llm auth status, run from the repo’s lib folder, or Settings → AI accounts & models in the app.
Fix the first failing one (login, key, local server), or add a fallback: in the app, per task; in the persona, fallback: [codex, ollama] on the route.

A notice like llm-fallback (“Claude Code failed; Codex answered instead”) is not an error. The work was done by the next provider in the chain.

The app checks Claude Code with a one-line round trip, because its own status can still say “logged in” after the token expired.

In the app: Settings → AI accounts & models → Log in again. The login runs in a terminal inside the app.
In a terminal: run claude, then /login (or claude auth login).

The engine never sees your password or token. Nothing is lost when a run stops on an expired login: sign in, run it again, and finished stages are not redone.

Batches

A batch is slow

Run estimate before a large batch. After one pilot on an otherwise idle Mac, estimates use your machine’s measured speed (bench prints the table).
Transcription runs one at a time (whisper owns the GPU); rendering runs a few in parallel. Lower --concurrency cpu-render=2 if your Mac is also busy with other work; raise it only if it isn’t.
VSTUDIO_H264_ENCODER=h264_videotoolbox uses the Mac’s hardware encoder for every encode.
Retouching long videos: the fast preset is about three times faster than quality.

The disk is full

python3 -m vstudio.batch du --batch <batch>      # what uses the space
python3 -m vstudio.batch clean --batch <batch>   # delete regenerable intermediates

clean keeps JSON, covers, sheets and previews. For delivered client batches, cleanup-sources lists source files past their cleanup date (a dry run) and only deletes with the code it printed. Your own recordings outside the batch folder are never deleted.

Captions and cuts

Names or terms are spelled wrong in captions

Speech recognition writes names by sound.

Proofreading with entity checks runs before captions are burned: place names are matched to their standard spelling, and your glossary terms are applied the same way everywhere. Run it, don’t skip it.
Add your own fixes to persona.local.yaml: subtitles: {term_fixes: [["[Tt]runking", "chunking"]]} (a regex and its replacement). Client batches use the client’s glossary.
Pass key terms to transcription up front (asr: {prompt: "LangChain, RAG"}).
Fix a single caption in review. A change that adds or drops a spoken word is refused unless a re-hearing of that audio confirms it.

Cleanup cut too much

python -m vstudio.cleanup verify OUTPUT re-transcribes the cut. A lost content word fails with its source and output time.
Keep the edit that covers it: reply 保留 7 (or --keep 7) and apply again.
If it keeps happening, use a gentler profile (gentle keeps more pause and breath), or add words to cleanup.never_cut in the persona.

In the app, filler cuts that might be real words wait for your yes in the Inbox; answer “keep” for the ones you want.