(Down)Load FasterWhisper Model

The parenthetical in the display name is the honest part: "(Down)Load FasterWhisper Model" downloads the weights the first time you run it. There's no separate model download step in this pack's README because the loader handles it - click Queue, watch it fetch a few hundred megabytes from Systran's Hugging Face org, and from then on it loads from disk. That's a genuinely nice property in an ecosystem where "download the checkpoint, put it in the right folder, realize it's the wrong folder" is the usual ritual.

It's the front door of the whole pack. It produces the FASTERWHISPERMODEL object that FasterWhisperTranscription needs, so every transcription workflow starts here.

What you set. Two dropdowns, that's it:

• model - a choice list of every stock Whisper size (tiny, base, small, medium, large-v1/v2/v3, plus the .en English-only variants and faster-whisper's distil-* speedups and the large-v3-turbo). The dropdown is built live from faster_whisper.available_models(), so it tracks whatever the library knows.
• device - cuda, cpu, or auto. auto lets faster-whisper pick. If you have a GPU, use cuda; if you're on a laptop without one, cpu still works because CTranslate2 is efficient enough on CPU that small/medium models are usable - just slower.

How the mechanism works. The loader isn't loading from the usual ComfyUI models/checkpoints folder. It constructs faster_whisper.WhisperModel(model_size_or_path, device=..., download_root=ComfyUI/models/faster-whisper, local_files_only=False), so the first run streams the CTranslate2-converted model into ComfyUI/models/faster-whisper and caches it. CTranslate2 models are quantized int8 by default - this is why "faster" Whisper is both faster and leaner than the PyTorch original, at near-identical accuracy. One limitation worth knowing: the node doesn't expose faster-whisper's compute_type option, so you get the library default (int8_float16 on GPU, int8 on CPU) and can't force fp16 or fp32 from the graph.

Which size do you actually want? Roughly: base for quick checks, small for decent captions fast, medium for a serious language other than English, large-v3 (or the newer large-v3-turbo) when accuracy is the point and you can wait. The included example workflow ships with large-v2, which is a reasonable default for subtitle work. If you're fine-tuning a custom Whisper, drop the converted model into models/faster-whisper and this node picks it up too - the loader scans that folder and adds anything it finds, stripping the models--Systran--faster-whisper- prefix that Hugging Face's cache layout leaves behind.

Install & troubleshooting. Same as the rest of the pack - Manager search "ComfyUI-faster-whisper", or clone into custom_nodes and pip install -r requirements.txt (faster-whisper==1.0.3 and soundfile are the whole dependency list). Model files land in ComfyUI\models\faster-whisper; if you'd rather place them by hand than download, grab the faster-whisper-* CTranslate2 folders from Systran's HF org and drop them there. The classic failure is a first run with no network - you get a download error, not a missing-model error, which confuses people the first time. And the large models are ~3GB on disk, so make sure you're not downloading it onto a nearly-full drive.