Model downloads (Hugging Face Hub)

Lens's local backends and curator's converters fetch weights from the Hugging Face Hub on first use and cache them under HF_HOME. Mount that cache into containers so the download happens only once.

Used by

Package	What for	Reads
argus-lens (opens in new tab)	WD14 (`SmilingWolf/wd-vit-tagger-v3`), RAM++, Florence-2 (`florence-community/Florence-2-base`) and BLIP-2.	`[local]`, `[wd14]`, `[torch]`, `HF_HOME`, `WD14_MODEL_REPO`, `WD14_MODEL_DIR`, `RAM_MODEL_DIR`, `HF_TRUST_REMOTE_CODE`
argus-curator (opens in new tab)	The `*-convert` commands download and convert MiVOLO, Places365 and Fashionpedia.	`mivolo-convert`, `places365-convert`, `garments-convert`
argus-forge (opens in new tab)	The training image downloads the base checkpoint. Mount the HF cache to keep it.	`HF_HOME`

Official guides

Minimal setup

Pick a cache location and set HF_HOME (default ~/.cache/huggingface). In containers, mount it as a volume so each model downloads once.
Install the extras for the backends you want:
- argus-lens[wd14]: WD14 on CPU onnxruntime, no torch.
- argus-lens[torch]: RAM++, Florence-2, BLIP-2.
- argus-lens[local]: all of them.
For a gated model, create a read token and export HF_TOKEN. The Hub client picks it up from the environment.

Lens loads florence-community/Florence-2-base by default, which needs no remote code. The original microsoft/Florence-2-base runs only with HF_TRUST_REMOTE_CODE=1. See trust_remote_code and Florence-2 (opens in new tab) before turning that on.