mix image_vision.download_models (image_vision v0.2.0)

Copy Markdown View Source

Pre-downloads the default models used by Image.Classification, Image.Segmentation, Image.Detection, Image.Background, Image.Captioning, and Image.ZeroShot so that first-call latency is eliminated and the application can run fully offline.

By default, every category's models are fetched. Pass one or more of --classify, --segment, --detect, --background, --caption, --zero-shot to limit the scope. Any category whose optional dependency is not loaded is skipped with a notice rather than treated as an error.

ONNX weights for segmentation and detection are stored under the ImageVision.ModelCache cache root (see that module for cache configuration). Bumblebee classifier and embedder weights are stored under Bumblebee's own HuggingFace cache (controlled by BUMBLEBEE_CACHE_DIR / the standard HF cache env vars).

Usage

mix image_vision.download_models
mix image_vision.download_models --classify
mix image_vision.download_models --segment --detect
mix image_vision.download_models --background --caption --zero-shot

Configuration

The task respects user overrides for the Bumblebee classifier and embedder:

config :image_vision, :classifier,
  model: {:hf, "facebook/convnext-large-224-22k-1k"},
  featurizer: {:hf, "facebook/convnext-large-224-22k-1k"}

Configured values are downloaded; unset values fall back to the library defaults.

Options

  • --classify downloads the classifier and embedder Bumblebee models. Requires :bumblebee and :nx.

  • --segment downloads the SAM 2 and DETR-panoptic ONNX weights. Requires :ortex.

  • --detect downloads the RT-DETR ONNX weights. Requires :ortex.

  • --background downloads the BiRefNet lite ONNX weights for background removal. Requires :ortex.

  • --caption downloads the BLIP image-captioning model, featurizer, tokenizer, and generation config. Requires :bumblebee.

  • --zero-shot downloads the CLIP model, featurizer, and tokenizer for zero-shot classification. Requires :bumblebee.