Crjstobal/siglip2-base-patch16-256-coreml

SigLIP 2 base/16 @256 — Core ML (int8)

Core ML conversion of google/siglip2-base-patch16-256 for on-device image–text search on Apple silicon. Used by the Organizer in Awesome Video Studio (search your footage by what is seen).

Files

File	What	Size
SigLIPImage.mlpackage	Image encoder. Input image 256×256 RGB, output embedding (768, L2-normalised)	89 MB
SigLIPText.mlpackage	Text encoder. Input input_ids int32 [1, 64], output embedding (768, L2-normalised)	270 MB

The tokenizer is not included: use the original files from google/siglip2-base-patch16-256 (tokenizer.json, tokenizer_config.json, special_tokens_map.json). Text is lower-cased, encoded without <bos>, followed by </s> and padded with <pad> (id 0) to 64 tokens. Images are resized to 256×256 without cropping.

Changes from the original (Apache 2.0 §4)

• Converted from PyTorch to Core ML with coremltools (fp16), image and text towers exported as separate models, embeddings L2-normalised in the graph.
• Weights quantised to int8 (linear symmetric, per-channel).
• No retraining or fine-tuning. Cosine similarity checked against the PyTorch model during conversion.

License

Apache License 2.0, same as the original model (see LICENSE). Original model © Google LLC.