nufrnd/lvc-siglip2-base-coreml

SigLIP 2 Base (Core ML) for Lossless Video Cutter

The model Lossless Video Cutter downloads for Find by Description's most accurate mode. It is Google's SigLIP 2 Base, patch 16, 256 pixels (commit 3f9f96cb90da5dbc758b01813f2f6f1aee24c1ab), converted to Core ML with Float16 weights by macos/scripts/convert_visual_search_models.py:

• Image.mlpackage: 256 × 256 RGB in, L2-normalized 768-value vector out (normalization built in).
• Text.mlpackage: 64 token ids in, L2-normalized 768-value vector out.
• tokenizer-vocab.json, tokenizer-merges.bin: the SentencePiece BPE tokenizer in the app's format.
• provenance.json: source commit, tool versions, and the cosine to the PyTorch model on real frames and sample descriptions (0.99998 or better).

Licensed under the Apache License 2.0, like the original model.