LucidexGroup/threadfinder-siglip2-base-patch16-224

SigLIP 2 Base (patch 16, 224) for Core ML

Core ML conversion of google/siglip2-base-patch16-224 (revision 75de2d55ec2d0b4efc50b3e9ad70dba96a7b2fa2) used by ThreadFinder, a macOS app, for on-device photo search by description. This is not an official Google release. Licence: Apache-2.0 (see LICENSE, which also states the changes made).

File	Weights	Size
SigLIP2ImageEncoder.mlpackage	palettized8	95.0 MB
SigLIP2TextEncoder.mlpackage	palettized8	282.9 MB

Both towers output a unit-length 768-dimension embedding named embedding; image-to-text similarity is their dot product. manifest.json lists every file with its size and SHA-256 and states the model contract in full.

Image tower

Input image: an RGB image of exactly 224x224. Resize the picture to that size without keeping the aspect ratio and without cropping, with bilinear filtering, as SigLIP's image processor does. The model applies the scaling itself (x / 127.5 - 1).

Text tower

Input input_ids: 64 int32 token ids from the Gemma SentencePiece model in tokenizer.model. Lowercase the text, encode it without a BOS token, keep at most 63 pieces, append EOS (id 1) and pad with id 0 to 64. SigLIP 2 was trained on lowercased text padded to 64 tokens; there is no attention mask.

Precision

The weights of the linear, matmul and gather operations are stored and computed at reduced precision; every other operation runs in float32. The conversion fails unless each tower's cosine against the PyTorch model is at least 0.99 on every synthetic drawing and caption it is checked with, with Core ML on all compute units and on the CPU only. The similarity floor computed from synthetic unrelated pairs is 0.010098.

Citation

@misc{tschannen2025siglip2multilingualvisionlanguage,
      title={SigLIP 2: Multilingual Vision-Language Encoders with Improved Semantic Understanding, Localization, and Dense Features},
      author={Michael Tschannen and Alexey Gritsenko and Xiao Wang and Muhammad Ferjad Naeem and Ibrahim Alabdulmohsin and Nikhil Parthasarathy and Talfan Evans and Lucas Beyer and Ye Xia and Basil Mustafa and Olivier Hénaff and Jeremiah Harmsen and Andreas Steiner and Xiaohua Zhai},
      year={2025},
      eprint={2502.14786},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2502.14786},
}