SortaPhotos/siglip2-base-patch16-224-onnx

SigLIP 2 base (224 px), ONNX fp16 — mirror for Sorta

This repository mirrors the files the Sorta desktop app downloads for on-device photo tagging and content search. They are unmodified copies of:

• Model: Google's SigLIP 2 base, patch 16, 224 px (commit 75de2d55ec2d0b4efc50b3e9ad70dba96a7b2fa2), Copyright Google LLC, licensed under the Apache License 2.0.
• ONNX conversion: onnx-community/siglip2-base-patch16-224-ONNX (commit ba1f3b0843f24bc5417d38e19c37b287d719b2f4), 16-bit floating-point version.

The tokenizer (tokenizer.json, Gemma vocabulary) is the one in Google's repository, byte for byte.

Files

File	Size (bytes)	SHA-256
onnx/vision_model_fp16.onnx	186,039,516	a1959f7bd3993a607e48839f6d01e25b876fe76afda301b028b78eef68aabd95
onnx/text_model_fp16.onnx	564,862,230	711da56ada0a4aa11c7dd3320df741081a3cae4f0ae1b5e5c6d5b294738d0eb0
tokenizer.json	34,363,039	cb9140fae3ac5122c972d37adf83e1248471a38147ad76f8215c8872c6fd8322

config.json, preprocessor_config.json, tokenizer_config.json and special_tokens_map.json come from the onnx-community repository, for reference.

Use

• Image: resize to 224×224 (bilinear, no crop), RGB, (x / 255 - 0.5) / 0.5, input pixel_values [1, 3, 224, 224] float32.
• Text: lowercase, tokenize, append <eos> (id 1), pad with <pad> (id 0) to 64 tokens; input input_ids [1, 64] int64.
• Output: pooler_output [1, 768]; L2-normalize both sides.
• Probability: sigmoid(exp(4.724453449249268) * cosine - 16.771724700927734) (the model's logit_scale and logit_bias).

License

Apache License 2.0; see LICENSE. Credit: Google (SigLIP 2) and onnx-community (ONNX conversion). See NOTICE.