siglip2_deploy
Copyright 2025 The Google Research Authors
Copyright 2024 The HuggingFace Inc. team

This product includes software developed by Google Research, published as SigLIP
2 (https://github.com/google-research/big_vision) and described in "SigLIP 2:
Multilingual Vision-Language Encoders with Improved Semantic Understanding,
Localization, and Dense Features" (Tschannen, Gritsenko, Wang, Naeem,
Alabdulmohsin, Parthasarathy, Evans, Beyer, Xia, Mustafa, Henaff, Harmsen,
Steiner, Zhai; 2025, arXiv:2502.14786).

The modules here are extracted from transformers/models/siglip, which is
licensed under the Apache License 2.0; see the LICENSE file in this directory.
Google's own release is JAX/Flax, and the PyTorch implementation is the one whose
numbers the published PyTorch checkpoints reproduce. See PROVENANCE.md.

The authors state of their SigLIP releases: these models are not official Google
products and were trained and released for research purposes.

ON THE WEIGHTS
--------------
All fifteen of Google's published checkpoints are Apache-2.0, stated on each
model card and served ungated -- including the ten mozo does not carry. Unlike
CLIP, nothing here is inferred from silence.

THE VOCABULARY
--------------
assets/gemma_bpe.json.gz is derived from tokenizer.json as published in
google/siglip2-base-patch16-224 -- 256,000 pieces, 580,604 merge rules and 249
added tokens, re-encoded so that merges are pairs of ids rather than pairs of
strings. One vocabulary serves every variant; tokenizer.model is
byte-identical across all fifteen Google publishes (sha256
61a7b147390c64585d6c3543dd6fc636906c9af3865a5
548f27f31aee1d4c8e2, 4,241,003 bytes).

It is Gemma's vocabulary. Gemma's own weights are published under the Gemma Terms
of Use behind a gate, and google/gemma-2b, gemma-7b and gemma-2-9b-it are all
gated. Google publishes this copy elsewhere: inside every google/siglip2-*
repository, ungated, under Apache-2.0. That is an affirmative grant by the rights
holder covering the artifact actually taken, not an inference from silence. It is
written down here because it looks alarming until checked.

NOT INCLUDED
------------
The two -naflex variable-resolution variants, the sigmoid training loss, the
captioning decoder and self-distillation heads, position-embedding interpolation,
and the tokenizer's silent truncation. See PROVENANCE.md for the full list and
for why.
