| --- |
| license: apache-2.0 |
| library_name: onnx |
| base_model: google/siglip2-base-patch16-224 |
| tags: |
| - clip |
| - siglip2 |
| - semantic-search |
| - camstack |
| --- |
| |
| # SigLIP2 B/16-224: CamStack build |
|
|
| Google SigLIP2 base (patch 16, 224×224) image and text encoders, re-packaged for |
| the CamStack inference pool and embedding encoder. 768-dimensional embeddings. |
| Selectable on the `clip-embedding` step as `siglip2-b16-224` (image) with the |
| text encoder `siglip2-b16-224-text`. |
|
|
| | Tower | Format | File | Precision | |
| | --- | --- | --- | --- | |
| | vision | ONNX | `onnx/camstack-siglip2-b16-224-vision.onnx` | fp32 (reference; not in the catalog) | |
| | vision | OpenVINO IR | `openvino/camstack-siglip2-b16-224-vision.{xml,bin}` | fp16 | |
| | vision | CoreML | `coreml/camstack-siglip2-b16-224-vision.mlpackage` | fp16 | |
| | text | ONNX | `onnx/camstack-siglip2-b16-224-text.onnx` | fp16 weights, fp32 I/O | |
| | text | OpenVINO IR | `openvino/camstack-siglip2-b16-224-text.{xml,bin}` | fp16 | |
| | text | CoreML | `coreml/camstack-siglip2-b16-224-text.mlpackage` | fp16, `input_ids` int32 [1,64] | |
| | tokenizer | HF `tokenizers` | `onnx/camstack-siglip2-b16-224-tokenizer.json` | — | |
|
|
| ## Source |
|
|
| - Model: `google/siglip2-base-patch16-224`, |
| <https://huggingface.co/google/siglip2-base-patch16-224>, revision |
| `75de2d55ec2d0b4efc50b3e9ad70dba96a7b2fa2`. Licence: Apache-2.0 (model card |
| `license: apache-2.0`). |
| - Code: <https://github.com/google-research/big_vision> (Apache-2.0). |
| - Paper: Tschannen et al., "SigLIP 2: Multilingual Vision-Language Encoders with |
| Improved Semantic Understanding, Localization, and Dense Features", 2025, |
| <https://arxiv.org/abs/2502.14786>. |
| - Training data: WebLI (Google internal, never released). |
|
|
| ## Modifications |
|
|
| These files are a modified version of the upstream model (Apache-2.0 §4(b)): |
|
|
| 1. **Vision input contract.** The vision graph takes RGB float in [0,1] |
| (NCHW, N×3×224×224) and applies SigLIP's `(x - 0.5) / 0.5` = `2x - 1` |
| inside the graph, ahead of the patch embedding. The output is the single |
| pooled `embedding` [N,768], not L2-normalised. |
| 2. **Text slice.** The text graph takes `input_ids` int64 [1,L] and slices it to |
| the model's 64 positions (`input_ids[:, :64]`) before the embedding lookup, so |
| a caller that pads to CLIP's 77 tokens gets the same embedding for any query |
| under 64 tokens. |
| 3. **Lower-casing tokenizer.** `camstack-siglip2-b16-224-tokenizer.json` is the upstream `tokenizer.json` with a |
| `Lowercase` normalizer prepended: the model was trained on lower-cased text |
| and the upstream processor does not lower-case. |
| 4. **Format conversion and precision.** OpenVINO IR and CoreML at fp16; the text |
| ONNX weights converted to fp16 (inputs and outputs stay int64/fp32). |
|
|
| Built and verified by `scripts/build-siglip2-model.py` in the CamStack |
| repository (https://github.com/camstack/server): cross-format cosine against the |
| fp32 reference, tokenizer ids equal to the reference processor, and the |
| image-text embedding space checked on reference images. |
|
|
| ## Licence |
|
|
| Apache License, Version 2.0 — <https://www.apache.org/licenses/LICENSE-2.0>; |
| full text in [`LICENSES/Apache-2.0.txt`](../../LICENSES/Apache-2.0.txt). The |
| upstream model is © Google LLC. CamStack's changes are listed above. Every |
| model's attribution is in [`THIRD_PARTY_MODELS.md`](../../THIRD_PARTY_MODELS.md). |
|
|
