SigLIP 2 base-patch16-384 (ONNX, image and text towers, fp16)

SigLIP 2 by Google (google/siglip2-base-patch16-384).

Exported for Imagent from google/siglip2-base-patch16-384 at commit f775b65a79762255128c981547af89addcfe0f88 by export_siglip.py, which is in this repo: it states the graphs' inputs and outputs, and checks each graph against its source before writing it. The weights run in fp16, with fp32 inputs and outputs. parity.json holds token ids and text embeddings from the reference tokenisation - ASCII lower-casing, as SigLIP 2 was trained, then Gemma's tokenizer - for checking other implementations. All credit for the model belongs to Google; please cite their work, not this repo.

File Size SHA-256 From
vision.onnx 186,613,790 bytes 61c4f999570ef03cdd69cbd8b7c748f6f76d4250ffc15e4986319812a72ed225 export_siglip.py
text.onnx 564,758,407 bytes f482f2763b56505e81f6f2614abc507e3f0bb33eb7e81e0d581cd2799619e02d export_siglip.py
vocab.json 6,940,807 bytes e1b2ad7cd775bc43b7573e6fda635960476bf0275ffd7c2de4f4aea6d18acabb export_siglip.py

Licence: apache-2.0, as upstream - see LICENSE.

Downloads last month

-

Downloads are not tracked for this model. How to track
Inference Providers NEW
This model isn't deployed by any Inference Provider. ๐Ÿ™‹ Ask for provider support

Collection including imagent/siglip2-base-384-onnx