Pavel Iakubovskii

btw, also observed "." and capitalized template influences the confidence quite a bit

Not sure what's up as I'm not familiar with this codebase (and no time to dig in), but for siglip what you're supposed to do is do sigmoid(zimg @ ztxt * temperature + bias)

from what you describe, I would bet the bias and/or temperature are missing?
The ground-truth reference code is https://colab.research.google.com/github/google-research/big_vision/blob/main/big_vision/configs/proj/image_text/SigLIP2_demo.ipynb

Hey @giffmana , temperature and bias are applied under the hood, see

Siglip
https://github.com/huggingface/transformers/blob/17792556b21b4da0dbb9e4b59b39fb34aae4047c/src/transformers/models/siglip/modeling_siglip.py#L1411-L1417

Siglip2
https://github.com/huggingface/transformers/blob/17792556b21b4da0dbb9e4b59b39fb34aae4047c/src/transformers/models/siglip2/modeling_siglip2.py#L1459-L1465