Hugging Face's logo

google
/
siglip2-base-patch16-224

Zero-Shot Image Classification
Transformers
Safetensors
siglip
vision
Model card Files
xet
 Community
19

question about 'model_type' in config.json

#5
by XA-hyy - opened
Discussion
XA-hyy

why param "model_type" in config.json is siglip,siglip_text_model and siglip_vision_model, not siglip2, siglip2_text_model and siglip2_vision_model？I noticed existed model class 'Siglip2Model',but still use 'SiglipModel'
qubvel-hf

Hey @XA-hyy , Siglip2Model is only for -naflex checkpoints. Other checkpoints are compatible with Siglip (v1)
XA-hyy
•
edited Mar 3, 2025

Hello @qubvel-hf ，Could you provide me a text script to reproduce the flickr30k indicator in the paper（T2I Recall@1 is 85%）? My test indicator T2I Recall@1 is only 77%(siglip2-large-patch16-384)，thanks！
JohnsonWythe

Hi, I still got a question here. I found it is not compatible with siglip v1. patch embedding uses conv2d in v1 but uses dense in v2.

Sign up or log in to comment