Sigmoid Loss for Language Image Pre-Training
Paper
•
2303.15343
•
Published
•
8
SigLIP release
SigLIP improves upon CLIP with a sigmoid loss. Both English-only and multilingual checkpoints are released.
-
-
google/siglip-base-patch16-224
Zero-Shot Image Classification • 0.2B • Updated • 373k • 59 -
google/siglip-base-patch16-256
Zero-Shot Image Classification • 0.2B • Updated • 11.6k • 6 -
google/siglip-base-patch16-384
Zero-Shot Image Classification • 0.2B • Updated • 3.25k • 10 -
google/siglip-base-patch16-512
Zero-Shot Image Classification • 0.2B • Updated • 41.8k • 24 -
google/siglip-large-patch16-256
Zero-Shot Image Classification • 0.7B • Updated • 25.6k • 12 -
google/siglip-large-patch16-384
Zero-Shot Image Classification • 0.7B • Updated • 10.8k • 9
google/siglip-base-patch16-256-multilingual
Zero-Shot Image Classification • 0.4B • Updated • 18.2k • 46Note This is a multilingual variant trained on WebLi without language filtering.
google/siglip-so400m-patch14-384
Zero-Shot Image Classification • 0.9B • Updated • 2.23M • 573Note This one has a scaled-optimized architecture, obtaining the best performance in English.