Sigmoid Loss for Language Image Pre-Training
Paper
•
2303.15343
•
Published
•
4
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 • Updated • 323k • 25 -
google/siglip-base-patch16-256
Zero-Shot Image Classification • Updated • 44.7k • 3 -
google/siglip-base-patch16-384
Zero-Shot Image Classification • Updated • 21.1k • 8 -
google/siglip-base-patch16-512
Zero-Shot Image Classification • Updated • 44.2k • 15 -
google/siglip-large-patch16-256
Zero-Shot Image Classification • Updated • 75.5k • 9 -
google/siglip-large-patch16-384
Zero-Shot Image Classification • Updated • 12.1k • 6
google/siglip-base-patch16-256-multilingual
Zero-Shot Image Classification • Updated • 14.6k • 36Note This is a multilingual variant trained on WebLi without language filtering.
google/siglip-so400m-patch14-384
Zero-Shot Image Classification • Updated • 2.87M • 292Note This one has a scaled-optimized architecture, obtaining the best performance in English.