SigLIP 2: A better multilingual vision language encoder

Hugging Face Blog · 1 year ago

Google released SigLIP 2, an improved family of multilingual vision-language encoders that extend the original SigLIP training approach with additional objectives for semantic understanding, localization, and dense feature extraction. The release includes four base variants at patch16-256 resolution plus dynamic resolution "naflex" variants, with the largest model containing 1 billion parameters. SigLIP 2 outperforms the original SigLIP across zero-shot classification, image-text retrieval, and transfer learning tasks, making it suitable for downstream applications including Vision-Language Models like PaliGemma.