Google SigLIP 2: Multilingual vision-language encoder with dynamic resolution and improved objectives

AI Impact Summary

Google introduces SigLIP 2, a multilingual vision-language encoder that extends the sigmoid-based training of SigLIP with additional objectives for better semantic understanding, localization, and dense features. The release adds dynamic resolution (naflex) variants and a new 1B-scale options, with reported gains across zero-shot classification, image-text retrieval, and cross-model transfer when extracting visual representations for Vision-Language Models. They also describe a decoder-assisted setup and self-distillation with Global-Local and Masked Prediction losses to improve localization and local semantics, with some losses only activated late in training to save compute. For practitioners, this means upgrading to SigLIP 2 checkpoints (e.g., base/large/so400m/giant-opt, including -naflex variants) and using the Siglip2Model path in Hugging Face transformers to leverage dynamic resolution and improved encodings in VLM pipelines.