Google DeepMind Research Releases SigLIP2: A Family of New Multilingual Vision-Language Encoders with Improved Semantic Understanding, Localization, and Dense Features

Google DeepMind's SigLIP2 introduces new multilingual vision-language encoders with enhanced semantic understanding, localization, and dense feature extraction. Built on Vision Transformers, it enhances compatibility without requiring system overhauls. Key improvements include a switch to sigmoid loss, better image captioning and localization through decoder-based loss, and the NaFlex variant to maintain native aspect ratios. SigLIP2 shows notable performance gains across various benchmarks, emphasizing fair representation and multilingual support.