ViT-B-16 SigLIP2 256

ViT-B-16-SigLIP2-256 is a zero-shot image classification and image-text retrieval model that builds on the SigLIP 2 architecture, trained on the WebLI dataset with a contrastive sigmoid loss. It uses a Gemma tokenizer (256k vocabulary) and has 86M parameters. This model is hosted as a managed API on gigarouter, providing endpoints that return classification or similarity scores for image-text pairs without requiring local infrastructure.

Key benchmark results from the SigLIP 2 paper and Big Vision repository:

Capabilities and improvements over SigLIP 1

SigLIP 2 models outperform their SigLIP counterparts at all model scales in zero-shot classification, image-text retrieval, and transfer performance when extracting visual representations for Vision-Language Models (VLMs). The unified training recipe incorporates captioning-based pretraining, self-supervised losses (self-distillation, masked prediction), and online data curation. These additions yield significant gains on localization and dense prediction tasks. Furthermore, the diverse data mixture includes de-biasing techniques, resulting in better multilingual understanding and improved fairness.

The ViT-B-16 variant operates at a 256-pixel input resolution and is the smallest of four released sizes (86M, 303M, 400M, 1B parameters), offering a cost-effective balance of performance and inference speed.