SigLIP 2 So400m

google/siglip2-so400m-patch16-512 is a zero-shot image classification and vision-language encoder model that extends the SigLIP pretraining objective with captioning-based pretraining, self-supervised losses (self-distillation, masked prediction), and online data curation into a unified recipe for improved semantic understanding, localization, and dense features.

This model, part of the SigLIP 2 family, is the So400m variant with approximately 1.14 billion parameters. It is licensed under Apache 2.0. The model supports zero-shot image classification and image-text retrieval, and can serve as a vision encoder for Vision-Language Models (VLMs). It was pretrained on the WebLI dataset using up to 2048 TPU-v5e chips.

Key Strengths

• Improved performance over original SigLIP across all model scales in zero-shot classification, image-text retrieval, and transfer for VLMs.
• Significant improvements on localization and dense prediction tasks.
• Supports multiple resolutions and preserves the input's native aspect ratio.
• Trained on a diverse data-mixture with de-biasing techniques for better multilingual understanding and improved fairness.

Evaluation Results

The following table from the SigLIP 2 paper shows evaluation results for the model family:

Additional Details

• The So400m variant has 1,136,556,698 parameters (F32).
• The full model family includes four sizes: ViT-B (86M), L (303M), So400m (400M), and g (1B).
• Training incorporated captioning-based pretraining, self-distillation, masked prediction losses, and online data curation.
• Variants support multiple resolutions and preserve the input's native aspect ratio.
• Training data included de-biasing techniques for improved multilingual understanding and fairness.

For further details, refer to the SigLIP 2 paper and the SigLIP documentation.