SigLIP 2 So400m

google/siglip2-so400m-patch16-naflex is a zero-shot image classification and vision-language encoding model that extends the SigLIP pretraining objective with captioning-based pretraining, self-supervised losses (self-distillation, masked prediction), and adaptive resolution handling for improved semantic understanding, localization, and dense features. It supports multiple resolutions while preserving the input's native aspect ratio via the NaFlex (non-uniform aspect ratio flexible) variant.

SigLIP 2 models are trained on the WebLI dataset (Chen et al., 2023) using up to 2048 TPU-v5e chips. The So400m/16 variant contains 400 million parameters and uses a Gemma tokenizer with a 256k vocabulary. The model can be used for zero-shot image classification, image-text retrieval, or as a vision encoder for vision-language models (VLMs).

Evaluation Results

The following table from the SigLIP 2 paper reports zero-shot ImageNet accuracy and COCO retrieval scores for the So400m/16 variant at different input resolutions:

• Zero-shot ImageNet accuracy: 83.4% at 256px, 84.1% at 384px, 84.3% at 512px.
• COCO Text→Image retrieval: 56.0 at 512px.
• COCO Image→Text retrieval: 71.3 at 512px.

Compared to the original SigLIP, SigLIP 2 shows consistent improvements across zero-shot classification, retrieval, and dense prediction tasks, with stronger multilingual understanding and improved fairness due to de-biasing techniques in the training data mixture.