SigLIP Base

google/siglip-base-patch16-512 is a zero-shot image classification model that uses a sigmoid loss function for language-image pre-training (SigLIP). Unlike standard contrastive learning with softmax normalization, the sigmoid loss operates on individual image-text pairs and does not require a global view of pairwise similarities for normalization. This design simultaneously enables scaling to larger batch sizes while improving performance at smaller batch sizes.

Overview

The model is based on the CLIP architecture but with a more effective loss function introduced by Zhai et al. (2023). It was pre-trained on the English image-text pairs of the WebLI dataset at a resolution of 512×512 pixels. Text inputs are tokenized and padded to 64 tokens; images are normalized across RGB channels with mean (0.5, 0.5, 0.5) and standard deviation (0.5, 0.5, 0.5). Training used 16 TPU-v4 chips over three days.

Performance

The paper reports that a related Locked-image Tuning variant (SigLiT), trained with only four TPUv4 chips in two days, achieves 84.5% zero-shot top-1 accuracy on ImageNet. For the standard SigLIP pre-training procedure, a batch size of 32k is sufficient; benefits diminish quickly beyond that, even when scaling to one million.

An evaluation comparing SigLIP to CLIP is shown below (source: paper Sigmoid Loss for Language Image Pre-Training).

Architecture Details

The model uses a Vision Transformer (ViT) base-sized patch-16 architecture with 512px input resolution. It is designed for zero-shot image classification and image-text retrieval tasks without requiring task-specific fine-tuning.