SigLIP Base (Patch16-224)

SigLIP (google/siglip-base-patch16-224) is a zero-shot image classification and image-text retrieval model that uses a sigmoid loss function for language-image pre-training, introduced in the ICCV 2023 Oral paper "Sigmoid Loss for Language Image Pre-Training". Unlike standard contrastive learning with softmax normalization, the sigmoid loss operates solely on image-text pairs and does not require a global view of pairwise similarities for normalization, enabling scaling to larger batch sizes while maintaining strong performance at smaller batch sizes.

Model Description

SigLIP follows the CLIP multimodal architecture but replaces the softmax-based contrastive loss with a pairwise sigmoid loss. This design decouples batch size from loss computation, allowing the model to benefit from a batch size of 32k without diminishing returns (pushed to 1 million in experiments). The base variant is pre-trained at 224x224 resolution on the English image-text pairs of the WebLI dataset.

Training Details

Images are resized to 224x224 and normalized with mean 0.5 and standard deviation 0.5 per channel. Text is tokenized and padded to a fixed length of 64 tokens. The base model was trained on 16 TPU-v4 chips for three days. A related SigLiT configuration (SigLIP with Locked-image Tuning) achieved 84.5% ImageNet zero-shot accuracy using only 4 TPUv4 chips in two days.

Benchmark Results

The following comparison of SigLIP versus CLIP is taken from the paper:

Further analysis in the paper demonstrates that the sigmoid loss maintains effectiveness at both small and large batch sizes, with optimal performance at 32k pairs.