SigLIP Base Multilingual

google/siglip-base-patch16-256-multilingual is a zero-shot image classification and image-text retrieval model that uses a pairwise sigmoid loss function 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.

Pre-trained on the WebLI dataset at 256x256 resolution, the model processes images resized and normalized across RGB channels (mean 0.5, std 0.5) and tokenizes text to 64 tokens. It was trained on 16 TPU-v4 chips over three days.

Key strengths

• Uses a sigmoid loss function that decouples batch size from loss computation, allowing effective training at batch sizes from small to 32k (optimal) and up to one million.
• Multilingual support via training on WebLI without language filtering.
• Presented as an Oral paper at ICCV'23.

Benchmark results

A SigLiT variant (SigLIP with Locked-image Tuning) achieves 84.5% ImageNet zero-shot accuracy when trained on only 4 TPUv4 chips for two days. The paper includes comparisons of SigLIP against CLIP across multiple benchmarks, as shown below.

This model is hosted on gigarouter as a managed, OpenAI-compatible API. No local installation or model loading is required; users send inference requests directly to the API endpoint.