SigLIP ViT SO400M 384

ViT-SO400M-14-SigLIP-384 is a contrastive vision-language model for zero-shot image classification that uses a sigmoid loss function for language-image pre-training, enabling it to classify images into arbitrary text-defined categories without task-specific fine-tuning. The model employs a Shape-Optimized 400 million parameter Vision Transformer (ViT-SO400M) with a patch size of 14 and 384x384 input resolution. It was trained on the WebLI dataset using the SigLIP approach, which replaces the standard softmax normalization in contrastive learning with a pairwise sigmoid loss. This design eliminates the need for global pairwise similarity normalization, allowing effective training at both small and large batch sizes. Key strengths include efficient training with modest compute: the SigLiT variant (SigLIP combined with Locked-image Tuning) achieves 84.5% zero-shot accuracy on ImageNet using only four TPUv4 chips trained over two days. The sigmoid loss also enables scaling experiments up to batch sizes of one million, though benefits diminish beyond 32k. The model is available under Apache 2.0 license and was presented as an Oral at ICCV 2023. The model supports both image-text contrastive tasks (via OpenCLIP) and standalone image embedding extraction (via timm). It is hosted on gigarouter as a managed API, providing OpenAI-compatible endpoints for zero-shot image classification without requiring local infrastructure.