SigLIP (Sigmoid Loss for Language-Image Pre-training) is a vision-language model from Google that improves on CLIP's training approach. Like CLIP, SigLIP learns a shared embedding space for images and text through contrastive learning on image-text pairs. The key difference is the loss function: CLIP uses a softmax-based contrastive loss that requires computing scores across all pairs in a batch, while SigLIP uses a sigmoid loss that treats each image-text pair independently. This seemingly small change has practical consequences.

The sigmoid loss lets SigLIP scale to much larger batch sizes without the memory overhead of the full softmax matrix. SigLIP 2, released by Google DeepMind in February 2025, introduced multi-resolution training, captioning-based pre-training alongside contrastive learning, and self-distillation. SigLIP models serve as the vision encoder in PaliGemma and PaliGemma 2, meaning they form the visual backbone of Google's VLM family. Available variants range from ViT-B/16 (86M parameters) to ViT-SO400M (400M parameters), with the larger models used in production VLMs.

SigLIP is used as a drop-in replacement for CLIP wherever a vision-language encoder is needed: VLM backbones, image retrieval, zero-shot classification, and multimodal embedding generation. For practitioners fine-tuning PaliGemma or building custom VLMs, understanding that SigLIP is the vision encoder helps with debugging, choosing the right model variant, and understanding resolution and performance tradeoffs.