SigLIP

SigLIP (Sigmoid Loss for Language-Image Pre-Training), published by Google DeepMind in March 2023, keeps CLIP's contrastive image-text paradigm but replaces the softmax contrastive loss with a per-pair sigmoid loss. The change eliminates the expensive global batch normalization, scales cleanly to huge batch sizes, trains more stably at small batches, and generally matches or beats CLIP at similar size.

It became the default vision encoder of the modern open VLM stack: SigLIP backbones power PaliGemma and PaliGemma 2, and variants range from ViT-B/16 up to SoViT-400M, including multilingual i18n versions. SigLIP 2 (February 2025) added multi-resolution training, captioning-based pretraining, and self-distillation. All released under Apache 2.0.

Use it as a drop-in CLIP replacement for zero-shot classification, image-text retrieval, and VLM vision towers.