SigLIP

Google researchers built SigLIP as a CLIP-style pair of image and text encoders trained with a sigmoid loss, scoring each image and caption pair on its own as a match or mismatch rather than normalizing across the whole batch with a softmax. That change makes training cheaper to scale and holds up better when batches are small. The encoders produce aligned embeddings for zero-shot classification and image-text search, and the shape-optimized So400m vision tower is the image backbone of Google's PaliGemma and many other open VLMs. Base, large, and So400m checkpoints at 224 to 512 pixels, including a multilingual one, load in Hugging Face Transformers, with ports in open_clip and timm. SigLIP 2 (2025) adds captioning and self-supervised objectives, broader multilingual coverage, better localization, and NaFlex variants that keep an image's native aspect ratio. Original checkpoints, Colab demos, and training configs sit in Google's big_vision JAX codebase, which accepts few outside contributions. Code and checkpoints are Apache 2.0, allowing commercial use. Useful for retrieval, tagging, and VLM builders.