siglip2-base-patch16-512-vision

Vision tower of google/siglip2-base-patch16-512

Note: This is a stripped-down copy of google/siglip2-base-patch16-512 containing only the vision tower. It encodes images / video frames into the shared embedding space and is meant to be loaded on its own at indexing time, separately from the other tower. Splitting the towers lets you load only the half you need, saving memory and load time.

The text tower (used at retrieval time) lives in Veritone/siglip2-base-patch16-512-text.

---

Original model card

SigLIP 2 Base

SigLIP 2 extends the pretraining objective of SigLIP with prior, independently developed techniques into a unified recipe, for improved semantic understanding, localization, and dense features.

Intended uses

You can use the raw model for tasks like zero-shot image classification and image-text retrieval, or as a vision encoder for VLMs (and other vision tasks).

Here is how to use this model to perform zero-shot image classification:

from transformers import pipeline

# load pipeline
ckpt = "google/siglip2-base-patch16-512"
image_classifier = pipeline(model=ckpt, task="zero-shot-image-classification")

# load image and candidate labels
url = "http://images.cocodataset.org/val2017/000000039769.jpg"
candidate_labels = ["2 cats", "a plane", "a remote"]

# run inference
outputs = image_classifier(image, candidate_labels)
print(outputs)

You can encode an image using the Vision Tower like so:

import torch
from transformers import AutoModel, AutoProcessor
from transformers.image_utils import load_image

# load the model and processor
ckpt = "google/siglip2-base-patch16-512"
model = AutoModel.from_pretrained(ckpt, device_map="auto").eval()
processor = AutoProcessor.from_pretrained(ckpt)

# load the image
image = load_image("https://huggingface.co/datasets/merve/coco/resolve/main/val2017/000000000285.jpg")
inputs = processor(images=[image], return_tensors="pt").to(model.device)

# run infernece
with torch.no_grad():
    image_embeddings = model.get_image_features(**inputs)    

print(image_embeddings.shape)

For more code examples, we refer to the siglip documentation.

Training procedure

SigLIP 2 adds some clever training objectives on top of SigLIP:

1. Decoder loss
2. Global-local and masked prediction loss
3. Aspect ratio and resolution adaptibility

Training data

SigLIP 2 is pre-trained on the WebLI dataset (Chen et al., 2023).

Compute

The model was trained on up to 2048 TPU-v5e chips.

Evaluation results

Evaluation of SigLIP 2 is shown below (taken from the paper).

BibTeX entry and citation info

@misc{tschannen2025siglip2multilingualvisionlanguage,
      title={SigLIP 2: Multilingual Vision-Language Encoders with Improved Semantic Understanding, Localization, and Dense Features}, 
      author={Michael Tschannen and Alexey Gritsenko and Xiao Wang and Muhammad Ferjad Naeem and Ibrahim Alabdulmohsin and Nikhil Parthasarathy and Talfan Evans and Lucas Beyer and Ye Xia and Basil Mustafa and Olivier Hénaff and Jeremiah Harmsen and Andreas Steiner and Xiaohua Zhai},
      year={2025},
      eprint={2502.14786},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2502.14786}, 
}