See our collection for all versions of SigLIP 2.

Run SigLIP 2 with Keras 3: JAX, PyTorch, or TensorFlow

kerasformers/siglip2_base_p16_256

Paper: SigLIP 2: Multilingual Vision-Language Encoders with Improved Semantic Understanding, Localization, and Dense Features (arXiv:2502.14786) · HF Papers

SigLIP 2 keeps SigLIP's sigmoid loss and adds captioning-based pretraining, self-distillation, and masked prediction for stronger dense features. It uses a 256k multilingual Gemma vocabulary, so many languages work without a separate multilingual checkpoint.

For more details on the model, please go to the upstream model card.

Pure-Keras 3 conversion of google/siglip2-base-patch16-256 for kerasformers. One implementation runs unmodified on TensorFlow / Torch / JAX.

This is a zero-shot image-text checkpoint (SigLIP2ZeroShotClassify): pass image(s) and text prompts at inference time.

✨ Quick start

import os
os.environ["KERAS_BACKEND"] = "torch"  # or "jax" / "tensorflow"

from kerasformers.models.siglip2 import (
    SigLIP2Processor,
    SigLIP2ZeroShotClassify,
)

processor = SigLIP2Processor.from_weights("kerasformers/siglip2_base_p16_256")
model = SigLIP2ZeroShotClassify.from_weights("kerasformers/siglip2_base_p16_256")

labels = [
    "a photo of a cat",
    "a photo of a dog",
    "a photo of a car",
    "a photo of a living room",
]
inputs = processor(text=labels, image_paths="your_image.jpg")
output = model(
    {
        "images": inputs["images"],
        "token_ids": inputs["input_ids"],
    }
)
print(output["image_logits"].shape)

Load any SigLIP 2 variant the same way with from_weights("kerasformers/<variant>"):

Variant	Hub
`siglip2_base_p16_224`	`kerasformers/siglip2_base_p16_224`
`siglip2_base_p16_256`	`kerasformers/siglip2_base_p16_256`
`siglip2_base_p16_384`	`kerasformers/siglip2_base_p16_384`
`siglip2_base_p16_512`	`kerasformers/siglip2_base_p16_512`
`siglip2_base_p32_256`	`kerasformers/siglip2_base_p32_256`
`siglip2_large_p16_256`	`kerasformers/siglip2_large_p16_256`
`siglip2_large_p16_384`	`kerasformers/siglip2_large_p16_384`
`siglip2_large_p16_512`	`kerasformers/siglip2_large_p16_512`
`siglip2_so400m_p14_224`	`kerasformers/siglip2_so400m_p14_224`
`siglip2_so400m_p14_384`	`kerasformers/siglip2_so400m_p14_384`
`siglip2_so400m_p16_256`	`kerasformers/siglip2_so400m_p16_256`
`siglip2_so400m_p16_384`	`kerasformers/siglip2_so400m_p16_384`
`siglip2_so400m_p16_512`	`kerasformers/siglip2_so400m_p16_512`

Tips

Set KERAS_BACKEND before importing Keras / kerasformers.
Prefer Processor.from_weights(...) so image size and tokenizer match the variant.
Map processor input_ids to model token_ids. No padding mask is required.
Tokenizer is Gemma-based (multilingual); prefer Processor.from_weights.
See SigLIP 2 docs and Loading Weights.
Community / upstream safetensors still work via the hf: prefix, e.g. SigLIP2ZeroShotClassify.from_weights("hf:google/siglip2-base-patch16-256").