siglip2

[sailfish009]

Replace old models with new ones to fine-tune training.

https://huggingface.co/google/siglip2-so400m-patch14-384
https://arxiv.org/pdf/2502.14786

Backward compatibility: SigLIP 2 is de-signed to be backward compatible with
SigLIP by relying on the same architecture. This allows existing users to simply swap out
the model weights and tokenizer (which is now multilingual) to get improvements on a
wide range of tasks.

[Mactor2018]

I have pretrained, finetuned and tested on my VQA task and got better performance compared to siglip, with nearly 1% increase.