Benchmarking on VQA tasks

[Masaaki-75]

Hi,

Thanks for sharing this work! I ran a simple benchmarking on multiple CFP classification datasets, and found that RET-CLIP beats some other fundus photography foundation models (e.g. EyeCLIP, RetiZero, etc) in terms of classification accuracy. Congrats on that!

Apart from that, I am wondering if RET-CLIP still leads in multimodal tasks like VQA (since it is trained in a CLIP-like manner). Would be grateful if there are some pilot validation results that can share with us!

[sStonemason]

Apologies for the delayed response.

Thank you for your support! And we're really glad to contribute to the community.

Following RET-CLIP, we continued this line of work and have since developed our latest model, ReVision, which further builds on the CLIP-style pretraining paradigm. As part of that effort, we have also explored VQA tasks, which may be directly relevant to your question. Preprint versions of the corresponding papers are available below, and we hope you find them useful!

the ReVision model: https://arxiv.org/abs/2512.14499
ReVision for VQA: https://arxiv.org/abs/2603.07131