[Feature] DINOv3 integration

[MkengineTA]

I have searched the existing feature requests, both open and closed, to make sure this is not a duplicate request.

• Yes

The feature

I would like to propose adding support for the new DINOv3 vision models from Meta AI for the Smart Search feature.

DINOv3 represents the current state of the art in vision models. They generate extremely high-quality "dense features," which could lead to significantly more precise and contextually relevant search results.

To start, support could be implemented for one of the smaller, yet still very powerful, models to evaluate the benefit. Good candidates would be:

facebook/dinov3-vits16-pretrain-lvd1689m

facebook/dinov3-convnext-tiny-pretrain-lvd1689m

Resources:

GitHub Repo: https://github.com/facebookresearch/dinov3

Hugging Face Collection: https://huggingface.co/collections/facebook/dinov3-68924841bd6b561778e31009

Thank you for taking the time to consider this proposal.

Platform

• Server
• Web
• Mobile

[bo0tzz]

I'm not well versed in these models, but from a quick look my impression is that this is a very general base model, and not something we could meaningfully use as-is in Immich. Access to it also seems pretty restricted, requiring people to share a bunch of personal information with Meta. cc @mertalev for thoughts.

[mertalev]

They don't necessarily need to be fine-tuned (the models in our catalog haven't been either). That being said, I'm not sure if DINOv3 is actually an improvement over SigLIP2 for retrieval purposes.