[stormfather]

I would try the Qwen models before LLaVa

Do you need the embeddings to be private? Or just the photos?

[msgodel]

For photo indexing I'd run CLIP directly and save on compute, no need to use a whole language model.