Provide faces of family members for detection?

[SteveDinn]

I would like to be able to have LLM Vision be able to determine whether someone approaching my door is a resident, known person, or unknown person. Would it be possible to seed the model with photos of selected faces for it to make this determination? It would be cool if the generated text could say "Steve is home" instead of "A man in jeans and a red hoodie approaches the door".

[valentinfrlch]

Memory does that. You can set it up in the LLM Vision Settings. Memory is still in Beta at the moment and might not work correctly. Here's the relevant documentation: https://llmvision.gitbook.io/getting-started/setup/settings#memory-beta

[SteveDinn]

What is the proper format of the images provided to the memory feature? I tried paths (local to within my HA container) and URLs, but the config editor rejected both, saying they were not valid paths.

I am selecting user profile images that match the path format /config/images/[GUID]/512x512.
I've also tried /api/image/serve/[GUID]/512x512 and that same URL with the full https://domain.com/...?authSid=asdfasdfasdfasdf.
Nothing works.

[MaartenPauchet]

wouldn't mind some clarity on that subject also. Are folders relative from a certain path? where does the memory module expect to find the pictures? can we configure the path?