Auto-Classification examples

[DrSpaldo]

I have swapped over to Noodle Gallery and would like to see what everyone is using Auto-Classification for and would like to see what examples people have?

[Deeds67]

@DrissiReda is a heavy user of this feature, I think he posted his prompts on discord. I'll see if I can find it when I'm home

[DrSpaldo]

Great, thanks for the post, that should tighten mine up a little. I will add them and re-scan. The first go, with the examples from the docs did not tag any of my 75000 photos, so I am not sure if something went wrong or not

[Deeds67]

Great, thanks for the post, that should tighten mine up a little. I will add them and re-scan. The first go, with the examples from the docs did not tag any of my 40000 photos, so I am not sure if something went wrong or not

The model you use is very important. I'm not home and don't know it off the top of my head, but if you join discord and search, you can find the models that people are using with good results. I'll update the documentation for this tonight.

[DrSpaldo]

I could only find one post from you on Discord.

Looks like you said; ViT-SO400M-14-SigLIP-384__webli

I have been using; ViT-B-16-SigLIP2__webli

I will also try; ViT-SO400M-16-SigLIP2-384__webli

[Deeds67]

I could only find one post from you on Discord.

Looks like you said; ViT-SO400M-14-SigLIP-384__webli

I have been using; ViT-B-16-SigLIP2__webli

I will also try; ViT-SO400M-16-SigLIP2-384__webli

Look at what RedBlood posted on discord, he experimented with a lot more models

[DrSpaldo]

Look at what @redblood posted on discord, he experimented with a lot more models

Interesting, I can't find it. Only that he mentions one:

[6:24 AM]Redblood: that's a cpu starter model, good as default for immich since they don't have heavy uses. They just do face extraction/recognition + OCR
[6:24 AM]Redblood: I use this one although I have a GPU: ViT-SO400M-16-SigLIP2-384__webli

[DrSpaldo]

I am really not sure what I am doing wrong. The auto classification isn't tagging anything. I have tried my original model. Then I tried ViT-SO400M-16-SigLIP2-384__webli. It didn't do anything, so after changing the model, I ran the smart search job on "ALL" and then ran the auto classification again. Not one photo has been tagged out of 70000

[Deeds67]

What are you using for Similarity? You need to play around with that value until it matches the results you're looking for. If you lower the number, it will match more. If you increase the number it will match less.

If you lower it too much, it will match too many unrelated things. If it's too high, it will not match enough (or nothing at all). Each model has different weights, so you'll need to play around and find what gives you the best results.

[DrSpaldo]

28, the default. But I mean, not a single photo has been tagged. So it has to be something strange?

[DrissiReda]

not necessarily everyone's library is different, the resolution, quality and content of your library is different, what I posted as examples were templates from my own library and it catches a lot, barring a few false positives. But by adding other filters like "contains known people" it will be further improved. use a more powerful model if you have the juice.

[Deeds67]

28, the default. But I mean, not a single photo has been tagged. So it has to be something strange?

It's not strange, just lower the value until you get a result that you're happy with. Those values are just a guideline for the default model. Every model is different and your library is different, so you need to tune it.

[DrSpaldo]

I am now running ViT-SO400M-16-SigLIP2-384__webli . Which is about the max (RAM wise) as I can run, it is using around 4.6GB. I will lower the values and run again