My tests with AI Auto Tagger and Metadata Layout

[sinus]

I use my own metadata-template.
Now I "lern" a bit with AI and the Auto Tagger.

If I am not wrong, this can import 3 different things, Keywords, Description and Landmarks.
For this, IMatch created 3 different Metadata-Tags.
I hope, this makes sense. 

On my first run, IMatch did override my real keywords, what I do not want of course in a test-phase.
I was able to say IMatch, do nothing override in my other tags. 

Now I have added the 3 AI-tags at the start for my normal Layout. If there is something written, the background turns yellow, gives me attention.

Cool is, I can let add the AI several time writes somehing in these fields, and can say, override the old stuff. Hence I can let run the AI several times.
One run for one picture takes about 5 seconds. 

I use Ollama with LLaVA 7b.

***
The results are quite good. But very different. Sometimes very good, sometimes not good. 
Maybe the results in English are better as I let translate the AI in German. I have the impression, but I do not know.

If I would send such AI-created texts out of my hand, it is very clear for me, I have to check and (mostly) edit them. 
But hey, if it is now so good, how good it will be in one, two, five years? 

[axel.hennig]

But of course there are IMatch Traits. So you can have much more AI tags in your metadata template.

[sinus]

Thanks, Axel, I am still on say step 3 of 100 

[Mario]

The results are quite good. But very different. Sometimes very good, sometimes not good. 
Maybe the results in English are better as I let translate the AI in German. I have the impression, but I do not know.

This is true. Smaller models (like LLaVA 7b) are reduced in size from much larger models. And the ability to return answers in different languages is one of the things that suffers. LLaVA 13b (needs 16GB of RAM on the graphic card) is better than LLava 7b, and probably also better with translating.

But the 300 or 500 billion models OpenAI and Mistral provide via their APIs to AutoTagger are of course superior in any discipline, including returning descriptions and keywords in non-English languages. I'd recommend giving Mistral a try. It's from France and supports all European languages of course.

By default IMatch provides 3 AI Tags: AI.description, AI.keywords and AI.landmark. You can add any number of additional AI tags as explained in IMatch Traits

[sinus]

Thanks, your explanation is as usual outstanding.
Because it does explain the stuff and even gives ideas and so on.

And btw: Your help is really really super and I think, this is very rare, compared to a lot of software. I was astonishe to see such a lot of information about AI and Auto Tagger. 
The first thing, what I did, after I realise the amount of help-infos: I went to the coffee-machine and made a fine coffee (with manual portafilter) and then I started to read the stuff. And looked at your videos. 

[Mario]

I was astonishe to see such a lot of information about AI and Auto Tagger. 

That's just because AutoTagger is such a huuuge feature.
The basics are simple. If you have Ollama and the LLava model installed, the AutoTagger default settings work out of the box and you get results.

If you don't have a powerful enough graphic card, opening an account with OpenAI or Mistral is the first step.
Then the defaults for these AIs in AutoTagger will work and produce results.

I've tried to make the initial setup as easy as possible - considering that we work with the latest high-tech in this case.

Next step is maybe selecting where AutoTagger stores results (normal tags or AI tags).
I've selected defaults that don't require to learn first-time users about AI tags already, and that they can immediately see the AutoTagger results in the File Window and Metadata Panel.

Next step is playing with prompts to get "better" results for your images and needs.

That's fun and can take a few days. I've tested and provided lots of examples and background info in Prompting for AutoTagger
That's a lot of stuff to read and play with.

That's already a lot more than you can do with "other" software.

Then there is the ability to use IMatch variables in prompts makes IMatch unique and prompts so much more powerful.
Again, this is fun and you can spend a lot of time trying out variables to feed the AI data you already have to get better results.

Then there's keyword mapping and the ability to use your thesaurus.

It took me many months to understand, design, implement, test and document all of this.
I expect it will take users days or weeks to learn all of this - if they need any of this.

I guess many users will maybe need to change a prompt a bit, but then already get very useful or "good enough" results.