IMatch 2025 has released and brings with it AI photo keywording. I’m sending my photos to the LLaVA 7b model via Ollama so nothing leaves my computer, and it returns alt-text ready descriptions and a set of keywords describing the image.
I am completely confused.
Photo keywording concepts1 tells me “Keywords are bits of metadata that characterize and categorize images in ways that captions don’t. …They help us find stuff. Other people will rarely even see them.”
Until now I’ve made use of IMatch’s Thesaurus feature. It is a Controlled vocabulary from which I select the keywords I want to add to my photos. The thesaurus is hierarchical. Adding the “mountain” keyword also adds “landform” and “nature”. In IMatch, this is specified as nature|landform|mountain.
The advantage of the Controlled vocabulary is the control through consistency. The disadvantage is every photo needs to be manually keyworded. The thesaurus provides a structure, but it also relies on me to add keywords in a consistent manner. Most of my 35,000+ images are not keyworded for this reason.
Finding a photo means knowing where and when it was taken, or who was there. My existing category structures help me narrow in. I’ve never been 100% comfortable with that approach. It’s not neat and complete enough for me. Why bother keywording just 20% of my photos. Searching helps me find stuff. The time cost of keywording 30,000 images means it will stay lost.
AI tagging offers a way forward. There are two approaches I can take.
- Blindly accept all keywords the AI model suggests.
- Use the AI keywords to suggest matches in the thesaurus.
I have been fighting with option 2 all day and can’t get it to work. Imagine I gave three people 100 books and asked them to categorise to their heart’s content. Each would emphasise different points across story, character, location, theme, style, genre and length. Then, if I asked them to repeat the exercise the answers would differ again. My approach of shoe-horning the AI keywords into my existing thesaurus structure hasn’t worked. There are too many new choices for my head to deal with today. I’ve taken the original problem of manually applying keywords and made it multiple times worse.
In fairness, IMatch has a keyword mapping function designed to help with this problem. I’m writing this post because I haven’t yet got my thinking right and I feel I’m being too restrictive.
The alternative is to take whatever AI gives me and use it as is. Don’t worry if there is a difference between “ruin”, “historic ruin” or “abandoned ruin” and just go with it. If I ever search for “ruin”, all three will be listed. If the word matches something in my existing thesaurus, I can use the preferred word instead. It requires me to do some mapping up-front. What confuses me is that I don’t know what the options are until I’ve queried them, and at that point they are applied to an image. Do I clean them out and try again, hoping matches will occur and the AI won’t return something different? It might well be the AI maxes out on the words it can return for the image styles I have.
If I use option 1, why bother with a thesaurus at all? Perhaps my better option is to lurk in the IMatch community forums to see what others are doing, and keep in my what I’m ultimately trying to do with keywording.