Scripting and the command line is an option. It may seem a bit daunting at first but can be a lot faster. Doing AI training stuff I tried brute forcing for awhile but it is just too tedious. A model like a Qwen 3 can be instructed to act like a photo captioning model. It is quite good at optical text recognition in images too.

You can script it to caption the whole image or any number of elements like color, location, subjects, etc., while also limiting the length of text. Then modify the image metadata with the text. It is also possible to feed it a set of json keys and it will fill in the values for the dictionary.

If you go this route, then you start using tools like ripgrep and it becomes possible to find and manipulate many thousands of images quickly to find niche sets in the tens to hundreds in just a single scripted command. You can do stuff like grep -ril foo | xargs | sed ‘s/foo/bar/’ or something like that. Not at my comp right now to pull up the command flags, but that is how to search and edit to change all instances of words in a bunch of files. That one is technically incorrect but it makes far more intuitive sense than alternatives.

Something like Qwen requires enthusiast level hardware, or if you have to, a free google colab instance. If you search for captioning models there are much smaller ones like the ancient BLIP that are less accurate but fast and run on anything. Anyways, I know this reply is not exactly what you wanted, but it is an option, and one I thought was beyond me until I tried it. Hopefully it is useful digital neighbor.