These experiments are a bit different from the others. I’ve just tried different words and Danbooru tags and then sorted them into the major form classes of English; Adjective, Adverb, Interjection, Noun, Preposition, and Verb. There are a couple of reasons for this experiment. First and foremost is that I am curious to see how they affect the final outcome if used as tags. Second, I’m working on a generator for the root page of this site which could use some extra words to spice things up.
I’ve acquired data from Danbooru’s Tag Groups as well as just googled for list of common verbs and adjectives and used that to compile a list of 1500 words. In the end I got a whole jumble of different words and I had no idea how to sort them, so I opted to sort them by word classification which sent me down a rabbit hole of learning Python and how to use NLTK (Natural Language Toolkit). However this was later scrapped for playing around with a Dictionary API. This python script also helps me sort out images that look very much like the ground truth image.
These images have been generated in several batches using the same prompt template for each and every image. The prompt is deliberately left very open, while still hand holding enough to get replicable and similar results.
All images have been made using the same settings and prompt:
Photo of a person, {{{{{______}}}}}, realism, high quality
Model used is NAI Diffusion Anime (Full), Steps 28, Scale 11, Quality Tags On, Seed 882812400, Undesired Content Low Quality + Bad Anatomy.
I’ve trusted Dictionary API completely when it comes to word classification. If you find a word sorted under a class and you can’t for the life of you figure out why, it is because that’s how the site classified it for me. Me being Swedish (and sucking utterly at grammar) is the reason why I’ve not really bothered to try and classify 1500+ words by hand. Especially since some words fit into more than one classification.