Generated without specifying a specific face detail.
Danbooru - Character Tags
Diving deep into the abyss of AI art experimentation
Preamble
Since NovelAI's image generator is fine tuned using Danbooru images and tags, I wanted to see which ones it is able to reproduce. I went to the Danbooru wiki and extracted all the different face tags listed.
This is the same experiment that I did with the previous model, which can be found in the legacy experiments, but updated to use the V4 model. Also, I've tweaked the prompt since the V4 requires a bit of a firmer hand to give consistent results sometimes.
Generation
As opposed to my legacy pictures where I generated by hand, I wrote a Python script utilizing the NovelAI API to generate all these images. I first locked down a very generic prompt to allow the character tag to influence the image as much as possible and then set it up to iterate through each and every one of the tags. The script is constructed in such a way that it requires my input every time it generates a new image. So there have been a whole lot of button pressing. This is due to the NovelAI ToS, 9.1.5, which I have also confirmed with the NovelAI team.
Prompt:
girl, mature female, 1.25::______::, 1.5::arms at sides, portrait::, black hoodie, blonde hair, medium hair, blue eyes, white background, small breasts
Model used is NAI Diffusion V4 Full, Steps 28, Guidance 5.5, Quality Tags Enabled, UC Preset Enabled, Seed 2320725169.
Footnote
Some things to note:Candycane?
Nope. That's just some censoring of mine.
Weird casing on image titles!
NovelAI's V4 model is based on T5 and therefore the prompt is case sensitive1 However, all tags from Danbooru are kept in lowercase for consistency.
That looks weird!
Since everything is generated using the very same seed, there are quite a few tags that have the potential to look good but just don’t look good here. Keep that in mind if you use these tags to generate your own prompts. Rather than looking for what actually looks good, look for a style you want and then try it out with other seeds or using your own prompt.
