>>105919568
I am a bit peeved though their prompting seems to indicate how they did captioning which is tags only and they aren't really doing that much with the natural prompting training part and instead just making sure it still retains that ability from the base model's built in descriptors to make natural prompting work. Not going to blame them too much because Lumina's team did not open source the captioner they used and making it work with Danbooru tags is probably a nightmare. I am concerned though how much of it is throwing shit at a wall and seeing what sticks.

FYI for those people that were posting stuff for the Lumina fine tunes I linked. A guideline you can try and use as a base to mix things up to get a better prompt is looking at how base Lumina 2 was captioned. The finetune should still not have wiped enough of this information away.
https://arxiv.org/abs/2503.21758
I did get better results from the Illustrious team alpha tune than you guys based on this information but not drastically so. I haven't tried the neta tune yet since I've been busy this weekend and won't be free until next week. Good luck with the experimenting.