Attempts to replicate Aka's style. (we did it yay - still have some things to fix for next version)
Aka is a fun assassin encountered in The Last Sovereign by Sierra Lee.
Learn more about the exceptional source game here:
https://the-last-sovereign.blogspot.com/
Version 1.0 + Alternates deprecated. Still usable for some specific results, but they require more fine-tuning and are not recommended.
Description
Restricting to Original Aka
FAQ
Comments (2)
This LoRA seems to require too much keywords to recreate the character else it looks nothing like Aka.
I would recommend when captioning that you should exclude the character properties such as ['white hair', 'short hair',]
If there are multiple outfits, use tags such ['defoutfit', 'blue defoutfit'] while excluding [red_headband, pants, green turtleneck].
Also since this is an anime character, I recommend that you use an anime checkpoint like 'AnyLora' for training. and please do not include 'and' or 'break' in the captions.
I hope this can help you improve your LoRA training.
Good luck.
Yes, I've realized this and changed my tagging approach for the next characters. Still learning how everything works. Thank you very much for the feedback.
Check out the suggested Robin LoRA which does its job much easier. I'm also working on a 3rd character and testing out different tagging methods to get more desirable results in an easier way.
There were a few other mistakes on this first attempt, besides using the wrong training model. My training parameters were also weird, making the results inconsistent. I did not use and or break in the training, though. Just needed to in the generation prompts to get what I want.
There is also no actual training data that includes white pants, so getting them to appear was tricky. The character is quite niche. Can retrain now that I have extra images to work with.
I'll get back to this one next and retag it with my findings. It should be greatly improved and easily get both outfits with 1-2 tags each.