Nerura from Akiba Meido Sensou (Akiba Maid War)
Update v1b (3rd August 2023):
Retrained the model with a bit more anime art style.
(At first, I wanted to quickly retrain the model with the anime style. However, I realized that training characters with only a single outfit could make the model too inflexible, especially if I aim to retain higher outfit accuracy. Therefore, I decided to experiment with regularization images and also use the OUTALL blocks for further training.)
Manually removed image backgrounds from the dataset instead of trying to do so with differential training.
Experimented with Prodigy.
P.S. The badge on her dress and the legwear are still not accurate.
P.S.2 My 13-year-old hard disk is dead. I won't regather the datasets of the old models and retrain them unless there are real reasons to do so.
P.S.3 Due to RL reasons, such as the need to get a fking job, I will have less time to have fun with model training.
Prompts (v1b):-
Trigger Words:
nerura_olis, blue hair, single braid, side braid, hair bow, pink eyes
Outfit:
blue headwear, detached collar, neck ribbon, frilled dress, wrist cuffs, asymmetrical legwear
Extra:
1girl, solo
(please refer to my sample images)
Weight: 1 (you may experiment yourself)
Anime Style Model (v1):
I trained a anime style model for experimentation. It is a LoRA-C3Lier/Locon, but it seems to work properly without an extension? (tested on AUTOMATIC1111 WebUI v1.3.2) You can use both the character model and style model together to replicate her in anime style.
All my models were trained with 512x512 pixels.
For optimal results, please upscale your generated outputs with your preferred methods.
My tenth model.
I continued my experimentation with differential training. This time, I attempted to reduce the influence of the image backgrounds of the dataset by doing differential training with a style model trained specifically on those backgrounds. The results were positive, I guess. Besides, I once again did some simple block weight extraction, and trained an anime style model, aiming to minimize the anime style of the character model with differential training, followed by finetuning...blah blah blah
So, I have complicated my workflow again, and I don't think this is worth the time and effort. Many others can train better models without all these complexities.
Anyway, I found this process enjoyable. Hopefully some of you will like the models.