Maria Mizuse 水瀬マリア | Tomodachi Game

Maria Mizuse from Tomodachi Game

I hadn't noticed her hair and eye colors were different in half of the anime until I started working on this model. Haha!

Prompts:-

• Trigger Word:
  maria_olis

• Outfit 1:
  pink hair, hair ribbon, green eyes, neck ribbon, white shirt, black jacket, black skirt, garter straps, thighhighs

• Outfit 2:
  brown hair, hairclip, blue eyes, black serafuku, red neckerchief, name tag, pleated skirt, black thighhighs

• Extra:
  1girl, solo
  (please refer to my sample images)

Weight: 1 (you may experiment yourself)

All my models were trained with 512x512 pixels.****For optimal results, please upscale your generated outputs with your preferred methods.

My seventh model.

This time, I have decided to use the D-Adaptation optimizer, which can automatically choose a learning rate for me. In the past, I experimented with it but didn't like the results. However, my recent models were getting overcooked quite easily with the default learning rates. I also tried different learning rates but couldn't decide on the best ones. Therefore, I decided to give D-Adaptation another try.

I realized that I had wasted too much time on stuff I couldn't decide. Once again, I trained over 50 models for a single character. Perhaps I should just simplify my workflow, probably into three parts: preparing the dataset, training a few models with varying numbers of epochs using D-Adaptation, and finally choosing the "best" model by comparing the generations from the last epochs across multiple batches.

Btw, I noticed that I mistakenly disabled aspect ratio bucketing in my previous two models while experimenting with a batch size of 1. No wonder the models sometimes couldn't generate the character's face. Maybe I'll retrain them when I get bored.

Anyway. hopefully some of you will like this model.