Current Anima Lora:

The current Anima version only contains the girls of the "present" in the anime. I needed to get back into it with a smaller project with all the new models so only Nyu, Lucy, Nana, Mariko, Mayu, and Yuka have been trained.

Unlike earlier versions, I used a completely different approach to creating the dataset. Using Qwen models and reference images I created a small high resolution dataset rather than a large cropped dataset like in the past. Creating a dataset independent of the source material means that even at full strength it will not hard force the render into an anime style but the characters when using an anime model do come out near 1:1.

Older versions:

This is the reason I started trying to make LORA's. The anime may not have been for everyone, but it's been my favorite since I first saw it.

It was trained on about 275 images taken directly from series stills focusing on many of the main characters and a few of the supporting characters also. With that large pool of images it has a habit of leaning towards the same style as the anime but with some work it's still quite versatile.

Sometimes needs helper prompts to correct hair color/length and eye color. Start without them and add if needed. Often needs (cat ears, fur) in negative prompt, most checkpoints think the Diclonious horns are ears.

The (Sweater) trigger is the one Mayu wears.

The (MaskedLucy) trigger does not work well. There were very few scenes to pull that look from, but I felt it necessary to at least try. Maybe you'll have better luck.

V9 - was the ninth time I attempted... it kinda worked and is workable... mostly.

V3 - As in the 3rd release, was trained on all the same images and captions as the previous version but using the methods I learned making 3 of my other LoRA's. It is both better and not better than it's predecessor. It works better for semi-realism but not as well for "less" semi-realism.