Trained on clip skip 1 on top of ACertainty

Intermediate checkpoints can be found on huggingface https://huggingface.co/alea31415/kumabear-roukin-characters

The locon is extracted from the checkpoint

The first fours images are generated by the full checkpoint, and the remaining ones by respective model (but of course you can also use locon for roukin characters. I am just too lazy to prompt it)

Why do I train the two animes together?

I feel these two animes (light novels actually) have so much similarity that I really want to make some crossovers.

Moreover there is no reason to do single anime either. I plan to add shinmai renkinjutsushi no tenpo keiei next.

Trigger Words

KumaBear

• Atla

• Cliff

• Eleanora

• Fina

• Flora

• Gentz

• Misana

• Noire

• Shia

• Shuri

• Telmina

• Yuna

Roukin8

• Adelaide

• Beatrice

• Colette

• Sabine

• YamanoMitsuha

To get everything right you may need additional trigger words for outfits and ornaments. Here are some suggestions

• If you want to get the bear costume of Yuna you may add kigurumi, bear hood, animal hood, animal costume, hand puppet etc.

• Add Red bow for Fina/Shuri/Noire

• Add twin drill for Shia

• Add double bun for Flora

• Add scrunchie for telmina

Kumakyuu and Kumayuru are not tagged, but you may get something that look right by prompting with bears, stuffed animal etc.

Interestingly I can hardly take off the hood of Yuna during the early phase of training, but it becomes possible after longer training (actually now Yuna by default does not have hood though almost all the images of her have hood on!)

Many characters are missing from the two animes. I may update the KumaBear one at the end of the season with the following characters

Kumakyuu, Kumayuru, Lurina, Farrat (king), Kitia (queen), Karin, Sanya, Helen, Ans, Mylene, Cattleya

Dataset

KumaBear 5113

• anime screenshots 5042

• fanart 37

• official art 15

• light novel illustration 19

Roukin8 2948 (screenshots only)

Regularization ~30K

Training

• First trained for 9739 steps, resumed and trained for another 20494 steps

• clip skip 1, resolution 512, batch size 8, on top of [JosephusCheung/ACertainty](https://huggingface.co/JosephusCheung/ACertainty/tree/main)

• 2.5e-6 cosine scheduler, Adam8bit, conditional dropout 0.08

--

I don't need support or credit, but I would be glad to know that you are using the models I trained and find it useful.

Moreover, I would like to advocate for more franchise models.

You can take a look at my workflow https://github.com/cyber-meow/anime_screenshot_pipeline if you are interested.

I just want to spread the fact that there is no reason to encode a single concept in each lora.