It was trained with a mixed captioning method: tags + NL (natural language) + tags_nl + nl_tags.

Adaln was not trained, only self attention, cross attention and mlp.

It reproduces the style and should be able to work well together with different LoRAs. I've not tested it on custom Anima checkpoints, so keep in mind that it might be weaker there (depending on how said custom checkpoint was made and how strong its baked-in style is).