You can find the artist on pixiv.net by typing the japanese name.

All the image are taken from there, I did some light cropping / upscaling.

@crombobular I honestly think for style lora captioning is not very important (it is important for character / concepts / objects).

For style the model will learn what you purposefully forget to provide in the caption. So if I want it to learn a digital illustration style, I will provide all of the information about what is factually visible, but I won't mention that it is a digital illustration / anime style.

You can throw a trigger like "in the style of bikabaka." , but I don't feel like it improves the lora effects a lot, and it makes the captioning process ever so slightly more annoying.

80% of the work for models after SDXL (and its derivative) is finding good training params, and having a good balance in the images you throw at the model. As long as you do not provide outright bad captions, the loras will turn out fine.

This is an example of one of the caption for this dataset:

"A blonde fox-eared girl with fluffy orange-tipped ears sits in a blue-gray airplane seat beside the window. Black over-ear headphones rest atop her twin ponytails, while her left hand props her cheek against the sill. White off-shoulder sleeves with black cuffs drape over her arms, joined by a yellow ribbon bow at the black sailor collar and a short black pleated skirt. Golden sunlight beams from above, illuminating her curled fox tail against the seat. A compact gray MP3 player sits wedged between her thighs, as blue sky and white clouds stretch beyond the glass."

No mention of the style or medium, just a somewhat factual description of the content.