This is a preliminary version of a style LoRA based on the 1987 animated film "The Tale of Genji." It's the first LoRA I've ever tried to train, and it was trained on Anima Base using Anima TrainFlow.

Samples used Anima Aesthetic with the Anima Turbo lora set to 0.8.

Options:

Recommended Strength: 0.8 (1 is too high, lower loses details)

Activation Tag: @HG-Style (Seems to help with color and shading)

Optional Tag: Hikaru_Genji for the protagonist. There were a few other men in the training data, but 1boy tends to have a similarity.

Yasuhiro Nakura was character designer / animation designer, and the style of the LoRA is, thick-line characters with finely detailed hair, pale skin, small eyes, darkness, film grain, wide shots.

Clothing: Kariginu for males, Junihitoe for females. for single layers, you should use kimono because that's the model's language (kosode is unstable). Hakama pants, of course. Japanese Clothes and layered Japanese Clothes is probably best.

Composition: Has a tendency to follow the movie's style and go for wide shots, so put that in negatives, or really focus on close-up tagging to get a better version.

Settings: indoors or outdoors seem to be the biggest differences. You can try veranda or shitomi, but these are less stable. In some earlier versions, I tried to use palace, but it would make a palace appear in the background, so don't try it.

The movie and the book are set in the Heian Era, so one of the easiest things for the model to do is create aristocratic men and women of the period.

Cautions and Talk:

Due to captioning, training data selection, and Hikaru Genji's general vibe, it'll be difficult to get 1boy and 1girl to appear stably in the same scene here. The model easily confuses details between the two. I can probably fix this by using more solo pictures in the training data, but be sure to fully distinguish men and women characters through clothing and positioning, unless you want the mix-and-match look.

It's not too good at architecture. I tried previous versions with more of a focus on Shinden-zukuri architecture (the shitomi, the roofs, the verandas, the round pillars, etc), and Anima Base did not seem to adapt well. Since the movie and the Heian era really delighted in how weird these palaces are, it's tough to tell Anima to forget Sengoku->Edo architecture. It might be better just to prompt for Shinden-zukuri features (asking for wood floors), but I haven't given up on potentially making cleaner training data for the weirdness.

For the LoRA I've uploaded, I think I was using Prodigy as the optimizer, Rank 64 (because the style needs to be deeper than Rank 32), and I might have adjusted the learning rate to .5. I am aware that Prodigy is really, really good at frying style because of how quickly it picks up details, but I made around 8 other combinations of AdamW, Learning Rate, Rank, before coming to a relatively stable outcome here.

I intend to return and improve this model, but

I understand what other users talk about when they upload something and say "It's broken, oops," or "I don't really know why it's like this." There's a lot of different ways to mess up a LORA, and you might realize after an hour that a) the training data was bad, b) captions were bad, c) settings were bad, etc. It's an improvement, and it's stable enough and unique enough that I uploaded it.