Type | |
Stats | 17 7 2 |
Reviews | (2) |
Published | Oct 6, 2024 |
Base Model | |
Training | Epochs: 5 |
Usage Tips | Strength: 0.8 |
Trigger Words | utenaki |
Hash | AutoV2 9CC06D338A |
Revolutionary Girl Utena (少女革命ウテナ)
A shoujo manga and anime series that started with the comic in 1996, followed by the anime series in 1997. The manga was by author Chiho Saito (さいとうちほ), who was part of the Be-Papas (ビーパパス) production group responsible for the anime's production.
Utena changed a lot of things, and it came at a time when its thought-provoking and very personally interpretable writing was especially appreciated. The elements of things like an apocalypse were also big on the public consciousness with the coming millennium.
How to Use This LoRA
I trained this LoRA to capture the aesthetic of the series. I went through and took hundreds of screencaps of the remastered Blu-ray release, then distilled them down to what I thought would broadly represent the feel, or "ki", of Utena.
Originally, I trained this LoRA for 1.5, and it was really robust there, but it especially benefits from Pony's greater intelligence and depth of training possible. I redid most of the images from the original posting's gallery with the new training, on a single checkpoint for consistency. I only upscaled them, I didn't edit any, and it was even on default Latent upscaler. I used a different scheduler on some of them, as AYS doesn't always do well on backgrounds, and Pony in general sometimes struggles with not putting people in anyway. Some checkpoints will be more difficult than others, but I recommend adding the tags NO-HUMANS with a weight of 1.1 or higher at the start, and following it closely with background focus if you want to make a background image.
As you may notice in the gallery, it does have capabilities to recognize some characters, though it benefits from guidance with those characters, as you will also notice. It's a little overactive sometimes with the bindi, or forehead decoration, that a couple of characters have. This can usually be avoided by putting bindi or forehead dot in negative prompts, but it's also super easy to just edit out in a finished picture. Characters were not the main focus here, however.
Independently, it does help with creating bishounen and ikemen characters, especially when used with a complementary checkpoint such as an anime style one. The only thing I did with the images in the gallery was to upscale them, as usual, so you can see what the LoRA does without much else.
I recommend using it at a weight anywhere from 0.7-1.0. A lot will depend on how well the checkpoint and the LoRA work together.
Enjoy your own private revolution...your romantic apocalypse!
As usual, if you'd like to support my work, please feel free.