Type | |
Stats | 570 |
Reviews | (116) |
Published | Aug 22, 2023 |
Base Model | |
Training | Steps: 2,580 Epochs: 2 |
Usage Tips | Clip Skip: 2 |
Trigger Words | lady_furina focalors_outfit |
Hash | AutoV2 8F62FB3B7F |
Lady Furina from Genshin Impact, dress her up however you like!
Recommended parameters
Sampler: DPM++ 2S a Karras;
Steps: 35-45;
LoRA weight: 0.85-0.9
Base model for training: MeinaMix V11
Prompting guide
lady_furina
will give you Furina herself. If you use highres fix, try not to make denoise strength higher than 0.35, or the heterochromia detail might get lost. A potential fix is to use ADetailer, write (heterochromia, dark blue and light blue eyes)
in positive prompt, adjust the weight as needed.
focalors_outfit
will give Furina her default in-game look. I recommend using it in tandem with lady_furina
, with 1.5 weight, such as 1girl, lady_furina, (focalors_outfit:1.5)
, but 1girl, focalors_outfit
would also work. I find that with lady_furina
the model is more expressive when it comes to Furina herself, but it's not conclusive.
As to other propmt tags, expression and action should get the most weight, clothing style behind that, and then background/location, perhaps 1.6, 1.3, 1.2 respectively. This is because the training images don't depict Furina with a lot of expressions, but backgrounds are rather flexible as I removed all of them before training.
You can use this LoRA on realistic style models, it will likely give you a Furina cosplayer :P
Training Method
Dataset: the dataset is separated into two concepts. lady_furina
and focalors_outfit
, all images come from in-game cutscene screenshots and fanart from Google, most of them generated by AI. I then remove the background and exaggerate her heterochromia. For images in focalors_outfit
, I fix the color of her gloves, as most of them were generated by previous AI models and the color were most likely wrong. While testing tagging mothods, some good results from the previous iterations of this model get processed and put into the dataset.
Tagging: in addition to deepbooru auto tagging, lady_furina
images get all biological feature tags removed, while keeping tags about her clothing. focalors_outfit
images get all biological feature tags and clothing tags removed. Composition tags such as full body
, camera from below
, front view
are added to mitigate potential stiffness in generated images.
Training: I tried the lion optimizer, turns out even at 1.5e-5 lr and 3e-6 text_lr it still learns too fast. Every epoch I set lady_furina
(21 images) as 30 and focalors_outfit
(11 images) as 60, but after 3 epochs the model got very stiff and stick only to what it was trained on. This model is the 2nd epoch result.
If anyone wants to share their two cents on how to improve the training procedure, please don't hesitate to comment, I am eager to learn better methods. Meanwhile for everyone else, enjoy playing dress-up with Furina :P