README.md

metadata

license: mit
datasets:
  - CyberHarem/saber_fatestaynightufotable
  - BangumiBase/fatestaynightufotable
pipeline_tag: text-to-image
tags:
  - art
  - not-for-all-audiences

LoRA model of Saber/セイバー/アルトリア/士剣 (Fate Stay Night [UFOTABLE])

What Is This?

This is the LoRA model of waifu Saber/セイバー/アルトリア/士剣 (Fate Stay Night [UFOTABLE]).

How Is It Trained?

• This model is trained with kohya-ss/sd-scripts, and the test images are generated with a1111's webui and API sdk.
• The auto-training framework is maintained by DeepGHS Team. The architecture of base model is is SD1.5.
• Dataset used for training is the stage3-p480-1200 in CyberHarem/saber_fatestaynightufotable, which contains 654 images.
• The images in the dataset is auto-cropped from anime videos, more images for other waifus in the same anime can be found in BangumiBase/fatestaynightufotable
• Trigger word is saber_fatestaynightufotable.
• Trigger word of anime style is is anime_style.
• Pruned core tags for this waifu are blonde hair, ahoge, green eyes, ribbon, hair ribbon, braid, blue ribbon, short hair. You can add them to the prompt when some features of waifu (e.g. hair color) are not stable.
• For more details in training, you can take a look at training configuration file.
• For more details in LoRA, you can download it, and read the metadata with a1111's webui.

How to Use It?

After downloading the safetensors files for the specified step, you need to use them like common LoRA.

• Recommended LoRA weight is 0.5-0.85.
• Recommended trigger word weight is 0.7-1.1.

For example, if you want to use the model from step 6076, you need to download 6076/saber_fatestaynightufotable.safetensors as LoRA. By using this model, you can generate images for the desired characters.

Which Step Should I Use?

We selected 5 good steps for you to choose. The best one is step 6076.

870 images (837.90 MiB) were generated for auto-testing.

Here are the preview of the recommended steps:

Step	Epoch	CCIP	AI Corrupt	Bikini Plus	Score	Download	pattern_0_0	pattern_0_1	pattern_0_2	pattern_1_0	pattern_1_1	pattern_2	pattern_3_0	pattern_3_1	pattern_4_0	pattern_4_1	pattern_5_0	pattern_5_1	pattern_6	pattern_7	pattern_8	pattern_9	pattern_10	pattern_11_0	pattern_11_1	pattern_12_0	pattern_12_1	pattern_13_0	pattern_13_1	portrait_0	portrait_1	portrait_2	full_body_0	full_body_1	profile_0	profile_1	free_0	free_1	shorts	maid_0	maid_1	miko	yukata	suit	china	bikini_0	bikini_1	bikini_2	sit	squat	kneel	jump	crossed_arms	angry	smile	cry	grin	n_lie_0	n_lie_1	n_stand_0	n_stand_1	n_stand_2	n_sex_0	n_sex_1
6076	28	0.989	0.990	0.815	0.931	Download
1302	6	0.988	0.996	0.821	0.775	Download
3472	16	0.988	0.997	0.820	0.730	Download
5208	24	0.988	0.994	0.816	0.728	Download
3906	18	0.988	0.994	0.815	0.716	Download

Anything Else?

Because the automation of LoRA training always annoys some people. So for the following groups, it is not recommended to use this model and we express regret:

1. Individuals who cannot tolerate any deviations from the original character design, even in the slightest detail.
2. Individuals who are facing the application scenarios with high demands for accuracy in recreating character outfits.
3. Individuals who cannot accept the potential randomness in AI-generated images based on the Stable Diffusion algorithm.
4. Individuals who are not comfortable with the fully automated process of training character models using LoRA, or those who believe that training character models must be done purely through manual operations to avoid disrespecting the characters.
5. Individuals who finds the generated image content offensive to their values.

All Steps

We uploaded the files in all steps. you can check the images, metrics and download them in the following links:

• Steps From 2604 to 6510
• Steps From 434 to 2170