Dataset used for training is the stage3-p480-800 in CyberHarem/yuki_touhou, which contains 178 images.
Batch size is 4, resolution is 720x720, clustering into 5 buckets.
Batch size for regularization dataset is 16, resolution is 720x720, clustering into 20 buckets.
Trained for 1800 steps, 40 checkpoints were saved and evaluated.
Trigger word is yuki_touhou.
Pruned core tags for this waifu are blonde_hair, hat, yellow_eyes, short_hair, bow, ribbon, hat_bow, black_headwear. You can add them to the prompt when some features of waifu (e.g. hair color) are not stable.
How to Use It?
If You Are Using A1111 WebUI v1.7+
Just use it like the classic LoRA. The LoRA we provided are bundled with the embedding file.
If You Are Using A1111 WebUI v1.6 or Lower
After downloading the pt and safetensors files for the specified step, you need to use them simultaneously. The pt file will be used as an embedding, while the safetensors file will be loaded for Lora.
For example, if you want to use the model from step 720, you need to download 720/yuki_touhou.pt as the embedding and 720/yuki_touhou.safetensors for loading Lora. By using both files together, 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 720.
1520 images (1.39 GiB) were generated for auto-testing.
Metrics Plot
The base model used for generating preview images is Meina/MeinaMix_V11.
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:
Individuals who cannot tolerate any deviations from the original character design, even in the slightest detail.
Individuals who are facing the application scenarios with high demands for accuracy in recreating character outfits.
Individuals who cannot accept the potential randomness in AI-generated images based on the Stable Diffusion algorithm.
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.
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:
