wa Style LoRA
v4
Machi
LoRA model to help imitate the style of the artist wa, also known as washizutan, and genryusui. He's best known for the doujinshi series WASANBON, which features mainly Kantai Collection characters drawn in various full-body and close up poses. Notable characteristics of his style include high detail in characters, especially hair, while also not straying too far from the original character artist's artstyle. His art frequently features Akagi, Kaga, Hiryuu, and Souryuu, as well as some other Shibafu ships.
Use
Model was trained on AOM2-nugmegmixGav2. Anything related to Anything-v3 should work, but my testing is limited.
No activation tokens are required. I recommend model weights of 0.5-0.9. Adjusting TEnc weight separately from UNet weight can also have sometimes desirable effects on posing as well (see example image).
Due to their prevalence in the training data, this LoRA can improve feature recall for some characters (see example images).
Popularity
Info
Latest version (v4): 1 File
About this version: v4
Example Images
All images were generated using AOM2-nutmegmixGav2, with some being upscaled by Ultimate SD Upscaler (while also applying this style LoRA). Inpaint was used to fix minor error in hands and matching eye color.
Dataset Info
934 manually pruned images from Danbooru, with images cleaned, cropped and divided where appropriate to remove text and skew the set towards single character focus.
Tagging was done via wd14-swinv2-v2 with threshold 0.15.
Training parameters
"net_dim": 128,
"alpha": 128.0,
"scheduler": "cosine_with_restarts",
"cosine_restarts": 5,
"warmup_lr_ratio": 0,
"learning_rate": 0.0001,
"text_encoder_lr": 1e-05,
"unet_lr": 0.0001,
"shuffle_captions": true,
"keep_tokens": 2,
"train_resolution": 768,
"lora_model_for_resume": null,
"unet_only": false,
"text_only": false,
"vae": null,
1 Version
Go ahead and upload yours!
