Komowata Haruka Chibi Style版本Komowata Haruka Style LoRA Model v2 (ID: 8929)

For imitating the style of prolific visual novel SD artist Komowata Haruka.

Use

Model was trained on AOM2-nugmegmixGav2. Anything related to Anything-v3 should work. Seems to work well on other models as well, but my testing isn't exhaustive.

Please do not use this model to directly compete with Komowata Haruka. Also, play Riddle Joker.

Notes on Prompting

No activation tokens should be necessary to apply the style.

If you desire more realistic body proportions, put "chibi" in the negative prompt. If the effect is still too strong, lower the model weight.

Conversely, if you want to create SD character art consistently, put chibi in the prompt and use a model weight of around 1.

描述:

Example images

Model is fairly versatile, the images are mostly Akagi eating a riceball because it's cute.

All images except one were generated using AOM2-nutmegmixGav2. Pastelmix-fp16 was used for the other.

Negative prompts utilizes bad-artist and bad-artist-anime embeddings from https://huggingface.co/NiXXerHATTER59/bad-artist

Dataset Info

126 manually pruned images from Komowata Haruka's Pixiv and Twitter, with images cropped and divided where appropriate.

1698 manually pruned images from 22 visual novels with SD art done by Komowata Haruka, biased towards more recent works. Some cropping was applied to skew the set towards higher-detail single character images.

Tagging was done via wd14-swinv2-v2 with threshold 0.15. Irrelevant character tags were (mostly) pruned.

Training parameters

"net_dim": 128,

"alpha": 128.0,

"scheduler": "cosine_with_restarts",

"cosine_restarts": 3,

"warmup_lr_ratio": 0.1,

"learning_rate": 0.0001,

"text_encoder_lr": 1e-05,

"unet_lr": 0.0001,

"shuffle_captions": true,

"keep_tokens": null,

"train_resolution": 768,

"lora_model_for_resume": null,

"unet_only": false,

"text_only": false,

"vae": null,

训练词语: