This LoRA facilitates imitating the style of Kujo Ichiso (Kujou Ichiso, kujouitiso, 玖条イチソ), an illustrator best known for their work on character art and illustrations for the browser game and media franchise Kantai Collection.

There are several complementary versions of this LoRA. If you're unsure which to get, I recommend starting with v5-NAI or v5-Counterfeit.

Use & Prompting

No activation tokens are required to use these models.

If unwanted rigging is appearing, try adding machinery to the negative prompt (sometimes I need >1.5x emphasis to get it to go away).

These LoRA were trained at 960x960 for finer detail, so they should also work well for highres fix upscaling.

v5 Series

The v5 series is mainly intended to imitate the style, although it can assist in Ichiso character feature recall.

v5-NAI model

Unless you're generating on a model very closely related to NAI, compatibility won't be quite as good as a LoRA trained directly on your preferred checkpoint, so weights between 0.4 - 0.8 are a good starting point. If combining with other LoRA models, you may need to use lower weights.

On Anything v3-1 about 0.6-0.8 is usable for initial generation with upscale at 0.6 weight with no LoRA, and around 0.4 with another LoRA. YMMV.

v5-Counterfeit model

I recommend using weights between 0.4-1.0, depending on how stylized you want it to be. If combining with other LoRA models, or using with models besides Counterfeit, you may need to use lower weights.

v10 Series

The v10 series is slightly more focused on character feature recall. However, as a result these can be harder to use with more artifacting and/or extraneous features appearing. I recommend using v5 if you don't intend to depict Ichiso ship girls.

Base/remodels forms were tagged with no overlaps. So a Yuudachi Kai Ni image was not tagged with "yuudachi \(kancolle\)", only "yuudachi kai ni \(kancolle\)". As such, you can prompt for a specific form of an Ichiso ship girl by prompting for the specific form and putting alternate forms in the negative if needed (see also: example Yuudachi image in v10-Counterfeit).

Ariake, Yuugure, and Shigure Kai San are still difficult since they have 5, 2 and 3 images in the training set respectively, and little/no representation in base model training sets. Everyone else works fairly well to very well.

v10-AnyPastel model

This model is intended to be widely compatible with many anime models, but I can't guarantee it will work on your favorite merge (AnyPastel is merged from Anything-V4.5, NAI and PastelMix). For those interested I have a comparison of some select models' performances here.

If using on AnyPastel directly try weights from 0.6-0.9.

For other models, or combining with other LoRA/etc., try starting at 0.6 and lowering if the quality is not good.

v10-Counterfeit model

Weights between 0.4-0.8 seem to work well on Counterfeit-V2.5. If combining with other LoRA/etc., you may need to use lower weights.

Dataset

About 338 images of artworks by Kujo Ichiso were scraped from Danbooru, Pixiv, Twitter, and Kancolle Wiki. Some artworks were rotated to be more upright orientation. Duplicates, or closely related images (for instance, intermediate drafts of later artworks, crops of other artworks) were pruned. Most text was cleaned or cropped out. Images containing multiple characters were selectively segmented and cleaned to create single character images where feasible.

Tagging was done via wd14-swinv2-v2 with threshold 0.15. KanColle character tags were manually curated afterwards.

描述:

If you don't particularly care about depicting an Ichiso ship girl, I recommend using v5-Counterfeit instead, as it's a bit easier to use.

Trained on Counterfeit-v2.5_pruned. Since Counterfeit is fairly different from other anime models, unusual things can happen when using LoRAs trained on other models with it, so this LoRA was trained specifically for it.

Example Images

Made using Counterfeit-v2.5_pruned. Images were produced by txt2img, with or without highres fix upscaling, then upscaling with Ultimate SD Upscale and no further processing.

Image grids are txt2img only.

Negative prompts utilize bad-artist and bad-artist-anime embeddings from https://huggingface.co/NiXXerHATTER59/bad-artist, the bad-hands-5 embedding from https://huggingface.co/yesyeahvh/bad-hands-5/tree/main, and also veryBadImageNegative v1.2-6400.

Hiryuu image uses a non-public LoRA.

Training Parameters

"net_dim": 128,

"alpha": 128.0,

"scheduler": "cosine_with_restarts",

"cosine_restarts": 1,

"warmup_lr_ratio": 0,

"learning_rate": 0.0001,

"text_encoder_lr": 5e-05,

"unet_lr": 0.0001,

"shuffle_captions": true,

"keep_tokens": 1,

"train_resolution": 960,

"lora_model_for_resume": null,

"unet_only": false,

"text_only": false,

"reg_img_folder": null,

"clip_skip": 2,

"vae": null,

训练词语: