[SDXL] Akagi (New Year Kimono) - Kantai Collection - Kancolle | 赤城【晴着】mode - 艦隊これくしょん 艦これ

Experimental SDXL LoRA model based on the New Year Haregi Mode outfit of Akagi from Kantai Collection

I trained this model to experiment with a small dataset (18 images of outfit plus 18 character reference) to see how versatile LoRA could be, and to test Animagine XL v3-like tagging methods on a manually manageable dataset. The quality and versatility isn't as good as a model with 100s or 1000s of images trained on, but it's usable.

Recommended settings

• LoRA network weight 0.7-0.9 for Animagine XL v3, or very closely related models.

  • Depending on the relation of the checkpoint, higher weight may be needed.

• Trained on Animagine XL v3. May not work well on models not closely related to Animagine XL v3.

• I highly recommend using highres fix with this LoRA, as well as quality tags such as masterpiece, best quality, etc.

• If using with other LoRA or heavy style tags, weight may need to be lowered further if it is clashing.

Suggested prompting

New Year Haregi Mode Akagi

1girl, akagi \(kancolle\), kantai collection, brown eyes, 
red kimono, print kimono, floral print, 
hair flower, hair ornament,
obi, yellow bow,
furisode, wide sleeves, long sleeves,
white socks, footwear,

• I strongly recommend prompting for a specific style as well

• If you want to use this outfit with another character, replace the character and series tags. This works relatively well with other characters

• Angles from behind tend to look okay but go rather off-model, due to limitations of the dataset

Hairstyle

hair up,
hair down,

• Hair up doesn't work very consistently if viewing from the rear, due to limitations of the dataset

Additional useful tags

pink pouch,
fur trim, fur-trimmed kimono,
hamaya, holding arrow,
shrine, outdoors,
winter, snow,
open kimono, off shoulder,
kimono lift,

• The pouch is somewhat inconsistent, it may need increased weight.

• For generating the arrow I recommend sampling using DPM++ 2M SDE Karras with at least 64 steps for the initial generation. Euler A tends to break it apart more often.

Bonus: Onsen Akagi

1girl, akagi \(kancolle\), kantai collection, hair up, brown eyes, nude, onsen, partially submerged, towel,

• For generalization I included several nude images in training and many of them were onsen/bathing, so you can use it as a more general Akagi LoRA as well (it doesn't know her regular outfit though).