Hugging Face's logo

khanon
/
lora-training

Model card Files Community
2
lora-training / hibiki /README.md
khanon's picture
khanon
adds retrained Hibiki LoRA
1350f17
preview code
|
raw
history blame
2.93 kB

Nekozuka Hibiki (Blue Archive)

猫塚ヒビキ (ブルーアーカイブ) / 네코즈카 히비키 (블루 아카이브) / 猫塚响 (碧蓝档案)

Download here.

Table of Contents

Preview

Hibiki portrait Hibiki 1 Hibiki 2 Hibiki 3

Usage

※ Hibiki's outfits are rather complicated, so I have documented all of her tags here for your reference. It may not be necessary to use all of them.

Use any or all of these tags to summon Hibiki: hibiki, halo, dog ears, bright pupils, goggles on head, black hair, dog tail

For her normal outfit: twintails, hair bobbles, engineering goggles, purple choker, multicolored jacket, black camisole, grey shirt, black shorts, fishnet pantyhose, dog tags

For her cheerleader alt: cheerleader, ponytail, crop top, wing collar, criss-cross halter, pleated miniskirt, pom pom (cheerleading)

  • Use cheerleader to strengthen or weaken the effect.
  • If the stickers appear on incorrect outfits, add star sticker, sticker on face, sticker on arm to the negative prompt.
  • If her name tag and collar appear on incorrect outfits, add wing collar to the negative prompt.

Calm expression: light smile, expressionless

Embarrassed expression: blush, embarrassed, wavy mouth, open mouth

Weapon: mortar (weapon), rocket, mortar shell, suitcase, m logo

Weight 0.9 - 1.0 should be ideal. The LoRA is slightly overfit, I'll probably retrain it with slightly fewer steps in the future. It's still a substantial improvement over the previous version and is reasonably flexible with Hibiki's outfit.

Training

Exact parameters are provided in the accompanying JSON files.

  • Trained on a set of 162 images; 82 cheerleader, 80 normal.
    • 8 repeats for normal outfit
    • 7 repeats for cheerleader outfit
    • 3 batch size, 4 epochs
    • (82*7 + 80*8) / 3 * 4 = 1619 steps
  • 0.0905 loss
  • 832x832 training resolution
  • constant_with_warmup scheduler
  • Initially tagged with WD1.4 swin-v2 model, then heavily edited
  • Used network_dimension 128 (same as usual) / network alpha 128 (default)
    • Resized to 32dim after training
  • Trained without VAE.
  • Training dataset available here.

Revisions

  • v3 (2023-02-13)
    • Retrained from scratch since the original LoRA was not very good.
    • Added many new images to the dataset, particularly for Hibiki's regular outfit.
    • Completely re-tagged the dataset. Used WD1.4 swin-v2 as a starting point, then performed extensive editing and cleanup for accuracy and consistency. Added many new tags to better describe Hibiki's outfits and expressions.
    • Uses improved hyperparamters from my more recent models.
  • v1 (2023-01-10)
    • Initial release.