Hugging Face's logo

khanon
/
lora-training

Model card Files Community
2
lora-training / kokona /README.md
khanon's picture
khanon
updates Kokona README
862db6c
preview code
|
raw
history blame
2.48 kB

Sunohara Kokona (Blue Archive)

春原ココナ (ブルーアーカイブ) / 스노하라 코코나 (블루 아카이브) / 春原心奈 (蔚藍檔案)

Note: This is an older model and does not perform as well. It is somewhat difficult to change Kokona's outfit due to poor tagging and a small dataset. I will re-train the model at some point to resolve these issues.

Download here.

Table of Contents

Preview

Kokona Preview

Usage

Use any or all of these tags to summon Kokona: 1girl, halo, animal ears, orange eyes, long hair, grey hair

For her normal outfit: china dress, vertical-striped dress, white skirt, frilled skirt, off-shoulder, jacket

Somewhat frequently, a strange, mysterious man in a black suit likes to appear out-of-frame beside her. You can banish him with 1boy in the negative prompt. Not sure why this happens as there was nothing like it in the training data.

The AI is generally pretty good at picturing her, but it struggles with her pout. Try various combinations of pout, :t, pouting and negative prompt wavy mouth to try to get it to appear correctly, or just use img2img without the LoRA. It seems to be related to the LoRA since vanilla NAI can do pouting just fine.

Her stamp is in the training data and it somewhat knows what it is, but is probably too small to be recognized by the AI.

Weights close to 1 work well. It is recommended that you only use NAI/Anything3.0 VAE, otherwise colors will be highly oversaturated.

Training

All parameters are provided in the accompanying JSON files.

  • Trained on 95 curated images, repeated 7 times (665 total images / 3 batchsize = ~222 iterations per epoch, 3 epochs)
  • 1.5e-5 unet learning rate, 1.5e-4 text learning rate
  • Dataset was tagged with WD1.4 interrogator. Shuffling was enabled.
    • kokona, blue archive were added to the start of each caption. keep_tokens=2 was set to prevent shuffling those tokens.

Revisions

  • v1 (2023-01-12)
    • Initial release.
    • Note: this is an old model from 2023-01-12 and is not as flexible as my newer models, however it still works pretty well. I'll try to re-train it soon with improved tagging and training parameters so that Kokona can be customized further.