Modified Illustrious-XL for LoRA training.
It reduces unintended style changes.
If you make a LoRA with this checkpoint, do not use "--noise_offset"!
Note: This model is not suitable for image generation. The model tends to generate ugly images.
LoRA学習用のIllustrious-XLです。
意図しない画風の変化を軽減し、精度も改善します。
これでLoRAを作る場合、「--noise_offset」は使用しないように!
また、noise offsetの効果を打ち消したため、マージ用途の使用は推奨しません。
Training information:
Finetuned Illustrious-XL by repeating the training and merging a DoRA 2 times with sd-scripts.
sd-scriptsで二回DoRAの学習とマージを繰り返すことでファインチューンされました。
Style training has not been tested, but some users reported that it can be learned successfully.
画風学習は未確認ですが、問題なく学習できるという報告があります。
Training parameters(do not use):
Optimizer: AdamWScheduleFree
Resolution: 1024(w/ bucketing)
LR: 0.0005
LR Scheduler: Constant with warmup(warmup=100)
Batch size: 2
Epoch: 1
Rank/Alpha: 16/8
Conv Rank/Alpha: 2/0.5
Debiased Estimation: True
Dataset information:
Dataset size: 5120
Gelbooruから収集した5120枚の画像で学習しました。
なお、以下のタグがあるか条件を満たす画像は除外しました。
filetype:gif, score:<0, mpixels:<1048576, tagcount:<16, \*_artifacts, adversarial_noise, greyscale, monochrome, digimon, photophop_(meidum), ai-generated, duplicate, bad_\*, off-topic, cropped, resized, reversed, rotated, third-party_edit, screenshot, tagme, real_life, watermark, 3d, koikatsu_(medium), mikumikudance, twitter_username
Training script(source code)
sd-scripts( https://github.com/kohya-ss/sd-scripts )
Notice/おしらせ:
This model is licensed under Fair AI Public License 1.0-SD( https://freedevproject.org/faipl-1.0-sd/ ).
If you make modify this model, you must share both your changes and the original license.
このモデルはFair AI Public License 1.0-SD( https://freedevproject.org/faipl-1.0-sd/ )でライセンスされています。
このモデルを改変する場合、変更点および元のライセンスを明記する必要があります。
Description
Base model is Illustrious-XL-v0.1
FAQ
Comments (3)
What settings do you use to train LoRA in Illustrious XL? I see the "noise offset" needs to be at 0, but do you have suggestions for other settings like Network DIM and others?
hi, I used your checkpoint to train lora with kohya, but the result is even worse than training with pony checkpoint. The face is totally different from my dataset image https://imgur.com/a/R2UDpDP
May I know if my kohya settings need to be different?
best training model, especially for rare concepts
Details
Files
Available On (3 platforms)
Same model published on other platforms. May have additional downloads or version variants.
