
Milk Loraシリーズ第3弾
SDXLで美しい日本人女性が出せるように調整したロラです,
前回の「CondenseMilk」がSD1.5モデル用だったため同じテーマでSDXLの使用を目指し学習しました,SDXLの潜在能力により、前回モデルより精度が向上しましたが、手や細部の崩れなどは避けられないためご理解の上使用して下さい,
学習モデルで使用したCyberrealistic XL V22で良好な結果が出ます,
トリガーワード不要,ウェイトは1.0〜0.6推奨
SamplerはDPM++2M SDE Karass推奨ですが、お好みでいろいろ試してみてください,
ベースモデルの日本人をFine tuningするという考えのためプロンプトでの"日本人"は不要のはずですが入れても問題はありません,
顔以外にLoRAの影響が出るため気になる方はLora Block Weightを使用してください,
私の活動を支援してくださる方はフォローをお願いします,
謝意
素晴らしいツールを提供してくださる、偉大な先人たち、
Lora Trainer作者の"Hollowstrawberry"氏
Cyberrealisticの作者、"Cyberdelia"氏に感謝いたします,
✣ ✣ ✣ ✣ ✣
Milk Lora Series 3rd edition
This is a Lora adjusted to produce beautiful Japanese women with SDXL.
Since the previous "CondenseMilk" was for the SD1.5 model, I aimed to learn how to use SDXL with the same theme. Due to the potential of SDXL, the accuracy has improved from the previous model, but please understand that some crumbling of the hands and details is unavoidable.
Good results are obtained with Cyberrealistic XL V22, which was used in the learning model.
No trigger words are required. Weight 1.0 to 0.6 is recommended.
Sampler DPM++2M SDE Karass is recommended, but please try various things to your liking.
The idea is to fine-tune a Japanese base model, so the "Japanese" prompt should not be necessary, but there is no problem if you include it.
Since LoRA affects things other than the face, if you are concerned about it, use Lora Block Weight.
If you would like to support my activities, please follow me.
Acknowledgements
I would like to thank the great predecessors who have provided me with these wonderful tools: "Hollowstrawberry", the creator of Lora Trainer, and "Cyberdelia", the creator of Cyberrealistic.
描述:
训练词语:
名称: SkimMilk.safetensors
大小 (KB): 124735
类型: Model
Pickle 扫描结果: Success
Pickle 扫描信息: No Pickle imports
病毒扫描结果: Success