※ 以前のバージョンの説明はAbout this versionから確認してください。
※ Please refer to "About this version" for details on the previous version.
noobaiの系列、illustriousの系列を問わず新しいモデルのいくつかを軸に一からマージをし直しました。そのためV4以前とは全く違う挙動をします。
新しいモデルを取り入れたため縦長ならば少し大きいサイズの出力でも問題がありません。(横長は胴が長くなりやすいです)
I have merged several new models, regardless of whether they belong to the Noobai or Illustrious series. As a result, the behavior is completely different from V4 and earlier versions.
Due to the incorporation of new models, vertically long images can now be generated with relative stability, even at slightly larger sizes. (However, in horizontally long images, the torso tends to appear elongated.)
目が崩れやすいので余裕があればADetailerの顔への適用をオススメします。
また確認できる問題点として暗闇の描写がやや苦手です。
Since the eyes can easily become distorted, if possible, it is recommended to apply ADetailer to the face.
One noticeable issue is that the model struggles somewhat with depicting darkness.
推奨する品質プロンプト/Recommended quality prompt
masterpiece, best quality, 推奨するネガティブプロンプト/Recommended negative prompt
worst quality, low quality, realistic, mutated hands, light particles, jpeg artifacts, lowres, signature, twitter username, company name, logo, sidelighting, heart-shaped pupils, 推奨するプロンプトをまず使用し、そこから調整をしてください。
また、特定のサンプラーや特定のスケジュール、CFGスケールが7を超えるとよくないためサンプルで使用した出力設定も参考にしてください。
例としてDPM++ 2M SDE HeunではNormalやUniformのスケジュールタイプを使用すると強烈にノイズが乗ります。
First use the recommended prompt, and then adjust it from there.
Additionally, certain samplers, specific schedules, and CFG scales above 7 do not work well, so please refer to the output settings used in the samples.
For example, with DPM++ 2M SDE Heun, using the Normal or Uniform schedule types results in excessive noise.
VAE内蔵済みです。
The VAE is already integrated.
サンプルで使用した出力設定/Output settings used for the samples
Sampling method:28
Output size:1040*1520
Schedule type:DPM++ 2M SDE Heun
Sampling steps:SGM Uniform
CFG Scale:5 ※7以上は推奨しません/A CFG scale above 7 is not recommended
Random number generator source:GPU
Emphasis mode:no norm
Eta noise seed delta:31337
CLIP skip SDXL:2
サンプルを出力した環境/Environment used to output samples
reForge
RTX3060
マージに使用したモデル/Models used for merging
Magic_ILL
waiNSFWIllustrious_v120
hikariNoobVPred_120
illustriousXL10_v10に手を改善するloraをマージしたモデル/A model merged with a LoRA that improves hand depiction, based on IllustriousXL10_v10
noobai0.75をベースに画風の学習をした自作のモデル/A custom model trained on art style, based on Noobai 0.75
songmix_v201
マージに使用したモデル及び、学習のベースとなったモデルの制作者達に感謝いたします。
I would like to express my gratitude to the creators of the models used for merging and the models that served as the base for training.
merge recipe
Magic_ILL x (1-alpha) + waiNSFWIllustrious_v120 x alpha = A
1,1,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,1A x (1-alpha) + hikariNoobVPred_120 x alpha = B
0,0,0,0,0,0,0,0,0,1,0,0,0,0,0,0,0,0,0,0B x (1-alpha) + illustriousXL1.0の手改善モデル x alpha = C
0,0,0,0.4,0,0.4,0,0.4,0,0,0.4,0,0,0,0,0,0.4,0.4,0,0C x (1-alpha) + noobai0.75をベースに画風の学習をした自作のモデル x alpha = D
0,0,0.6,0,0,0.6,0,0,0.6,0.6,0,0.6,0.6,0,0,0.6,0,0,0.6,0D x (1-alpha) + songmix_v201 x alpha = E
0,0,0,0,0,0,0,0,0.2,0.4,0.4,0.6,0.6,0.2,0,0,0,0,0,0E.fp16 x (1-alpha) + C x alpha = LibrasIllustriousXL_V5
0,0.4,0.4,0.4,0.4,0.4,0.4,0.4,0.4,0.4,0.4,0.4,0.4,0.4,0.4,0.4,0.4,0.4,0.4,0.4Description
一からマージをし直したのでV2.0とは雰囲気が変わってしまっています。
今回は前のV2モデルの上位互換となるようには調整していないので、V2.0と比較し使い分けていただけると幸いです。
サンプルでは引き続き顔にADetailer(デフォルト設定)を使用しています。
I re-blended from scratch, so the feel has changed from Version 2.0.
This version is not intended as an upgrade to the previous V2 model, so please consider using them separately based on your needs.
ADetailer (default settings) is still used for faces in the sample images.
推奨する品質プロンプト/Recommended quality prompt
masterpiece, best quality, 推奨するネガティブプロンプト/Recommended negative prompt
worst quality, bad hands, realistic, bad quality, jpeg artifacts, bad ???????, scan, jaggy lines, unclear, text, signature, twitter username, blurry, white background, simple background, ※ 雰囲気の濃密さを出したい場合はネガティブプロンプトの realistic を外してください。
※ To enhance the dense atmosphere, please remove "realistic" from the negative prompt.
※ ネガティブにrealisticがある場合はbad ???????をセットで入れてください。ネガティブにrealisticがあると手足が崩れる確率が上がります。
※ When "realistic" is included in the negative prompt, please also add "bad ???????." Having "realistic" in the negative prompt increases the chances of hand and foot deformities.
サンプルプロンプトでは 1boy, male focus, と入力しても男性のように見える人物が極端に出にくくなるようなので気になる場合、下記のプロンプトを試してください。(これはillustriousの特徴のようです。)
In sample prompts, even if you input "1boy, male focus," it tends to be difficult for a male-looking character to appear. If you're concerned about this, please try the following prompt.(this seems to be a characteristic of illustrious.)
Quality prompt
best quality, Negative prompt
bad quality, bad ???????, bad hands, jpeg artifacts, lowres, scan, jaggy lines, unclear, (sketch:0.4), text focus, text, signature, twitter username, mascara, これでも男性のように見える人物が出にくいと感じる場合、一例となりますが (muscular male:0.6), 等を指定するといいかもしれません。
If you still feel that a male-looking character is difficult to generate, you might want to try specifying something like "(muscular male:0.6)" as an example.
VAE内蔵済みです。
The VAE is already integrated.
サンプルで使用した出力設定/Output settings used for the samples
Sampling method:28
Schedule type:Euler a
Sampling steps:simple
CFG Scale:6
Random number generator source:NV
Emphasis mode:no norm
Eta noise seed delta:31377
CLIP skip XL:2
サンプルを出力した環境/Environment used to output samples
reForge
RTX3060
マージに使用したモデル/Models used for merging
Illustrious-XL-v0.1
knkHelioblend V2
copycatIllustrious V2
XL_e4P_Mix
私のよく利用するコミュニティで配られたモデル/A model distributed in a community I frequently use
私が作成したillustriousXLv0.1をベースとした画風モデル(胴長対策用)/A style model I created based on illustriousXLv0.1(For addressing long torso proportions)
マージに使用したモデル及び、学習のベースとなったモデルの制作者達に感謝いたします。
I would like to express my gratitude to the creators of the models used for merging and the models that served as the base for training.
merge recipe
Illustrious-XL-v0.1 x (1-alpha) + コミュニティで配られたモデル x alpha (0,0,0,0,0,0,0,0,1,0,0,1,0.6,0,1,0,0,0,0,0) = AA x (1-alpha) + knkHelioblend_v20 x alpha (0,0,0,0,0,0,0,1,0,0,0,0,0,0,0,0,0,1,0,1) = BB x (1-alpha) + copycatIllustrious_v02 x alpha (0,0,0,0,0,1,1,0,0,0,0,0,0,0,0,0,1,0,0,0) = CC x (1-alpha) + XL_e4P_Mix x alpha (0,0.1,0,0,0,0,0,0.5,0,0,0,0,0,0,0,0,0,0.5,0,0.05) = DD x (1-alpha) + IllustriousXLv0.1に学習を施したモデル(胴長対策用) x alpha (0,0,0,1,0,0,0,0,0,0,0,0.4,0,0,0,0,0,0,0,0) = EE x (1-alpha) + Illustrious-XL-v0.1 x alpha (0,0.25,0.25,0.25,0.25,0.25,0.25,0.25,0.25,0,0.25,0.25,0.25,0.25,0.25,0.25,0.25,0.25,0.25,0.25) = FF x (1-alpha) + copycatIllustrious_v02 x alpha (0,0.25,0.25,0.25,0.25,0.25,0.25,0.25,0.25,0.25,0.25,0.25,0.25,0.25,0.25,0.25,0.25,0.25,0.25,0.25) = LibrasIllustriousXL_V3








