プロンプトとモデル他．Prompts, checkpoints, etc.
使用したAI
Stable Diffusion
私の画像生成法は確立されていません。そのため、検証できていないことが多く参考程度に見てください。
画像1,2,3はシードや設定情報は同一でモデル違い。モデルはカスタムでマージ比率が違うだけです。
出来上がったものは構図は何となく似ていますが別物です。
次の比較は設定です。基準となる対照画像は３、画像4はDetail Enhancer(LoRA:Add more details)を除いたもの。ちなみに、このモデルは描画力が弱いため補っています。御覧の通りモデルの差ほどではありませんが大きく絵柄は変わります
さらに画像5はDynamic Thresholding (CFG Scale Fix)を除きます。若干ですが絵柄に変化が見られますね。
参考としてカスタムモデルの親になったものを画像5と同じ条件で出力。
マージ等の条件は伏せますが、画像8と画像9が出発モデルで段階的に現行モデルに近づきます。
再現できないことの証明みたいになり、納得がいかないので、時間があるときにどうすれば望む画像に近づくのか試したいと思います。
なお、プロンプトは抜粋でウエイトや画質関係の情報は削除しています。参考程度に……
My image generation method is not established. Therefore, please see it only as a reference as I have not been able to verify it.
Images 1, 2, and 3 have the same seed and configuration information, but different checkpoints. The checkpoints are custom and the merge ratio is just different.
The resulting images are somewhat similar in composition, but different.
The next comparison is the settings. The reference control image is 3, and image 4 is without Detail Enhancer (LoRA: Add more details). Incidentally, this checkpoint is compensated for by the weak rendering power. As you can see, the difference is not as great as the difference between the models, but it is a big change.
Image 5 shows the model without Dynamic Thresholding (CFG Scale Fix). You can see a slight change in the pattern.
For reference, the parent of the custom model is output under the same conditions as in Image 5.
The merging and other conditions are not disclosed, but image 8 and image 9 are the starting models, and they approach the current model step by step.
It is like a proof that it cannot be reproduced, and I am not convinced, so I would like to try to see how I can get closer to the desired image when I have time.
Note that the prompts are excerpts and the weighting and image quality related information has been removed. For reference only: ......
画像1,2,3はシードや設定情報は同一でモデル違い。モデルはカスタムでマージ比率が違うだけです。
出来上がったものは構図は何となく似ていますが別物です。
次の比較は設定です。基準となる対照画像は３、画像4はDetail Enhancer(LoRA:Add more details)を除いたもの。ちなみに、このモデルは描画力が弱いため補っています。御覧の通りモデルの差ほどではありませんが大きく絵柄は変わります
さらに画像5はDynamic Thresholding (CFG Scale Fix)を除きます。若干ですが絵柄に変化が見られますね。
参考としてカスタムモデルの親になったものを画像5と同じ条件で出力。
マージ等の条件は伏せますが、画像8と画像9が出発モデルで段階的に現行モデルに近づきます。
再現できないことの証明みたいになり、納得がいかないので、時間があるときにどうすれば望む画像に近づくのか試したいと思います。
なお、プロンプトは抜粋でウエイトや画質関係の情報は削除しています。参考程度に……
My image generation method is not established. Therefore, please see it only as a reference as I have not been able to verify it.
Images 1, 2, and 3 have the same seed and configuration information, but different checkpoints. The checkpoints are custom and the merge ratio is just different.
The resulting images are somewhat similar in composition, but different.
The next comparison is the settings. The reference control image is 3, and image 4 is without Detail Enhancer (LoRA: Add more details). Incidentally, this checkpoint is compensated for by the weak rendering power. As you can see, the difference is not as great as the difference between the models, but it is a big change.
Image 5 shows the model without Dynamic Thresholding (CFG Scale Fix). You can see a slight change in the pattern.
For reference, the parent of the custom model is output under the same conditions as in Image 5.
The merging and other conditions are not disclosed, but image 8 and image 9 are the starting models, and they approach the current model step by step.
It is like a proof that it cannot be reproduced, and I am not convinced, so I would like to try to see how I can get closer to the desired image when I have time.
Note that the prompts are excerpts and the weighting and image quality related information has been removed. For reference only: ......
呪文
呪文を見るにはログイン・会員登録が必須です。
イラストの呪文（プロンプト）
jacket partially removed, heart in eye, burnt clothes, holding fishing rod, kanji, doujin cover, pentagram, tape gag, adjusting headwear, red socks, friends, cloud print, coke-bottle glasses, oral invitation, competition school swimsuit, barbell piercing, gradient legwear, prisoner, blood on breasts, wind chime, carrying over shoulder, tape measure, flaming weapon
イラストの呪文（ネガティブプロンプト）
入力なし
- Steps 60
- Scale 7.5
- Seed 2832278306
- Sampler DPM++ 2M Karras
- Strength
- Noise
- Steps 60
- Scale 7.5
- Seed 2832278306
- Sampler DPM++ 2M Karras
