NAI2-NAI35-LPCM版本v1.0 (ID: 881923)

Summary

An attempt to expand the CFG scale by mixing LCM with PCM.

Using Euler and Euler a samplers that can be used in both.

---

This is a fairly experimental merge model. The model itself is not very practical.

これはかなり実験的なマージモデルです,モデル自体はあまり実用にはなりません,

I have almost exclusively used LCM, but while LCM allows for high-speed generation, it has the problem that the CFG scale cannot be set large. Generally, CFG scale is considered to be a parameter that controls the fidelity to the prompt, but I don't think many people use it with that intention. Changing this parameter changes the tone of the image and the appearance of the burn, so there are many models where the setting range is limited and it cannot be changed significantly. In the past, it was possible, but nowadays, I rarely see people using anything other than 7.

私はほぼLCMしか使わなくなっていますが，LCMは高速生成が可能になる半面，CFG Scaleが大きくとれなくなるという問題を抱えています,一般にCFG sacaleはプロンプトへの忠実度を制御するパラメータとされていますが，そういう意図で使っている人はそう多くはないのではないかと思っています,これを変更することで画調や焼けの出方が変わるので，設定範囲が制限されて大きく変えることはできないモデルが多いからです,昔ならともかく最近では7以外で使う人はめったに見ません,

In LCM, the setting range of CFG scale is limited to 1 to 2. Even if the fidelity to the prompt can be managed to some extent by ingenuity of the prompt, what may be more of a problem is the effectiveness of the negative prompt (well, I rarely use negative prompts, but I'm sure there are many people who put various things into negatives). This is also affected by the CFG scale.

LCMではCFG scaleの設定範囲は1～2に限定されます,プロンプトへの忠実度はプロンプトの工夫などである程度なんとかなるとしても，もっと問題になるかもしれないのがネガティブプロンプトの効き具合です（まあ，私はめったにネガティブプロンプトは使わないのですが，いろんなものをネガティプに突っ込む人は多いでしょう）,こちらもCFG scaleの影響を受けます,

In any case, there is no doubt that it is advantageous for image generation to have a CFG scale as large as possible. LCM is a model that balances generation speed and fidelity in that respect and chooses speed. The idea is that even if the fidelity is half, you can get a good image by running it at twice the speed.

なんにせよ，CFG scaleをできるだけ大きくとれるほうがイメージ生成で有利になるのは間違いありません,LCMは生成速度とその辺の忠実性を秤にかけて速度を取ったモデルです,忠実度が半分でも倍速で回せばいい絵も出てくるみたいな考え方ですね,

There are various speed-up methods other than LCM. PCM is one of them. There is LoRA that generates images in 8 steps or 4 steps, but unlike LCM, it has the option to set a large CFG scale.

LCM以外の高速化手法もいろいろあります,PCMもその1つです,8ステップや4ステップで画像を生成するようなLoRAが出ていますが，LCMと違って，CFG Scaleを大きく取れる選択肢も用意されているのが特徴です,

As far as I tried, the 4-step one is useless in terms of quality. 8 steps is probably within the practical range. However, it is a little unstable compared to LCM, which produces a fairly stable image. It also produces burns. LCM was a sampler that hardly produced burns, so the model I created is based on that assumption and is packed with data with a fairly dense learning. It is not suitable for PCM.

試した限り，4ステップのものはクオリティ的に使いものになりません,8ステップなら実用範囲というところでしょう,ただ，LCMがかなり安定した絵を出すのに対してやや不安定です,焼けも出ます,LCMはほとんど焼けを出さないサンプラーでしたので，私が作成しているモデルはそれを前提にかなり濃いめの学習でデータを詰め込んでいます,PCM向きではありません,

So, this model is the result of doing something unconventional, applying LCM and PCM at 0.5 each. The data content is exactly the same as NAI2-NAI35, which has already been released.

ということで，LCMとPCMを0.5ずつ適用してみるという非常識なことをしたのが今回のモデルとなります,データ内容はすでに公開しているNAI2-NAI35とまったく変わりません,

Please use Euler for the sampler. Euler a can also be used. DDIM and LCM are also quite fine, and although the images produced by DDIM are quite unstable, they are interesting. The number of steps should be around 8 and the CFG scale should be around 3.5.

サンプラーにはEulerを使用してください,Euler aも使えます,DDIMやLCMでもわりと大丈夫で，DDIMの出す絵がかなり不安定なものの面白い感じです,Step数は8，CFG scaleは3.5くらいを基準にしてください,

Personally, I think LCM alone is more useful, but it's interesting that it can produce images that are a little different from usual. Setting Adetailer is difficult, and sometimes it seems better not to set it. The method itself can be used as it is with LoRA without merging it with the model. You can also change the blend ratio, but be aware that the useful sampler and CFG scale will change.

個人的にはLCM単体のほうが有用だと思いますが，いつもと少し違った感じの絵が出せるのが興味深いところです,Adetailerの設定が難しめで，設定しないほうがいいのではないかと思えることもあります,なお，手法自体はモデルにマージせずにLoRAのままでも利用可能です,ブレンド比を変えることもできますが，有用なサンプラーやCFG scaleは変動してきますので注意してください,

描述:

训练词语: