<h2 id="heading-27">Twinkle: Real Semi-Real LCM</h2>This is a realistic semi-realistic model created using <a target="_blank" rel="ugc" href="https://civarchive.com/models/280500/instantrealizer">InstantRealizer</a>. Unusually among my models, most of my models are Asian women.Since it is positioned as semi-realistic, the texture of the skin has not been improved that much. In terms of <a target="_blank" rel="ugc" href="https://civarchive.com/models/280500?modelVersionId=316811">IR02</a> parameters, it is 0.27.<a target="_blank" rel="ugc" href="https://civarchive.com/models/280500/instantrealizer">InstantRealizer</a>を使用して作ったリアル寄りのセミリアルモデルです。私のモデルでは珍しく，アジアン女性が中心になります。あくまでセミリアルという位置づけですので，肌の質感などはさほど上げておりません。IR02のパラメータで言うと，0.27です。This time, I have incorporated LCM from the beginning.This model requires LCM as a sampler. It can be obtained by expanding <a target="_blank" rel="ugc" href="https://github.com/guoyww/AnimateDiff">AnimateDiff,</a> but if you do not have it on hand, please use Euler a instead.The number of steps should be 10 and the CFG scale should be 2. Also, this time we have adjusted it with ADtailer in mind, so please use it if the resolution is low. There is also an example using Hires.Fix, but as you can see, Denoising is set quite low. By incorporating LCM, it is now practical to increase the raw resolution from the beginning. It's impossible to make the width 1024 or more, but you can use it with about 900 without much difficulty.今回はLCMを最初から組み込んであります。このモデルを使用するにはサンプラーとしてLCMが必要です。Animatediffを拡張することで入手できますが，手元にない人はEuler aで代替してください。ステップ数は10，CFG scaleは2を目安にしてください。また，今回はADtailerを念頭にして調整していますので，解像度が低い場合は使用してください。Hires.Fixを使った作例もありますが，見てわかるようにDenoisingをかなり低くしています。LCMを組み込んだことで，素の解像度を最初から上げておくことも現実的となっています。横1024以上にするのは無理でも900くらいならあまり無理なく使えます。This model is created based on <a target="_blank" rel="ugc" href="https://civarchive.com/models/233103/rm01-experimental-base-merge">RM01</a>. I selected some semi-realistic models (8 models) that looked good, converted them to Lora based on the differences with RM01, added InstantRealizer, and added Lora for adjustment based on that model. Since the merge test is performed on the prompt, it is easy to avoid the trouble of creating intermediate models. It would have been possible to determine the ratio in Lora and then merge the models themselves, but this time I tried merging only in Lora. This was also easy as I was able to merge nine models in one operation, including the base model RM01.このモデルはRM01ベースで作成しまします。よさそうなセミリアルモデルをいくつか（8モデル）選び，それをRM01との差分でLora化して，InstantRealizerを加え，それをベースに調整用Loraなどを加えて作りました。プロンプト上でマージテストを行っているので，中間モデル作成の手間が省けて楽です。Loraで比率を決めて，モデル自体をマージしてもよかったのですが，今回はLoraだけのマージでやってみました。ベースモデルのRM01を含めて9個のモデルのマージが1回の操作でできるので，これも楽でした。<img src="https://image.civitai.com/xG1nkqKTMzGDvpLrqFT7WA/c8b5a0f8-a3da-4061-a35f-12aaede14b8c/width=525/c8b5a0f8-a3da-4061-a35f-12aaede14b8c.jpeg" />(*During the process, the male character looked too bad, so I mixed the Blood &amp; Spirit Mix a little and adjusted it. Unfortunately, the output of the combined Lora on the prompt and the output of the merged model of Lora are different. )The faces of the semi-realistic models are more or less fixed, so they don't easily fall apart, so it was a lot easier. The most common faces are familiar to everyone. This girl was created by adjusting the other faces and blending ratio.セミリアルモデルは，それぞれ顔がだいたい固定されており，崩れることが少なくて，こちらもなにかと楽でした。一番多いのは皆さんよくご存じの顔ですね。それ以外の顔と混合比を調整して今回の女の子ができました。<s>This model is only intended to draw </s><s>1girl</s><s>. I'm not sure what happened, but the ability to follow prompts is low. At first, I couldn't even say I wanted to send a man or two girls together. After additional learning of attn, I finally got to a level where I can't get enough of it. Please use the sample image prompts as a guide. As with other models, the limbs and fingers may become strange, but sometimes the size of the neck and head becomes strange. please note.</s>FIXED.It seems that there was a memory shortage because too many Loras were being merged at once. I was able to create a <a target="_blank" rel="ugc" href="https://civarchive.com/models/298792?modelVersionId=335816">normal ckpt </a>by splitting and merging.Please <a target="_blank" rel="ugc" href="https://civarchive.com/models/298792?modelVersionId=335816">download V1.1</a>.<s>このモデルは1girlを出すことしか想定していません。どこでどうなったのか不明ですが，プロンプトへの追従性は低めです。最初は男性を一緒に出したい，女の子2人出したいといったこともできませんでした。attnの追加学習をしてようやく出なくもないレベルになりました。サンプル画像のプロンプトを参考にしてみてください。手足や指がおかしくなることがあるのは，ほかのモデルと同様ですが，たまに首や頭のサイズが狂います。ご注意ください。</s>

Twinkle: Real Semi-Real LCM

This version is buggy due to memory shortage. Plz download v1.1.

Files with SHA256: 20643948632f9c98bfe3177ab574c0ea48525fd1df0a3b3bc8dc4a104ea09e0d