<h3 id="ver.3.0-2ure9hvbc">ver.3.0</h3>This is a version that reflects the latest know-how.I've aimed for more detail and color saturation, and (where possible) better body stability and prompt fidelity, but it's still not enough.最近のノウハウを反映したバージョンです。精細感と色の濃さの向上，（可能であれば）身体安定性とプロンプト忠実性なども狙ったものですが，まだ十分ではありません。The model is made with non-LCM, but all samples use LCM LoRA. However, the sampler uses Kohaku_LoNyu_Yog. My recent personal trend is to use it with the same settings as the LCM sampler.モデルは非LCMで作られていますが，すべてのサンプルでLCM LoRAを使用しています。ただしサンプラーにはKohaku_LoNyu_Yogを使用しています。LCMサンプラーと同じ設定で置き換えて使うのが最近の個人的トレンドです。When I uploaded the file, I felt that the photo processing was a little sloppy depending on the prompt, but it's fine as it is. If I try to increase the photo processing one more step, the figure looks more like figurines, but it looks cheap. If I increase it one more step, it looks more like a doll than figurines. The texture of the hair and other parts becomes too realistic and it doesn't look like figurines.ファイルをアップロードしたところで，プロンプトによっては少し写真化処理が甘いなと感じましたが，まあこれはこれでまとまっているのでよしとしましょう。試しに写真化をもう1段上げるとフィギュア感は向上しますが，見るからにチープなフィギュアになります。さらに1段上げるとフィギュアというよりドール感が向上します。髪の毛などの質感がリアルになりすぎてフィギュアぽくなくなります。I was also creating additional learning LoRA for figure creation, but the quality was not so good so I didn't use it this time.フィギュア化用の追加学習LoRAも作成中だったのですが，クオリティがいま一つだったので今回は採用していません。<hr /><h3 id="ver.2.0-y16pmemes">ver.2.0</h3>PVC Figurerizer ver.2.0 is LCM specification. Please use LCM or Euler a for the sampler. The standard number of steps is 12 to 15 and the CFG scale is about 2.PVC Figurerizer ver.2.0はLCM仕様です。サンプラーにはLCMかEuler aを使ってください。Steps数は12～15，CFG scaleは2くらいが基準です。BTW What kind of figure is a good figure? A figure with big boobs? 　It may indeed sell better that way. But that's not the case. A regular figure expresses the charm of the character in a single pose. Models that reproduce this need to have a variety of poses and expressions. Furthermore, it is desirable that the picture as a whole has a sense of drama. In ver.1, there were major limitations in terms of backgrounds, etc. In ver.2, we have expanded this area and made sure that you can create pictures that are more realistic and have a higher degree of freedom. We've included ImageEnhance to help you compose dramatic scenes with the simplest possible prompts. I made the boobs bigger just in case.よいフィギュアとはどんなフィギュアでしょうか？　おっぱいが大きいフィギュア？　確かにそっちのほうがよく売れるのかもしれません。しかし，そういうことではありません。通常のフィギュアは，そのキャラクターの魅力を1つのポーズで表現したものです。それを再現するモデルには，ポーズや表情が多彩さが望まれます。さらに絵全体としてのドラマ性があるものが望ましいでしょう。ver.1では背景などで大きな制限があった。ver.2ではそのあたりを拡充し，よりリアルで自由度の高い絵作りができるように配慮しています。できるだけ簡単なプロンプトでドラマチックなシーンを構成できるようにImageEnhanceを組み込みました。念のためにおっぱいは大きめにしておきました。Note:If you think the face is not cute, please do not use ADtailer and HighRes.Fix at the same time. Both are great tools for increasing resolution, but they often destroy the picture. This model does not take into account facial details at medium distances, so it is essential to use one or the other. However, there are times when it is better not to use them at the same time. When using HighRes.Fix, set denoising and Hires steps to a low value.<img src="https://image.civitai.com/xG1nkqKTMzGDvpLrqFT7WA/a71767a6-0afb-4ae7-b208-590d536d0dc6/width=525/a71767a6-0afb-4ae7-b208-590d536d0dc6.jpeg" />注意：顔が可愛くないと思ったときにはADtailerとHighRes.Fixは同時に使わないでください。両者とも解像度を上げるには優れたツールですが，高い頻度で絵を破壊します。このモデルは中距離での顔の詳細について配慮していませんので，どちらかを使うことは必須となります。ただし，同時には使わないほうがよいときもあります。HighRes.Fixを使う場合は，denoisingとHires stepsは低めに設定してください。<h3 id="ver.1-lf70z5goq">ver.1</h3>This is a model that generates figure(figurine)-like images.At first, I made it 2.5D model based on Kawaii Animetic V2 and then combined 6 types of Lora in the optimal ratio.There are several figurine models, but this model does not paint the character to look like a figure, but rather makes the prompt a figure and vignette.フィギュアのような画像を生成するモデルです。Kawaii Animetic V2をベースに2.5D化を行い，6種類のLoraを最適な割合で組み合わせました。フィギュア化モデルはいくつか存在しますが，本モデルはキャラクターをフィギュアっぽく塗るのではなく，プロンプトをフィギュアとヴィネットにするような動作となります。In general, increasing the texture of PVC greatly affects the character's face, and even if you look at existing models, you can see that the character is prioritized in the creation of the image. However, there are cases where it is hard to say that anything other than the hair area is made of PVC. For this model, I have adjusted the balance a little closer to PVC, which I think is the best at the moment. The face is affected accordingly, so Adetailer or HiRes.Fix is required. You may also use the Detail version of Lora if you like.一般に，PVCの質感を上げるとキャラクターの顔などに影響が大きく出てしまい，既存のモデルを見てもキャラクターを優先した絵作りが行われていることがわかります。ただ，髪の毛のあたり以外はPVC化されているとはいいがたいものも見受けられます。このモデルではもう少しPVCに寄せて，現状ではこれがベストだと思えるバランスで調整しました。その分，顔に影響が出ていますので，AdetailerかHiRes.Fixは必須となります。また，好みでDetail系のLoraを使用するのもよいでしょう。Perhaps because Lora's training data is biased, the outputs are a little overexposed, but there is no problem I think.Loraの学習データに偏りがあるためか，やや露出の多い出力となりますが，とくに問題はありません。* figurerise(figurerize) is a word of Japanese-English. It is something that has been used for a long time. But, Bandai-Namco made "FIGURE-RISE" a registerd trade mark at toy area in Japan. Figurerize or figureriser or figurerizer are not registered yet, and it is not toy area. I dare to name this Figurerizer.

PVC Figurerizer

pvcFigurerizer_v30.safetensors

v3.0

v2.0

v1.0

# PVC Figurerizer v3.0

This is a checkpoint trained to generate high-quality PVC figure and figurine-style images. The model's goal is not just to apply a 'figure' texture to a character, but to turn the entire prompt into a complete figure and vignette scene.

This latest version (v3.0) has been enhanced to provide more detail, increased color saturation, and better body stability and prompt fidelity.

**Usage Recommendations:**
- **Base Model:** SD 1.5
- **Activation Words:** There is no single mandatory trigger word. For best results, include terms like `pvc figure`, `figurine`, `on a display stand`, and `diorama` in your prompt.
- **Essential Tools:** Using **ADetailer** or **Hires.Fix** is considered essential for good facial details, as the model prioritizes the PVC texture. The author recommends not using both at the same time.
- **Settings (for LCM workflow):** The author uses a separate LCM LoRA to generate the sample images for speed. When using that workflow, the recommended settings are:
    - **Sampler:** LCM or Euler a
    - **Steps:** 12-15
    - **CFG Scale:** ~2.0

Files with SHA256: 5f9a6aa6bf174f11019f23a136437e630bc374bb4c52022c257378e202736f80