⚠️This LORA is for WAN2.2 I2V video generation!!⚠️
The sample PNG image is meta information for video generation. Drag and drop it into ComfyUI to view the workflow and resources.
This model was trained in Iwara's HENTAI MMD dance.
https://www.iwara.tv/search?type=videos&page=0&query=How%20It%27s%20Done&sort=views
The training was done using clips created from MMD videos.
It is believed to be optimized for anime images.
Operation has not been checked with live-action images.
The sample video is set to 16 FPS.
The motion is limited to hip-swinging motions while standing.
We're not committed to exact, perfect motion reproduction.
training dataset is here.
https://mega.nz/folder/csohADKK#ga12soB_jn8aANWLWGn33Q
Basic Prompt (SIDE 2 SIDE)
She is swaying her hips rhythmically from side to side.Basic Prompt (BACK and FORTH)
She is swaying her hips rhythmically back and forth.Pose details Prompt (Her arms are on her hips.)
Intense dance motion
Front angle.
Shaking her hips from side to side
Her arms are on her hips.
Pose details Prompt (arms raised to show armpits.)
Intense dance moves.
Dancing with arms raised to show armpits.
Shaking hips from side to side.
Pose details Prompt (middle fingers raised.)
Intense dance movements.
Provocative dancing with the middle fingers raised on both hands.
Shaking hips up and down.
Pose details Prompt (Pseudo-masturbation)
Intense dance motions.
Pseudo-masturbation dance.
Shaking hips back and forth.
Shaking hips up and down.
■sorry Japanese text■
🎊IWARA復活おめでとう！！🎊
MMDのエロダンスが大好きです。
このモデルは、MMDモデルの無いキャラクターにIWARAのエロダンスを踊ってもらうために作りました。
当初は、中腰や仰向け・背面振り向き尻たたきなどを含めいた多様なポーズでデータセットを作成したのですが、モデル作成に失敗しました。
ので、今回は妥協して立ち状態の腰振りモーションに限定したトレーニングとなっています。
サンプル動画は16fpsに設定されています。
モーションは立った状態での腰振り動作に限定されています。
正確かつ完璧なモーション再現にはこだわっておりません。
実写画像からのI2V検証は行っていません。
「How it's Done」は、単純に私が大好きな動画シリーズなのでトレーニングしました。
IWARAやNICONICOだと本シリーズの動画しか表示されないのですが、You Tubeでは完全に検閲削除されてて、同姓同名の全く別曲の動画が該当するのがなかなか…面白いですね。
今後とも、よろしくお願いします。
モデル作成には、AIモデル作成支援ソフト『HENTAI APP』を利用しています。
詳細は以下記事を参照：https://civarchive.com/articles/26629/create-wan22-datasets-from-videos-optimized-software-has-been-released
📘rticle from the WAN2.2 IV2 Training Guide
https://civarchive.com/articles/27705
私は本LORAの利用に、制限・権利・責任を一切主張しません。
Description
FAQ
Comments (6)
Hello, I tested Lora. The dynamism of her body movements while dancing is great; the "Her arms are on her hips" and "Arms raised to show armpits" effects are perfectly rendered!
If the character in the original image isn't making a middle finger gesture, the animated version will show her index finger raised.
When using Pseudo-masturbation, if the original hand isn't placed on the genitals, the dance will become random and won't automatically extend to the genitals.
During video generation, the character has a chance to automatically change her hairstyle to a ponytail/twins and wear a headband.
If the character isn't standing, such as sitting or lying down, the dance movements will be lowered, making it appear as if she's actually sitting/lying down.
Thank you for your comment.
Thank you for your valuable feedback!
The LOW noise parameter on the detail display side might have been a little too high.
This will be helpful for future model creation! Thank you!
Based on my general understanding and experience, my approach is to lower the HIGHT intensity when motion or camera movement is erratic, and lower the LOW intensity when details or background movement is erratic.
I've posted examples of live action results. As the other commenter said, if the person doesn't already show middle fingers, she will not do it at all and if the person is not standing up the lora won't really work.
Thank you for your comment and video submission!
I haven't trained the motion of raising the middle finger from a non-upright position.
Also, initially, I was training actions like squatting, handstands, and back-slapping, but I removed them from training because the movements became very unstable.
Honestly, I don't know how to create a stable I2V model for WAN2.2.
It's commonplace for a model created with a 500-step video clip to work perfectly, while a 30-clip, 3000-step model completely breaks down. It's driving me crazy! 🤪 But, it's really fun to experiment without knowing the right answer! 🤪
Also, as usual with WAN 2.2, it tends to transform 2D into 3D, so this actually works better on live action than anime.
@Slowmoe
Thank you for your comment!
Based on my limited knowledge,
HIGH: Prioritizes movement, camera work, and other effects.
LOW: Prioritizes details such as style and facial expressions.
I'm deliberately sticking to hentai anime models, so I'm creating them using parameters that prioritize both HIGHT and LOW.
I don't think it's working effectively yet...
Training takes a tremendous amount of time—12 hours—so I've failed many times with complex models.
I failed to train the model yesterday and today 😂
But it's a lot of fun!🤣
Details
Files
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.