This model was trained in HENTAI MMD dance.
https://www.iwara.tv/videos?tags=dynamic_hip_swing_motion
🎥HIGH-WAN2.2-I2V➡video generation!!
🎥LOW-WAN2.2-I2V➡video generation!!
🖼️IL-Start Image➡I2V StartImage generaterion!!
🎥The sample PNG image is meta information for video generation. Drag and drop it into ComfyUI to view the workflow and resources.
🎥HIGH&LOW WAN2.2-I2V
This is a dance motion involving vigorous hip swinging.
Start from a half-crouching position with your legs spread apart.
The emphasis is on provocative motions that mimic male masturbation.
🖼️IL-I2V Start Image
We are also releasing the LORA model for creating the start image.
The "IL-StartImage" model is an image generation model for Illustrious.
🎥HIGH&LOW WAN2.2-I2V➡video generate
Basic Prompt (Dynamic Hip Swing)
She performs a dynamic hip-swinging dance.
She performs a dynamic hip-shaking dance.
Option Prompt (Masturbation gesture)
She performs a dynamic hip-swinging dance.
She performs a dynamic hip-shaking dance.
She dynamic shakes her right hand vigorously up and down.
She dynamic swings her right hand vigorously up and down.
🖼️IL-I2V Start Image➡Start Image generate
Basic Prompt (Dynamic Hip Swing)
<lora:MMD_DynamicHipSwing_Anime_IL_V02:1>armpits, looking at the viewers, arms raised, arm behind her head, standing, bow-legged, cowboy shot, smile, stage, standing, pantyshot, stage_lights, spread legs,
Option Prompt (Masturbation gesture)
<lora:MMD_DynamicHipSwing_Anime_IL_V02:1>armpits, looking at the viewers, arms raised, arm behind her head, standing, bow-legged, cowboy shot, smile, stage, standing, pantyshot, stage_lights, spread legs,
a hand on crotch, male masturbation gesture, penis masturbation gesture,
This model is a proof-of-concept model for the article we published recently.
📘WAN2.2 IV2 Training Guide：https://civarchive.com/articles/27705
High-priority model
5 clips from the dataset
Switch Every:5 Steps:500➡1000
Initially, I created it using STEP:500, but I felt it lacked expressive power, so I regenerated it using STEP:1000...
training dataset and Xml is here.
https://mega.nz/folder/Vg4HwCYL#-DaekjuCc2MTQmgs1HmWoA
I use the AI model creation support software "HENTAI APP" to create the model.
📘See the article below for details：
📘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
Video Start Image Generater
FAQ
Comments (3)
To be honest, the arm movements that I prioritized for simulating male masturbation are not reproduced very well.
I changed the STEP count from 500 to 1000 and regenerated the model, but the reproduction of the penis masturbation motion was not satisfactory.
For detailed individual movements, combining it with other LORA models might be more practical.
This is just a random idea and not based on any solid evidence, but what if we merged a Lora model that performs masturbation with one that moves its hips? Or what if we trained the model using masturbation footage from Dataset 1 and hip-moving footage from Dataset 2? Wouldn’t that result in a model that’s trained just right? This is just an idea I came up with based on my own speculation about what it means to specify multiple datasets in Lora training. Since it’s not an idea I got from reading official documentation or learning the correct way to train Lora, it’s highly likely that this isn’t the correct configuration. In my personal experience, if you overtrain the model with an extremely large step size (such as 4000), the generated videos tend to have exaggerated movements. Naturally, this causes the fine details of the motion to break down, but I sometimes adjust the LowNoise settings to reduce this breakdown and stabilize the excessive movements to some extent. I hope this idea proves useful to someone.
I did a little research, but I couldn’t quite figure it out. However, if you specify footage of someone masturbating while swaying their hips in Dataset 1, and then specify the desired supporting motion—in this case, footage of someone masturbating—in Dataset 2, you might be able to generate movements that complement Dataset 1. Please note, however, that this is based on ChatGPT’s response, as it is unfortunately not explicitly stated in the official documentation.