Wan2.2 Ass Slider LoRA
V1
Adjust body ass effortlessly:
Negative (-4 to 0): Small (use prompt to help)
Positive (0 to 6): from big to huge, (use prompt to help)
Note: This one may darken the skin color, a little bit, this may be a model bias in wan, I don't know why. It may slightly alter composition; avoid extreme settings for best results, test until you understand the impact of each value.
Note 2: You can add bigger values to High noise than low noise, usually I use same on both, but it works.
Description
First version
FAQ
Comments (17)
I adore your focus on sliders. This was much needed.
How do you train sliders in Wan? How do you train the opposite ends?
https://github.com/ostris/ai-toolkit basically you find something in the model and make a push on this thing, like ass size. There are no images and captions used adding nothing, just changing the model values.
@alcaitiff Huh, so is the training faster than using data? How long did yours take to train? Did you do it on your own hardware? I have a 3090, and I'm curious about doing training on that.
You can do it.
It don't take a long time. But, like any training, it is like cooking, each time is different, and several times it is not good enough. This one I did 5 or 6 times until it got on this stage.
I'll find a tutorial for you
@alcaitiff Thank you. The video says to make a generic dataset for the subject (like in the video maker's case, "man"). Did you skip the dataset, since you said there are no images and captions? Does that work? The video doesn't seem to talk about this as optional, though I could have missed something.
If you did make a dataset, what did you use as your dataset? If you just create a generic dataset for "woman" to match the original model's understanding, which is suggested in the video, it will make women, but it won't make women who are showing their asses. One option is to have the model make "woman in tight clothing standing on the ground, facing away", or something like that, but I don't know if that veers too far from the "woman" concept. Then again, I could be overthinking this, if you didn't use any dataset at all.
Thanks!
@Jellai you can use an image of a potato. It will work anyway and don't use it, for my ass slider I don't put any ass images just selected any previous dataset used, same thing for the others. I changed the datasets and retrained, it does not matter. Just put create a dataset with 5 generic person images and use it. It works.
Would love to pay you for some specific sliders like penis size, hip size, height, age.
penis size slider is a must have please
Different kind of training. This one is modifying what the model knows already - forcing the weights to rebalance around concepts. Because the model tries to avoid genitals by default, you have to fully train genitals and then also size. It's a bigger job.
@boobkake22 this
Can you do a skin color slider? Wan 2.1 could do more body types and skin colors with a simple prompt, now with 2.2 its really difficult to get darker skin tones or control the age of the subject.
This one has a bad side effect of making the characters oily. "The Ratio" by Playtime is better. Your body size (skinny) slider is better than theirs, so yours can actually be combined at skinny -1.5 and ratio 1 to make pretty decent pear-glass bodies.
It would still be much better with a dedicated microwaist slider that can be used at .25-.5, I have not had much luck making one of my own yet.
Details
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.