JOI LoRa for Wan2.2.
Typical prompt:
Her fist performs slow, sensual vertical pumping motion at the edge of a portrait shot, moving up and down with languid, deliberate tempo as it enters and exits frameThe T2V version generates very similar faces every time - I think it has to do with my captioning of the dataset not being great. You might have to play with prompt or have another LoRa to compensate for it. I2V version works fine.
Please support my work by leaving positive reviews and posting your generations to the model page.
As always, please use responsively
Description
FAQ
Comments (12)
based
It's really good when I can get the hand to appear but it seems like if the hand is not already visible it refuses to move it front of the person. Anyone figure out how to do that?
Ever heard of prompting?... If you want her hand to move up and appear at the bottom of the frame, before she moves it up and down, etc. write it down...
problem is they are moving their mouth and opening, but there is no sound
@MatorinaDaddy It's the whole point of JOI videos, they're speaking :D
why the f are their mouths moving if there is no sound
Presumably because it's been trained on videos where mouths move, and not trained on the audio, which isn't within the capacity of most video models.
@corve ok so someone smart trained this on dataset where the mouths are moving but this is a "JOI" video trained on a video model which has no capacity to train on sound, why? Just train the gestures and movements then, there is plenty for JOI, instead of this muted mouth opening thing
@MatorinaDaddy Do it yourself then, whiner
@joxaro9187376 wtf are you talking bout, your mouth open but i cant hear ya
Though it would have been great if you provided an actual workflow that works. She doesn't make a fist with the provided prompt for some reason. Just add the workflow along with the video so we can see how it works.
Now I can always win at rock paper scissors.