>>29744484
I think a Lora would be what is needed along with an image with it already there. If you are transforming to another image anyway then you might as well have that image already with the ball-gag in place rather then rely on a prompt to somehow do it. However if there is a prompt that would do it - maybe more detail is need. After all the video training on the Wan model would probably have ball-gags. But it helps if the image already has it anyway but it clearly does not understand "ball-gag" on it's own.