V2 seems better yes. You only trained this on girls right? Because if i use a man, he transforms into a woman lol

Thanks for your effort in making it work; I'm very interested in this. Apparently, this effect works much better on WAN 2.1.

V2 is much better at holding likeness. Essential for transitioning to other loras like pussyjob. I get good results running it at 0.6 in the high slot.

One of the main problems we run into with AI generation is when we have an object that is covered by a thin layer of something, and we want to know what the object looks like underneath the layer.

Now, we know that what's underneath the layer should be a hyperboloid of one sheet. However, Wan has a nasty habit of turning it into a plain cylinder instead, and it's been like half a year and still nobody's make a LoRA for that.

If Wan instantly transitions, then it is pretty much guaranteed to turn it into a cylinder and doesn't look at all like what's supposed to be underneath.
If the transition is gradual, however, then it would mostly preserve its underlying shape. First, the covering layer would become transparent, then disappear, offering very believable results.
And if ScaleROPE is used with scale_t of 2 or 3, then the transitions won't take so long and we'll be able to see underneath that much more quickly. I suppose a depth ControlNet, an inpainting mask, and maybe some segmentation would help but I haven't tried those yet.
I do congratulate you though on creating the best LoRA I've seen for this particular use case.