>>25638834Yeah but unlike Live2D, this uses Machine Learning.
My experience with lots of Machine Learning projects is that most end results run slow as fuck. Not to mention it uses Python. Python projects are notoriously slow and not meant for In-real time stuffs like this.
This thing is also public in Github for a very long time, but so far no pngtubers have been using it in place of a png, why?
Maybe the first time if I input a few image and keyframes the initial generation process is slow as fuck (as expected), but what if the compatible software to be used with this model is ressource heavy (as it is nowhere stated to have crosscompatibility with Live2D files).
will the rig generated by this project be compatible with existing softwares like VtuberStudio? Sure it will get there eventually, but I do not think fast enough for the results above.
How does the tech works? Live2D uses tons of separated layers + objects to generate different permutations (so low footprint). But this uses ML (GAN), so likely that each permutation rendered as a different file/instance which is what I think would take much more memory? My perception of visual GAN is that it would do a lot things, but never efficiently as opposed to manual tasks (there would be way too many 'trash/artifact' to clean up like those waifu generators).
