If animations, it's not consistent, but it's getting better and better by the time, might need a few years, haha. For live2d, I believe it can be done from the base image.
I've said for a long time that the whole "make a computer squint at random noise until it sees a waifu" approach is misguided. Look for future AI paradigms that include 3D shape, spatial awareness, and some form of cause-and-effect baked into the underlying system, not stapled on after the fact. For animation to truly work, it can't just "look good," it also has to "make sense."