Something to keep in mind is that part of the problem with these fakes is the shallowness of the training data. If rather than, say, 30k images being used - 3 million were instead - that would cause a great change in the quality.

But - that would also take a ton more processing; instead of a clip taking several hours to create, you might be looking at it taking a few days. That, or you'd have to throw a ton of money into better hardware to handle the extra load.