Sad part is that we need to train a generative model from scratch to use this one; i.e., we can’t fine-tune current models to use FFF.

Hope someone does it soon.