Is Flux refiner training possible?

[matabear-wyx]

Maybe some two-stage training like SDXL?

[bghira]

it is continuous time schedule so not really. flow-matching in general is a disappointment for these reasons

[matabear-wyx]

yet for a continuous time scheduler, the concept of a 'refiner' still holds, right?
Since there are no explicit discrete steps, there can be a seamless transition between the base generation and refiner stages. I think it needs to be redesigned in terms of dividing the time intervals and adjusting the noise reduction strategies.
But I don't really have any idea on how to design this for now :(