SR3: variable number of denoising steps?

[carbocation]

The SR3 paper doesn't seem to be very clear about its training setup or its architecture. Nevertheless, it does seem to imply that SR3 can be run with a varying number of denoising steps ("SR3 can generate high-resolution images, e.g., 1024×1024, but with a constant number of refinement steps (often no more than 100).")

However, that doesn't seem possible with the current approach, where the number of denoising steps is fixed at training time and cannot be modified during evaluation (since the shape of the model is a function of the number of denoising steps). I am curious if I misinterpreted the description in the SR3 manuscript.