the speed and quality of score based/diffusion depends on what sampler you use. If youre using euler's method to solve the ODE for example, that might be slower than some of the newer methods developed for diffusion models, like tero karass' ODE solvers. AFAIK there isnt consensus on what the best sampler to use is though.

i dont think it affects training convergence much though since its more or less the same objective.