This new paper speeds up diffusion models by analytically solving the linear part of the reverse ODE to make precise solvers.
I find it super cool that they use our adaptive step-size algorithm () to get their best results! 😸
@jm_alexia
While the higher order methods were slower, did you do a measure of strong and weak error? I would suspect they were just hitting a much lower error, since they tend to have not just a higher order but also a more accurate error estimator. Did you check work-precision?
@ChrisRackauckas
It's not quite what you ask, but we have Figure 1 in with x=Number of Function evaluations (time), y=FID (metric of quality/diversity of the generated data).
The DPM-Solver paper have this info instead in tables.