Generating new samples that look like real data often requires many expensive queries of a learned model. This becomes a bottleneck when the target distribution has a wide range of scales (a high condition number), like images with both fine and coarse features. The question is how to choose both the path from noise to data and the numerical solver to minimize the number of queries needed.
The authors propose a joint design: an "elliptic schedule" that defines the interpolation path, and a tailored discretization called Zoloflow, inspired by optimal rational approximations. For Gaussian distributions, this combination achieves minimax optimal performance, requiring only about log(condition number) times log(sqrt(dimension)/error) score evaluations.
This matters because it provides a principled, optimal strategy for the fundamental Gaussian case, and extends to strongly log-concave distributions with improved guarantees. The schedule itself is also found to be effective on image generation benchmarks even beyond the theoretical guarantees, suggesting practical value.
