Diffusion Quasi-Monte Carlo
arXiv:2608.11055
2026
Sampling
1 ideas extracted · analyzed Sep 1, 2026
What the math gives to ML
The paper provides a principled way to replace independent diffusion-model seeds with scrambled low-discrepancy seeds while preserving valid expectation estimates through importance weighting. Its transferable asset is the composition of inverse-Gaussian-coordinate transport and a discretized probability-flow ODE, together with an explicit correction for density distortion caused by approximate transport. This is most useful for diffusion ensembles whose outputs are aggregated into nonlinear metrics, where ordinary Monte Carlo creates large seed-to-seed variance. The strongest initial application is batched diffusion inference for uncertainty estimates, calibration, or scientific metrics rather than ordinary single-sample generation.
Ideas from this paper
✗ Failed on benchmark
2026
Use Owen-scrambled Sobol points instead of independent Gaussian seeds for batched diffusion sampling, mapping each cube point through the component-wise inverse Gaussian CDF and the model's probability-flow ODE. Estimate ensemble expectations with importance weights computed from the target-to-proposal density ratio, so the estimator remains valid despite finite-step and learned-score transport errors.
Useful7/10
Difficulty6/10
Novelty7/10