Research
High-accuracy sampling for diffusion models and log-concave distributions
arXiv:2602.01338v2 Announce Type: replace Abstract: We present algorithms for diffusion model sampling which obtain elta-error in polylog(1/elta) steps, given access to widetilde O(elta)-accurate scor
arXiv:2602.01338v2 Announce Type: replace Abstract: We present algorithms for diffusion model sampling which obtain elta-error in polylog(1/elta) steps, given access to widetilde O(elta)-accurate score estimates in L^2. This is an exponential improvement over all previous results. Specifically, under minimal data assumptions, the complexity is widetilde O(d_star polylog(1/elta)) where d_star is the intrinsic dimension of the data. Further, under a non-uniform L-Lipschitz condition, the complexity reduces to widetilde O(L polylog(1/elta)). Our approach also yields the first polylog(1/elta) complexity sampler for general log-concave distributions using only gradient evaluations.
Related
- When Langevin Monte Carlo Meets Randomization: Non-asymptotic Error Bounds beyond Log-Concavity and Gradient Lipschitzness
- HardFlow: Hard-Constrained Sampling for Flow-Matching Models via Trajectory Optimization
- Radial Load--Reserve Certificates for Wasserstein Propagation in Isotropic Diffusion Samplers
- Convergence Rates for Non-Log-Concave Sampling and Log-Partition Estimation
- Control Consistency Losses for Diffusion Bridges
Source: arXiv cs.LG | 2026-04-28