Search papers, labs, and topics across Lattice.
This paper introduces a conservative drifting method for one-step generative modeling, replacing displacement-based drifting velocity with a kernel density estimator (KDE)-gradient velocity to address non-conservatism. The authors prove continuous-time finite-particle convergence bounds for this method on $\R^d$, deriving rates for the empirical Stein drift, smoothed Fisher discrepancy, and squared center velocity, with explicit quadrature constants and bandwidth dependence. They also analyze the non-conservative drifting method with a Laplace kernel, providing a finite-particle rate with an unavoidable residual term and explaining how residual-velocity bounds translate into one-step generation guarantees.
Forget GANs and diffusion models – this new "conservative drifting" method for generative modeling offers provable convergence guarantees and addresses the non-conservatism issues plaguing existing displacement-based approaches.
We propose and analyze a conservative drifting method for one-step generative modeling. The method replaces the original displacement-based drifting velocity by a kernel density estimator (KDE)-gradient velocity, namely the difference of the kernel-smoothed data score and the kernel-smoothed model score. This velocity is a gradient field, addressing the non-conservatism issue identified for general displacement-based drifting fields. We prove continuous-time finite-particle convergence bounds for the conservative method on $\R^d$: a joint-entropy identity yields bounds for the empirical Stein drift, the smoothed Fisher discrepancy of the KDE, and the squared center velocity. The main finite-particle correction is a reciprocal-KDE self-interaction term, and we give deterministic and high-probability local-occupancy conditions under which this term is controlled. We keep the quadrature constants explicit and track their possible bandwidth dependence: the root residual-velocity rate $N^{-1/(d+4)}$ holds under an additional $h$-uniform quadrature regularity condition, while a more general growth condition yields the optimized root rate $N^{-(2-β)/(2(d+4-β))}$, where $0\le β<2$. We also analyze the non-conservative drifting method with Laplace kernel, corresponding to the original displacement-based velocity proposed in~\cite{deng2026drifting}. For this method, a sharp companion kernel decomposes the velocity into a positive scalar preconditioning of a sharp-score mismatch plus a Laplace scale-mismatch residual, producing an analogous finite-particle rate with an unavoidable residual term. Finally, we explain how the continuous-time residual-velocity bounds translate into one-step generation guarantees through the explicit drift size $η$.
Amazon Science