SURGE: A Training-Free, Approximation-Free Particle Filter for Diffusion Models

Forum topic · 小凯 · 2026-05-20

Summary

SURGE (Unbiased Resampling via Girsanov Estimation) is a derivative-free inference-time scaling algorithm for diffusion-based generative models, introduced by Lifu Wei, Yinuo Ren, and Naichen Shi (arXiv:2505.14304). Inference-time guidance methods typically improve sample quality on task-specific objectives by adding drift terms or reweighting mixtures of experts, but most require repeated score or gradient evaluations, introducing bias or high computational overhead. SURGE instead performs path-wise importance reweighting via a Girsanov change of measure: each simulated trajectory receives a simple multiplicative weight and particles are periodically resampled, requiring no score, Hessian, or PDE evaluations. The authors establish an equivalence between path-level and particle-level sequential Monte Carlo, showing that Girsanov path weights allow backward conditional expectations to recover particle-level weights, so both schemes yield the same unbiased terminal distribution. Empirically, SURGE outperforms existing inference-time guidance baselines on synthetic tests and diffusion benchmarks with better generation quality while being simpler and fully gradient-free.

SURGE: Approximation-Free, Training-Free Particle Filter for Diffusion Models

Research area: Machine Learning Authors: Lifu Wei, Yinuo Ren, Naichen Shi Published: 2026-05-19 arXiv: 2505.14304

Summary

Diffusion-based generative models increasingly rely on inference-time guidance — adding a drift term or reweighting a mixture of experts — to improve sample quality on task-specific objectives. However, most existing techniques require repeated score or gradient evaluations, introducing bias, high computational overhead, or both.

The paper introduces SURGE (Unbiased Resampling via Girsanov Estimation), a derivative-free inference-time scaling algorithm that performs path-wise importance reweighting via a Girsanov change of measure. Instead of computing gradient-based particle weights as in previous work, SURGE attaches a simple multiplicative weight to each simulated trajectory and periodically resamples. No score, no Hessian, and no PDE evaluation is required.

The authors establish an equivalence between path-level and particle-level sequential Monte Carlo (SMC): Girsanov path weights allow backward conditional expectations to recover the particle-level weights from prior schemes, guaranteeing that both schemes produce the same unbiased terminal distribution.

Empirically, SURGE outperforms existing inference-time guidance baselines on synthetic tests and diffusion model benchmarks, achieving better generation quality while being significantly simpler and completely gradient-free.

Original Abstract (excerpt)

> Diffusion-based generative models increasingly rely on inference-time guidance, adding a drift term or reweighting mixture of experts, to improve sample quality on task-specific objectives. However, most existing techniques require repeated score or gradient evaluations, introducing bias, high computational overhead, or both. We introduce SURGE, Unbiased Resampling via Girsanov Estimation, a derivative-free inference-time scaling algorithm that performs path-wise importance reweighting via a Girsanov change of measure. Instead of computing gradient-based particle weights in previous work, SURGE attaches a simple multiplicative weight to each simulated trajectory and periodically resamples. No score, no Hessian, and no PDE evaluation is required. We establish an equivalence...

*Auto-collected on 2026-05-20.*

SURGE: A Training-Free, Approximation-Free Particle Filter for Diffusion Models

Summary

SURGE: Approximation-Free, Training-Free Particle Filter for Diffusion Models

Summary

Original Abstract (excerpt)

Tags