Guiding diffusion models without computing gradients

Diffusion models often need guidance at inference time to match specific objectives—like generating images of a certain style. Current methods require repeated gradient or score evaluations, which is expensive and can introduce errors. URGE replaces this with trajectory reweighting: each generated sample path gets a simple multiplicative weight, and the algorithm periodically resamples based on those weights. No gradients, Hessians, or PDE solving needed. The method is mathematically proven equivalent to prior particle-filtering approaches but simpler to implement and faster in practice.