MM-SOLD method generates samples on CPUs without expensive training

Researchers Zhenyu Yao and Daniel Paulin have introduced MM-SOLD (moment-matched score-smoothed overdamped Langevin dynamics), a training-free generative sampling method that eliminates the expensive neural network training typically required by diffusion models. The approach builds on recent insights that score matching implicitly smooths the empirical score and that this smoothing bias captures low-dimensional data geometry. MM-SOLD uses an interacting particle system in which particles are constrained to match the empirical moments (mean and covariance) of the target data throughout the sampling trajectory. The authors prove that in the large-particle limit, the empirical density converges to a deterministic limit whose stationary marginal is a Gibbs–Boltzmann density obtained by exponentially tilting a naive score-smoothed diffusion target.

In experiments on 2D synthetic distributions and latent-space image generation, MM-SOLD produces samples with quality and diversity competitive with neural diffusion baselines—but requires no training phase and runs entirely on CPUs. This makes high-quality generative sampling accessible on commodity hardware without the need for GPU clusters or lengthy optimization. The method is particularly relevant for low-resource settings and applications where training data is limited or privacy-sensitive, as it avoids the data-hungry training loop typical of modern generative AI.