The article studies the problem of training diffusion models to sample from a distribution with a given unnormalized density or energy function. It benchmarks several diffusion-structured inference methods, including simulation-based variational approaches and off-policy methods. The study also proposes a novel exploration strategy for off-policy methods, which improves the quality of samples on a variety of target distributions. The code for the sampling methods and benchmarks studied is publicly available for future work on diffusion models for amortized inference.
Publication date: 8 Feb 2024
Project Page: github.com/GFNOrg/gfn-diffusion
Paper: https://arxiv.org/pdf/2402.05098
