Improved off-policy training of diffusion samplers

Neural Information Processing Systems 

We study the problem of training diffusion models to sample from a distribution with a given unnormalized density or energy function.