Report
WASABI trains diffusion models to sample from Boltzmann distributions
WASABI's researchers say trained models can generate new samples without further energy or gradient evaluations.
TLDR
WASABI uses a target distribution's energy function to train a diffusion model while avoiding a difficult calculation of the model's own log density. On alanine dipeptide, its authors report that it outperformed the best reported baselines using an estimated 3.3 times fewer energy evaluations. They also report competitive performance on interacting-particle benchmarks.
Combined views
—
2 Sources, first seen ago
