By modeling the score function instead of the density function, we can sidestep the difficulty of intractable normalizing constants. How?
Langevin dynamics?
https://yang-song.net/blog/2021/score/
