Sampling from the Boltzmann distribution is a fundamental task in computational chemistry. It is a probability distribution over a state space corresponding to the equilibrium distribution of states that the system tends to over time. The Boltzmann distribution is given by: $p(x) = \frac{1}{Z}e^{-\beta H(x)}, $where $Z = \int e^{-\beta H(x)}dx$ is the partition function, $\beta$ is the inverse temperature, and $H(x)$ is the Hamiltonian energy of the state $x$. However, direct sampling from the Boltzmann distribution is generally infeasible due to the difficulty of computing the partition function $Z$. Denoising diffusion probabilistic models have emerged as a powerful tool for generating samples from complex probability distributions. The fundamental idea behind these models is to transform a simple initial probability distribution into the target distribution through a sequence of small, reversible diffusion steps such that training amounts to better approximation to the target distribution. We derive a constraint expression for continuous graph generative models that formally samples from the Boltzmann Distribution even when the data distribution does not follow it.