Résumé : Diffusion models in the literature are optimized with various objectives that are special cases of a weighted loss, where the weighting function specifies the weight per noise level. Uniform weighting corresponds to maximizing the ELBO, a principled approximation of maximum likelihood. In current practice diffusion models are optimized with non-uniform weighting due to better results in terms of sample quality. In this work we expose a direct relationship between the weighted loss (with any weighting) and the ELBO objective. We show that the weighted loss can be written as a weighted integral of ELBOs, with one ELBO per noise level. If the weighting function is monotonic, then the weighted loss is a likelihood-based objective: it maximizes the ELBO under simple data augmentation, namely Gaussian noise perturbation. Our main contribution is a deeper theoretical understanding of the diffusion objective, but we also performed some experiments comparing monotonic with non-monotonic weightings, finding that monotonic weighting performs competitively with the best published results.