A lot of my research over the years has involved fitting images. Usually, I use the traditional assumption that the conditional prior for the data pixels given the model pixels is iid normal with mean zero. But in some cases one might want to use a correlated distribution instead. That’s a technical challenge which I’ve been trying to solve from several angles (though good solutions might be well known to people who know these things).
I got one of those working the other day, and was rather amused at some of the output. Check out these residuals of one posterior sample of a gravitational lens fit (bottom right). The “correlated noise” likelihood is quite happy to call this a decent fit! 🙂