We study generative modelling of correlated extragalactic foregrounds (tSZ plus CIB) on FLAMINGO simulations, separating by-construction versus learned statistics and supervised versus unsupervised use of truth ensembles. On twenty 5-degree patches at 150 GHz we benchmark scattering-transform (ST) synthesis against diffusion models (DDPM) and paired Gaussian fields. We introduce a phase-preserving joint multipole Cholesky match plus paired pixel histogram matching that forces agreement with truth on auto- and cross-spectra and one-point statistics, then compare non-by-construction ScatCov-only pipelines and an ensemble-mode variant usable without per-patch truth at inference. We discuss how microcanonical SC matching, trained DDPM, and our semi-supervised recipe sit on complementary axes for real-sky applications, and report a JAX implementation (jaxst) with large speed-ups over a PyTorch reference. See the PDF for equations, figures, and full numerical results.