Extragalactic foregrounds in cosmic microwave background (CMB) observations are both a source of information and nuisnace. Their effective modeling is increasingly important thanks to the low-noise observations from current and upcoming experiments.We explore the application of wavelet flow models for the novel task of training flow-based models on multi-component CMB foreground maps.We introduce a hybrid prior in the context of wavelet flows and show that they outperform conventional priors. Specifically, the mean power spectra of the generated maps are within few percent of the inputs' power spectra on all scales.