WNN (Seurat v4) vs. totalVI (scvi-tools) for CITE-seq data

I want to build a UMAP from CITE-seq data with a joint embedding of the scRNA-seq and protein ab data. What’s the ‘best’ method in terms of representing the most accurate embedding? In the totalVI paper, they say totalVI > Seurat v3. In the Seurat v4 paper, they say Seurat v4 > totalVI, but is there any independent benchmarks or other informed views? thanks