Yep, it is a very good approach to try on a small model and sample from it, then compare with the ADVI result. But be aware that even if it works in small model, in larger model ADVI with mean-field the variances of the parameters are likely underestimated. You should also try full-Rank ADVI.