Improbably huge log-fold change values

[LLansing]

In an ANCOM-BC2 output (res_pair table), a few of my taxa have improbably huge lfc values (-146 and -4696). In many of my samples the taxon has a 0 abundance, but there are still enough samples with non-zero values to pass the default prv_cut of 0.1.

By plotting the bias-corrected log abundances (image) and/or the raw abundances of these taxa across my group variable as box plots, I can see that the distributions don't seem to support such huge lfc values.

3/4 of the results with huge lfc also have significant adjusted p-values AND passed the pseudo-sensitivity test (diff & passed.ss):

What could be the source of my analysis producing these values? Does it have to do with low-numbers of non-zero samples within the groups? For reference, I have ~150 samples in this example and thus only a fairly small proportion have non-zero values, as shown by the grey dots on the box plots.