Test and Train models diverging

[RICE-G]

My research lab has been using cellbender for awhile as part of our pre-processing pipeline for our single cell RNA seq data. We've had no issues with it in the past but recently I am having trouble with the test/train models diverging and cellbender calling too many cells in 6/8 samples. The one difference in these samples is that they were generated using 10x Genomics' new On Chip Multiplexing technology, where 4 samples are pooled into a single set and then later demultiplexed with the 10x Genomics cell ranger multi pipeline (https://www.10xgenomics.com/support/software/cell-ranger/latest/analysis/running-pipelines/cr-multi). This is our first time using demultiplexed data with cellbender, which in theory, shouldn't be any different from what we've done before as the input of the individual samples is the same once we demultiplex with cell ranger. If anyone has had this issue before and has any suggestions of what has worked for them we would really appreciate it! Screenshots included are a summary of the cellranger output, the ELBO and UMI curve plots generated from cellbender, and a summary of the output metrics generated by cellbender.