Spatial transcriptomics assays measure a few hundred genes per cell, while single-cell RNA sequencing measures the full transcriptome but discards spatial context. Deep generative models such as gimVI bridge the two by learning a shared latent representation and imputing unmeasured genes in the spatial modality. We have identified failure modes in which imputation degrades systematically under conditions that are not currently diagnosed by standard benchmarks. This project characterizes those failure modes and develops corrections to the model architecture and training objective, evaluated against held-out ground truth. Development uses mouse tumor tissue, cell lines, and synthetic human single-cell data. No sensitive or personally identifiable data are processed. The improved method is intended for publication and open release.