Yes.
There are so many differences in the modeling, which results in a significant difference in the model's performance.
I can mention some differences (not all):
scPoli uses prototypes per cell type and employs a metric learning approach to form the latent space.
scPoli maps the covariates in a continuous latent space that reveals sample level variation.
I suggest reading the paper for more details.
Just curious, since they both based on cVAE, with cell-type data as conditioning input, is there any difference between their backbones or algorithms?