Sparse structureVAE train

When training Sparse structureVAE,
TRELLIS/TRELLIS/trainers/vae/sparse_structure_vae py
z, mean, logvar = self.training_models['encoder'](ss.float(), sample_posterior=True, return_raw=True)
True is used, which is a common way to train VAE. When training Sparse structureVAE,
TRELLIS/TRELLIS/trainers/vae/sparse_structure_vae py
z, mean, logvar = self.training_models['encoder'](ss.float(), sample_posterior=True, return_raw=True)
True is used, which is a common way to train VAE.

But when training the SparseStructureFlowModel model,

Here the latent is read from the processed file, but datasets uses False when reading the encoder latent, which means only taking the mean and not re-parameterizing.

So I took sample_posterior=False, got ss_latents, and then entered it into Sparse structureVAE's decoder, and got the following result

When I apply sample_posterior=True, the posterior is normal

I would like to ask how your SparseStructureFlowModel is trained. Is the result of sample_posterior=False used for gt? If this is gt, decoder should not be used, because the distribution fitted during training is different

Compare TRELLIS/TRELLIS/trainers/vae/sparse_structure_vae with the dataset path used by SparseStructureFlowModel, focusing on sample_posterior=True during training versus False when reading processed latents. Determine which latent is intended as ground truth and whether decoding it should reproduce the reported behavior; done means the training and decoder expectations are clearly resolved.