SC-VAE is the sparse compression autoencoder in TRELLIS.2 that maps O-Voxel data to and from compact structured latents.

Role
Encode and decode latents
Data
Sparse 3D structure and attributes
Spatial downsampling
16× in TRELLIS.2
Type
Variational autoencoder

Encode, generate, decode

The encoder compresses structured 3D data; generative models operate in that latent space; the decoder reconstructs shape or texture information for asset export.

Compression is a design tradeoff

A compact latent makes large-scale generation practical, while the autoencoder design determines which geometric and material details survive reconstruction.