The paper introduces TriFold, a native 3D Gaussian VAE designed to learn compact generative representations of complex 3D scenes. The authors note that native 3D Gaussian scenes are hard to encode because primitive counts vary widely, spatial density is nonuniform, and even neighboring primitives can differ in geometry and appearance. TriFold addresses this by imposing structural regularity on the latent space while keeping the spatial adaptivity of the Gaussians.

During encoding, TriFold uses cell-wise learned queries to aggregate projected Gaussians into multiple slots per cell, avoiding the information loss that comes from pooling a single feature per location. During decoding, occupancy prediction identifies active regions, and query-based set prediction with candidate selection reconstructs local Gaussian sets with spatially varying primitive counts.

Building on this representation, the authors train a latent flow-matching model for conditional scene-level 3DGS generation. Experiments reported in the paper show high-fidelity reconstruction and generation from a single image that provides only a partial view of the scene, which the authors say validates TriFold's effectiveness for native 3D Gaussian representation learning.