GS-Voxel: Fitting-Free Structured Latents for Large-Scale 3DGS Generation
The paper turns irregular 3D Gaussian reconstructions into sparse voxel latents without per-scene fitting.
GS-Voxel keeps selected Gaussian primitives’ sub-voxel positions and rendering attributes while organizing them into active voxels. Its factorized VAE encodes geometry and local attributes separately, so latent size scales with occupied voxels instead of a fixed primitive budget. The authors train image-conditioned flow models in that latent space for aerial 3DGS scene generation. They also use overlap-aware tiled inference to extend generation over larger areas from satellite-view conditioning images. HF Daily Papers' note
GS-Voxel keeps selected Gaussian primitives’ sub-voxel positions and rendering attributes while organizing them into active voxels. Its factorized VAE encodes geometry and local attributes separately, so latent size scales with occupied voxels instead of a fixed primitive budget. The authors train image-conditioned flow models in that latent space for aerial 3DGS scene generation. They also use overlap-aware tiled inference to extend generation over larger areas from satellite-view conditioning images. HF Daily Papers' note
score 4