SPAR3S: sparse autoregressive 3D scene generation from few views, no 3D supervision

kwangmoo_yi · x · 2026-09-05

Researchers from Naver introduce SPAR3S (arXiv:2609.03931), a sparse voxel-aligned 3D latent generative model that completes full 3D scenes from a few unconstrained views, decoded as 3D Gaussian Splats.

Core idea: formulate scene generation in a compact, structured voxel-aligned latent space where only occupied voxels are represented; completion reduces to predicting missing latent tokens and their spatial support.
No ground-truth 3D supervision needed: the sparse latent space is learned directly from multi-view images via photometric supervision through differentiable 3D Gaussian Splatting.
A masked autoregressive transformer jointly models voxel occupancy and latent token values, enabling efficient, spatially consistent generation of unseen regions.
Unlike feed-forward reconstruction (limited to visible content) or dense volumetric generative models (computationally costly, data-hungry), SPAR3S addresses both limitations at once.

Original post →

SPAR3S: sparse autoregressive 3D scene generation from few views, no 3D supervision

More from Multimodal