📝 Abstract:
3D Gaussian Splatting (3DGS) enables real-time novel view synthesis, but Gaussian representations reconstructed from low-resolution observations are typically sparse and lack fine geometric and appearance details, limiting the quality of high-resolution novel-view rendering. Existing 3DGS super-resolution methods either refine each scene through test-time optimization with external 2D priors, or learn feed-forward mappings from low-resolution images to high-resolution splats. Neither line of work learns a reusable cross-scene mapping that operates directly on pre-fit low-resolution Gaussian splat sets. This thesis proposes a generative two-stage framework for super-resolving pre-fit 3D Gaussian splat sets. In Stage I, a Point Transformer-based VQ-VAE learns a discrete latent codebook from high-quality Gaussian splats. In Stage II, a transformer predicts high-resolution latent codes from degraded low-resolution splat sets. Instead of directly regressing Gaussian attributes, the proposed approach formulates restoration as discrete latent code prediction, enabling feed-forward reconstruction without per-scene optimization or image-space priors. The proposed framework demonstrates the feasibility of generative super-resolution directly in the Gaussian parameter space, providing an alternative to optimization-based and image-driven approaches.