1 of 1

Generalizing 3D Gaussian Splatting for OOD Views

Srinath Ravi, Fernando de la Torre

Carnegie Mellon University

Motivation

SOTA Method

Dataset Creation

RefineGS

Treats artifacts as structured noise and learns a feed-forward correction model

Core Idea

Ablations

Limitations

​

​

Motivation

3D Gaussian Splatting (3DGS) produces high-quality novel views only within training distribution and fails under out-of-distribution (OOD) viewpoints

RefineGS is the first framework to reframe OOD correction as a 3D-consistent restoration task rather than a re-optimization problem. By decoupling geometry from appearance, the model leverages feed-forward spatial priors to anchor structural integrity while concurrently resolving photometric artifacts. This representation-agnostic approach recovers high-frequency detail in a single pass.

Lacks real-time usability (0.5 FPS) and fails to leverage reference-view context

Method

PSNR ↑

SSIM ↑

Inference speed (FPS)

Noisy Input (Perturbed 3DGS)

21.94

0.717

—

Difix3D+ (SOTA)

20.65

0.548

0.5

Ours

23.56

0.724

20.45

  • Requires high pixel-covisibility between input and reference views
  • FPS is limited by inference resolution
  • Can only refine but cannot recover missing geometry

Ours vs Diffix3D+