Search papers, labs, and topics across Lattice.
This paper introduces 3D Morphological Perturbations as an optimization-free regularizer for 3D scene representations, addressing the artifacts that arise in sparse-view settings without the need for costly per-scene reconstructions. By treating Gaussian components of 3D Gaussian Splatting as fundamental units, the method applies morphological perturbations to enhance spatial consistency across views. The approach not only reduces mean depth error by 12.5% compared to existing artifact refiners but also improves downstream robotics policy success rates by up to 8.0% across multiple manipulation tasks.
Eliminating the need for per-scene optimization, this method boosts 3D model performance while enhancing visual fidelity and spatial consistency.
3D scene representations like NeRF and 3D Gaussian Splatting (3DGS) suffer severe artifacts in sparse-view settings. Recent generative 3D artifact fixers attempt to address this, but rely on paired corrupted and clean renders requiring costly, per-scene reconstructions across varying view configurations. While 2D image augmentations act as instant regularizers, no explicit equivalents exist for 3D representations to preserve spatial consistency across views, an essential property for 3D-aware training. We propose 3D Morphological Perturbations as an optimization-free regularizer that preserves spatial consistency. Leveraging explicit 3DGS, we treat each Gaussian as a fundamental building block - analogous to a 2D pixel - and apply perturbations across its morphological parameter space via scale, rotation, and pruning. Our method eliminates per-scene 3DGS optimization loops from dataset curation while enabling models to learn stronger geometric priors than sparse-view baselines in diagnostic ablations conducted on a lightweight video diffusion sandbox. Scaled to a 14B-parameter video model via ControlNet, our approach maintains visual fidelity while reducing mean depth error by 12.5% over state-of-the-art image-to-image 3D artifact refiners, ultimately boosting downstream robotics policy success rates by up to 8.0% across 3 of 4 manipulation tasks.
