Existing methods for joint RGB and X-ray 3D reconstruction typically train separate models for each modality. That leads to redundant geometry, alignment artifacts, and weak coupling between external appearance and internal structure. X-Splat instead builds a unified representation where all Gaussian primitives share a single set of geometric parameters, and only surface-proximal primitives carry appearance attributes.

The authors avoid aligning two independently optimized sets of Gaussians. Instead, they perform cross-modal registration using clean multi-view stereo and X-ray geometric proxies, then jointly optimize the unified representation with supervision from both modalities. A surface-depth consistency loss derived from X-ray geometry further guides the RGB reconstruction.

In experiments on the InsideOut dataset and the authors' Flower & Insect dataset, the framework reports more robust cross-modal alignment, higher RGB rendering quality without compromising X-ray reconstruction, and a more compact representation than maintaining two independent geometries. The paper appears at SIGGRAPH Asia 2026.