Abstract
Recent years have witnessed remarkable success of 3D Gaussian Splatting(3DGS) in novel view synthesis, surpassing prior differentiable renderingmethods in both quality and efficiency. However, its training process suffersfrom coupled opacity-color optimization that frequently converges to localminima, producing floater artifacts that degrade visual fidelity. We presentStableGS, a framework that eliminates floaters through cross-view depthconsistency constraints while introducing a dual-opacity GS model to decouplegeometry and material properties of translucent objects. To further enhancereconstruction quality in weakly-textured regions, we integrate DUSt3R depthestimation, significantly improving geometric stability. Our methodfundamentally addresses 3DGS training instabilities, outperforming existingstate-of-the-art methods across open-source datasets.