GaussianPainter: Painting Point Cloud into 3D Gaussians with Normal Guidance

• DOI: 10.1609/aaai.v39i10.33172
• Publication Date: 2025-04-11T11:52:44Z
• Authors: Jingqiu Zhou; Lue Fan; Xuesong Chen; Linjiang Huang; Si Liu; Hongsheng Li

In this paper, we present GaussianPainter, the first method to paint a point cloud into 3D Gaussians given reference image. GaussianPainter introduces an innovative feed-forward approach overcome limitations of time-consuming test-time optimization in Gaussian splatting. Our addresses critical challenge field: non-uniqueness problem inherent large parameter space This space, encompassing rotation, anisotropic scales, and spherical harmonic coefficients, rendering similar images from substantially different fields. As result, networks face instability when attempting directly predict high-quality fields, struggling converge on consistent parameters for output. To address issue, propose estimate surface normal each determine its rotation. strategy enables network effectively remaining constrained space. We further enhance our with appearance injection module, incorporating image fields via multiscale triplane representation. successfully balances efficiency fidelity generation, achieving high-quality, diverse, robust content creation clouds single forward pass. A video is provided supplementary material more detailed explanation method.