Out-of-distribution (OOD) 3D relighting requires novel view synthesis under unseen lighting conditions that differ significantly from the observed images. Existing relighting methods, which assume consistent light source distributions between training and testing, often degrade in OOD scenarios. We introduce MetaGS to tackle this challenge from two perspectives. First, we propose a meta-learning approach to train 3DGaussian splatting, which explicitly promotes learning generalizable Gaussian geometries and appearance attributes across diverse lighting conditions, even with biased training data. Second, we embed fundamental physical priors from the Blinn-Phong reflection model into Gaussian splatting, which enhances the decoupling of shading components and leads to more accurate 3D scene reconstruction. Results on both synthetic and real-world datasets demonstrate the effectiveness of MetaGS in challenging OOD relighting tasks, supporting efficient point-light relighting and generalizing well to unseen environment lighting maps.

Forthenotations of directions, we use the convention that both the incident and outgoing rays point away from a scattering location. Spherical Harmonics (SH) are orthonormal basis defined on complex numbersovertheunitsphere. Since they were designed for scenes with solid objects, we adapt them to cope with participating media. Our implementation of the Neural Reflectance Field [2] baseline uses the same neural network architecture and positional encoding asinthe original paper. In addition, we employ a visibility MLP [3]tocompute a1-Dvisibility anda1-Dexpected termination depth.

Inthispaper,weaddress the"dual problem" ofmulti-viewscene reconstruction in which we utilize single-view images captured under different point lights to learnaneural scene representation. Different fromexisting single-viewmethods which can only recover a 2.5D scene representation (i.e., a normal / depth map for the visible surface), our method learns a neural reflectance field to represent the3Dgeometry andBRDFsofascene.

We cut off the gradient from the render loss to the visibility network. Our inference uses the same setting as the training. Specifically, we implement the dual-network design with a coarse network and a fine network. For the "env + point" illumination, we set the number of first indirect bounces to 32 . F .1 Scenes trained on the "point" Each test view has a new point light.