Artifact Elimination in 3D Gaussian Splatting via Multi-Scale Feature Maps

3D Gaussian Splatting(3DGS) has received widespread attention due to its high fidelity and efficient rendering in recent years. However, since 3DGS solely relies on image reconstruction loss without accounting for illumination variations, it tends to generate minor floating artifacts and exhibits measurable quality degradation in detail when reconstructing scenes under varying lighting conditions. To address this problem and improve the robustness, we propose a Gaussian Splatting reconstruction representation method based on multi-scale feature maps. Specifically, we propose (1) a feature map-based multi-scale upsampling module that takes as input both the feature maps generated by the 3DGS rasterizer and the ground truth from existing encoders, ultimately outputting a multi-scale feature loss for 3DGS training, and (2) an adaptive illumination model that dynamically regulates exposure to enhance robustness under varying lighting conditions. Experimental results on three public datasets demonstrate that our method effectively eliminates illumination artifacts, enhances reconstruction fidelity in detailed regions and exhibits better robustness, outperforming existing NeRF-based and 3DGS-based approaches.