Toward multi-view texture consistency in novel view synthesis from sparse-view 3DGS

3D Gaussian Splatting (3D-GS) and its variants achieve strong performance in real-time novel view synthesis. However, their performance drops significantly under sparse input views due to overfitting and limited visual cues. In particular, existing methods struggle to reconstruct detailed geometry and suffer from color artifacts in rendered novel views. To address these issues, we propose Multi-view Texture Consistent Gaussian Splatting (MTC-GS), which introduces effective regularization to ensure coherence in unseen views. Specifically, our main contributions are Local Geometry Regularization (LoGR) and Coherent Texture Regularization (CoTR). LoGR corrects inaccurate geometry to provide more reliable local geometric guidance, and CoTR is a novel coherent texture constraint that refines suboptimal Gaussians. Experiments on the LLFF and Mip-NeRF360 datasets show that MTC-GS achieves state-of-the-art performance, outperforming the baseline method FSGS by 0.37 dB and 0.15 dB in PSNR, respectively, and producing photorealistic renderings with minimal artifacts.