AdaptiveShader: Adaptive Forward and Deferred Shading for Gaussian Splatting

Recent advances in 3D Gaussian Splatting have enabled real-time high-quality rendering, yet existing methods struggle to faithfully capture reflective surfaces. Prevailing methods often hardwire the shading strategy as either single-branch or rigid dual-branch, which limits their performance on reflective surfaces. In this work, we revisit the overlooked potential of forward shading and propose AdaptiveShader, a pixel-level adaptive shading framework that dynamically balances forward and deferred rendering. We first introduce an oracle model that selects per-pixel shading results by comparing both branches against ground-truth images. This oracle reveals the untapped benefits of forward shading and motivates our design. Since such ground-truth supervision is unavailable at test time, we further develop an affinity-based selection mechanism, which learns to assign pixel-wise weights to forward and deferred shading outputs. By adaptively leveraging the complementary strengths of both branches, our method outperforms existing methods in average novel-view synthesis performance on the two synthetic datasets and achieves the best average PSNR and SSIM on the Ref-Real dataset. Extensive experiments demonstrate that our method delivers strong average rendering fidelity, while the Ref-Real LPIPS results expose a remaining perceptual limitation.