Novel view synthesis under sparse-view conditions remains a highly challenging task, primarily due to the difficulty of accurately modeling 3D scene structures while preserving fine details. In this paper, we propose a geometry-consistency-enhanced neural rendering framework based on random initialization to improve reconstruction accuracy and training stability. Our approach mainly involves designing a complementary supervision mechanism that combines a disparity consistency constraint with a pseudo-view depth supervision. The former introduces random translations to camera poses to construct pseudo-stereo pairs, guiding the model to learn cross-view geometric consistency; the latter generates auxiliary supervisory signals through depth back-projection, effectively reducing the direct reliance on the outputs of pre-trained depth networks. These two components work complementarily, significantly enhancing the model’s capacity to represent complex 3D structures and ensuring robust convergence. Extensive experiments conducted on two datasets, including LLFF and DTU, demonstrate the effectiveness and robustness of our method under sparse-view settings.