Abstract: To address the limitations of traditional Tensor Radiation Fields (TensoRF) in view synthesis tasks—specifically, their constrained expressiveness for high-dimensional data, weak capability in capturing high-frequency features, and significant memory overhead due to large model size—we construct a Lightweight Compact TensoRF (LC-TensoRF) model for fast, high-quality synthesis of novel views in complex 3D scenes. First, a novel tensor decomposition method, M2, is proposed to achieve low-rank tensor representation of high-dimensional data, enabling decomposition of 3D scenes in a more compact and flexible manner. Second, a Tensor Decoder Enhancement Module (TDEM) is developed. This module mitigates low-frequency bias during network training by reparameterizing MLP weights using fixed Fourier bases. It further incorporates gating mechanisms and residual network structures to enhance the model’s sensitivity to scene colors, thereby strengthening its ability to capture high-frequency information (particularly textural details) in scene appearance. Finally, storage overhead under M2 decomposition and TDEM is reduced by searching for the optimal voxel pruning rate, effectively minimizing model size while preserving high rendering quality. Extensive experiments on public datasets (Synthetic, Replica, and UrbanScene3D) demonstrate that the proposed model significantly improves rendering quality while reducing model size by approximately 50%.