<p>Neural radiance fields (NeRFs) revolutionized 3D reconstruction by learning continuous scene representations from sparse 2D images, yet face two limitations: (i) inadequate modeling of channel–space interactions by MLP-based architectures and (ii) ineffective use of collective supervision information between distant pixels in the optimization process. To address these challenges, we propose mixed feature radiance field (MixeRF), which combines multi-resolution hash encoding with a multi-dimensional feature interaction network. The hierarchical hash table improves the training efficiency, while the cross-feature mechanism enhances the channel–space interdependence for fine scene detail recovery. We further design distribution-aligned normal distribution residual loss based on distributional alignment to perform global pixel supervision by predicting the minimization of the distributional distance between residuals and Gaussian noise. To balance optimization stability with reconstruction quality, we introduce a smooth approximation technique for differentiable loss function fine-tuning. Experiments on mixed indoor/outdoor datasets demonstrate consistent performance gains: +2.13 dB higher PSNR than Nerfacto on the Lego dataset and 2.2 dB higher PSNR than Instant-NGP on the Fern dataset. The smooth approximation technique incurs merely 1.15 dB average cross-dataset PSNR degradation (range: 0.06–0.96 dB) while preserving optimization stability.</p>