Density-driven NeRF for Low-light Scene Synthesis

Neural Radiance Fields (NeRF) shows impressive results in novel view synthesis under normal illumination. However, in low-light scenes, the limited pixel values reduce the influence of volume density in rendering, leading to inaccurate radiance field learning. To address this issue, we propose a Density-Driven NeRF that improves performance in low-light conditions. Our method uses a pre-trained NeRF on normal-light images as a density teacher and applies the L2 loss to encourage the student network to match the teacher’s predicted density. This leads to accurate volume rendering despite limited visual cues. Experimental results on synthetic low-light datasets show that our approach improves RGB image quality evaluated by PSNR, SSIM and LPIPS, and yields better geometric consistency assessed by RMSE and δ-accuracy of the predicted density.