We present a super-fast convergence approach for reconstructing the per-scene radiance field from a set of images with known poses. This method achieves NeRF-comparable quality and converges rapidly in less than 15 minutes with a single GPU. Our approach uses a density voxel grid for scene geometry and a feature voxel grid with a shallow network for view-dependent appearance. We introduce post-activation interpolation on voxel density to produce sharp surfaces and robustify optimization by imposing priors. Evaluation on five inward-facing benchmarks shows our method matches or surpasses NeRF's quality with significantly faster training times. Our method uses a grid resolution of 160³, which is much lower than previous methods. We achieve visual quality comparable to NeRF at 45× faster rendering speeds and do not require cross-scene pretraining. Our key contributions include two priors to avoid suboptimal geometry in direct voxel density optimization and post-activated voxel-grid interpolation for sharp boundary modeling. Our method reduces training time from 10–20 hours to 15 minutes on a single GPU, achieving NeRF-quality results with much faster convergence. We also show that our method can achieve high-quality results with fewer voxels and better performance in real-world scenarios. Our approach combines the advantages of explicit and implicit representations, enabling efficient and accurate scene reconstruction.We present a super-fast convergence approach for reconstructing the per-scene radiance field from a set of images with known poses. This method achieves NeRF-comparable quality and converges rapidly in less than 15 minutes with a single GPU. Our approach uses a density voxel grid for scene geometry and a feature voxel grid with a shallow network for view-dependent appearance. We introduce post-activation interpolation on voxel density to produce sharp surfaces and robustify optimization by imposing priors. Evaluation on five inward-facing benchmarks shows our method matches or surpasses NeRF's quality with significantly faster training times. Our method uses a grid resolution of 160³, which is much lower than previous methods. We achieve visual quality comparable to NeRF at 45× faster rendering speeds and do not require cross-scene pretraining. Our key contributions include two priors to avoid suboptimal geometry in direct voxel density optimization and post-activated voxel-grid interpolation for sharp boundary modeling. Our method reduces training time from 10–20 hours to 15 minutes on a single GPU, achieving NeRF-quality results with much faster convergence. We also show that our method can achieve high-quality results with fewer voxels and better performance in real-world scenarios. Our approach combines the advantages of explicit and implicit representations, enabling efficient and accurate scene reconstruction.