This paper presents a super-fast convergence approach to reconstructing radiance fields from a set of images, achieving NeRF-like quality in less than 15 minutes with a single GPU. The method uses a density voxel grid for scene geometry and a feature voxel grid with a shallow network for complex view-dependent appearance. Key contributions include post-activation interpolation on voxel density to produce sharp surfaces and robust optimization techniques to avoid suboptimal solutions. Evaluation on five datasets shows that the method matches or surpasses NeRF's quality while reducing training time significantly. The approach is efficient and effective, making it suitable for real-time applications and novel view synthesis.This paper presents a super-fast convergence approach to reconstructing radiance fields from a set of images, achieving NeRF-like quality in less than 15 minutes with a single GPU. The method uses a density voxel grid for scene geometry and a feature voxel grid with a shallow network for complex view-dependent appearance. Key contributions include post-activation interpolation on voxel density to produce sharp surfaces and robust optimization techniques to avoid suboptimal solutions. Evaluation on five datasets shows that the method matches or surpasses NeRF's quality while reducing training time significantly. The approach is efficient and effective, making it suitable for real-time applications and novel view synthesis.