Neural Radiance Fields (NeRFs) have transformed 3D scene reconstruction, yet they struggle with scale-dependent artifacts. Standard NeRF models sample discrete points along rays, ignoring the conical volume a pixel actually represents. This leads to "aliasing": jagged edges in distant views and blurriness in close-ups. While Mip-NeRF addressed this using Integrated Positional Encoding, its reliance on slow Multi-Layer Perceptrons (MLPs) makes it computationally expensive. Conversely, modern grid-based methods like TensoRF offer 100x speedups but re-introduce aliasing by reverting to point-sampling. Mip-Grid bridges this gap by introducing anti-aliasing to high-speed grid representations.
Methodology: Multi-scale Generation via Convolution
The core innovation of Mip-Grid is the generation of multi-scale feature grids on the fly using learnable convolutions. Instead of storing multiple grids—which would be memory-prohibitive—the method utilizes a single shared high-resolution grid.
By leveraging factorized tensor representations (such as Vector-Matrix decomposition), Mip-Grid applies 1D or 2D depth-wise convolutions to the grid components. This approach is mathematically equivalent to 3D voxel convolution but significantly more efficient. During rendering, a scale-aware coordinate system interpolates between these generated features based on the pixel's footprint, ensuring smooth transitions across varying distances.
Experimental Results
The authors evaluated Mip-Grid on modified versions of the Blender and LLFF datasets, rendered at multiple resolutions (Full, 1/2, 1/4, and 1/8) to test multi-scale robustness. The primary implementation, Mip-TensoRF, was compared against the original Mip-NeRF and vanilla TensoRF.
Mip-Grid achieves state-of-the-art rendering quality while maintaining the efficiency of grid-based methods. Mip-TensoRF trains in approximately 16 minutes on a single GPU, compared to over 30 hours for Mip-NeRF, while delivering superior PSNR (35.23 vs 34.51). As shown in the quantitative comparison, Mip-TensoRF occupies the optimal region of the PSNR-vs-time plot. Qualitatively, the method effectively eliminates the "shimmering" and jagged edges common in point-sampled grid methods, particularly visible in thin structures at low resolutions.
Conclusion
Mip-Grid successfully integrates anti-aliasing into fast, factorized grid representations. By learning optimal blurring kernels and utilizing efficient convolutions, it provides a "plug-and-play" solution for high-quality, multi-scale neural rendering without the traditional trade-off between speed and visual fidelity.
Summary updated