Research map / 3D vision / Gaussian splatting
Mip-Splatting: Alias-free 3D Gaussian Splatting
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Mip-Splatting makes 3D Gaussian Splatting alias-free across sampling rates by adding a 3D smoothing filter bounded by training-view sampling frequency and replacing 2D screen-space dilation with a 2D Mip (box-approximating) filter.
Problem
3D Gaussian Splatting produces dilation, erosion, and aliasing artifacts when rendering at sampling rates (focal length or camera distance) different from those used in training, i.e., poor out-of-distribution generalization across scales.
Approach
Constrain each 3D Gaussian's maximal frequency with a 3D low-pass filter whose size is set from the maximal sampling rate of the training views (Nyquist), and replace the heuristic 2D screen-space dilation with a 2D Gaussian approximating the pixel-area box filter, enabling alias-free rendering at unseen scales without multi-scale training.
Key results
- Single-scale training with multi-scale testing on Blender: Mip-Splatting averages 31.97 PSNR vs 29.77 for 3DGS and 33.01 for 3DGS+EWA.
- On Mip-NeRF 360 single-scale training (1x) with 8x zoom-in testing, Mip-Splatting averages 27.37 PSNR vs 23.25 for 3DGS and 25.43 for 3DGS+EWA.
- Same-scale testing on Mip-NeRF 360 gives 27.79 PSNR, on par with 3DGS (27.21) and 3DGS+EWA (27.77).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add 3D smoothing and replace 2D dilation with a 2D Mip filter | 3DGS · Blender, single-scale train/multi-scale test; 1/2, 1/4, and 1/8 resolution | PSNR: 30.66 -> 34.00; 21.38 -> 31.85; 17.69 -> 28.67 | Helped |
| Add 3D smoothing and replace 2D dilation with a 2D Mip filter | 3DGS · Mip-NeRF 360, single-scale train/multi-scale test; 2×, 4×, and 8× resolution | PSNR: 23.50 -> 27.39; 20.71 -> 26.47; 19.59 -> 26.22 | Helped |
| Add 3D smoothing and replace 2D dilation with a 2D Mip filter | 3DGS · Blender, multi-scale train/test | PSNR (Avg.): 29.77 -> 34.56 | Helped |
| Add 3D smoothing and replace 2D dilation with a 2D Mip filter | 3DGS · Mip-NeRF 360, single-scale train/same-scale test | PSNR: 27.21 -> 27.79 | No clear effect |
Surprising findings
- Simply removing screen-space dilation from 3DGS causes optimization difficulties on complex scenes due to excessive numbers of small Gaussians.
Limitations
- Gaussian filter approximating the box filter introduces errors when Gaussians are small in screen space, increasing with zoom-out.
- Slight training overhead from recomputing per-Gaussian sampling rates in PyTorch every m=100 iterations.
- Rendering at higher resolutions is a super-resolution task; the method cannot hallucinate absent high-frequency details.
Experimental setup
Datasets: Blender dataset Mip-NeRF 360 dataset
Benchmarks: Blender dataset Mip-NeRF 360 dataset
Metrics: PSNR SSIM LPIPS
Baselines: NeRF Mip-NeRF Plenoxels TensoRF Instant-NGP Tri-MipRF 3DGS 3DGS + EWA Mip-NeRF 360 Zip-NeRF
Builds on: 3D Gaussian Splatting EWA splatting Mip-NeRF Zip-NeRF
Closest papers by meaning
- Anti-Aliased 2D Gaussian Splatting (NeurIPS 2025 · similarity 0.92)
- Multi-Scale 3D Gaussian Splatting for Anti-Aliased Rendering (CVPR 2024 · similarity 0.90)
- Frequency-Aware Density Control via Reparameterization for High-Quality Rendering of 3D Gaussian Splatting (AAAI 2025 · similarity 0.90)
- GaussianSpa: An "Optimizing-Sparsifying" Simplification Framework for Compact and High-Quality 3D Gaussian Splatting (CVPR 2025 · similarity 0.89)
- Speedy-Splat: Fast 3D Gaussian Splatting with Sparse Pixels and Sparse Primitives (CVPR 2025 · similarity 0.89)
- Generative Sparse-View Gaussian Splatting (CVPR 2025 · similarity 0.89)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.