Research Paper

Mip-Splatting: Alias-free 3D Gaussian Splatting

A quality-focused paper that adds 3D and 2D filtering to reduce aliasing, dilation artifacts, and scale-dependent rendering problems in 3DGS.

November 2023AntialiasingarXiv:2311.16493

Detailed Reading

Mip-Splatting analyzes vanilla 3DGS as a sampling problem. When training images observe the scene at one scale and testing views zoom in or out, Gaussian primitives can contain frequencies that the input never properly sampled. This creates aliasing, dilation, or shimmering artifacts.

The paper adds two filters. A 3D smoothing filter constrains primitive size based on the sampling frequency implied by the training views, while a 2D Mip filter replaces the cruder dilation used during rasterization. Together they make projected splats behave more like properly filtered image samples.

This is not a flashy application paper; it is a renderer-quality paper. Its value is that it explains why a scene can be high-PSNR yet visually unstable under camera motion. For web viewers and interactive tools, that scale robustness is often what makes the difference between a demo and a usable asset.

Mip-Splatting studies a rendering-quality problem that appears once 3DGS is used across very different camera distances. A Gaussian optimized for training views may look sharp nearby but alias or dilate at other scales. The paper reframes splatting as a filtering problem, much like mipmapping reframed texture sampling.

The method introduces filtering in both 3D and 2D. In 3D, it constrains the representation so primitives do not become degenerate in ways that create unstable footprints. In 2D, it applies screen-space filtering to make projected splats respect pixel coverage, reducing aliasing when views zoom out or move.

The algorithmic lesson is that real-time rendering quality is not only about optimization loss. A model can match training photos and still violate sampling theory. Mip-Splatting adds the missing scale awareness, making Gaussian footprints behave more predictably under camera changes.

The paper matters for production viewers because aliasing is immediately visible in navigation, video export, and web playback. Its tradeoff is that filtering can soften details if tuned aggressively, so the method is best understood as balancing sharpness with stability rather than simply increasing PSNR.

What The Paper Does

Mip-Splatting targets a common artifact in 3DGS: views can look good at the training scale but break when camera distance or focal length changes.

The paper introduces a 3D smoothing filter and a 2D Mip filter so the representation respects sampling frequency more carefully.

Core Ideas

  • Constrains Gaussian primitive size according to sampling frequency induced by input views.
  • Replaces the original 2D dilation behavior with a Mip-style filter.
  • Improves robustness when testing across scales and camera distances.

Why It Matters

  • It tackles a fundamental rendering-quality issue rather than a narrow application.
  • Many practical viewers involve zooming, changing resolution, and moving cameras, so scale robustness matters.
  • It influenced later work on production-quality splat filtering and rendering stability.

Read This If

  • Your splats shimmer, blur, dilate, or alias when viewed at different scales.
  • You are building a renderer or trainer and care about robust camera movement.
  • You want a principled explanation of sampling artifacts in vanilla 3DGS.

Limitations And Caveats

  • Filtering improves scale behavior but does not solve geometry extraction or semantic editing.
  • Some quality gains may depend on dataset scale and capture distribution.
  • It adds conceptual and implementation complexity to the baseline renderer/trainer.