Mip-Splatting: Frequency-Aware Rendering

Updated 5 April 2026

Mip-Splatting is a rendering technique that introduces matched 3D and 2D closed-form filters to guarantee alias-free, artifact-free 3D synthesis.
It integrates a band-limiting 3D filter with a pixel-integrating 2D Mip filter, ensuring frequency consistency across varying scales and camera intrinsics.
Empirical evaluations show that Mip-Splatting maintains robust image quality under dynamic view changes and arbitrary zoom, enabling real-time applications.

Mip-Splatting is a rendering and scene representation technique designed to solve the challenges of anti-aliasing, scale-robustness, and frequency consistency in 3D Gaussian Splatting (3DGS) for novel view synthesis. By introducing closed-form frequency-domain filtering in both world (3D) and screen (2D) space, Mip-Splatting guarantees alias-free, artifact-free rendering across a broad range of camera intrinsics and sampling rates. This technique has become a foundational component in state-of-the-art differentiable rendering and SLAM pipelines, notably enabling robust performance under arbitrary zoom, dynamic view changes, and real-time applications (Yu et al., 2023, Li et al., 7 Mar 2026).

1. Motivation and Aliasing in 3D Gaussian Splatting

Traditional 3DGS represents a scene as a superposition of anisotropic 3D Gaussian primitives, each parameterized by center, covariance, color, and opacity. Standard rendering pipelines for 3DGS involve projecting each primitive into 2D image space, then compositing the projected splats in front-to-back order. To avoid the collapse of Gaussians to subpixel extents (causing holes or “eroded” features), 3DGS applies a fixed-dilation term in screen space. However, this causes severe artifacts when rendering at scales differing from training, as the fixed dilation induces “gap” formation (zoom-in) or “halo” blooming (zoom-out), and fails to adhere to any principled frequency constraint (Yu et al., 2023).

These artifacts arise from two fundamental issues:

The absence of an explicit 3D band-limit on the Gaussian primitives.
The use of an unphysical and scale-inconsistent 2D dilation filter in screen space.

Mip-Splatting directly addresses both deficiencies by introducing matched 3D and 2D closed-form filters (Yu et al., 2023).

2. 3D Smoothing Filter: Band-Limiting Primitive Frequency Content

Classical sampling theory (Nyquist–Shannon) indicates that the maximal reconstructible frequency is half the sampling frequency imposed by the imaging setup. For each Gaussian primitive, the relevant sampling bound is set by the densest (highest-frequency) view in which it is visible:

$\nû_k = \max_n \left[ \mathbb{1}_n(p_k) \cdot \frac{f_n}{d_{kn}} \right]$

where $f_n$ is the focal length, $d_{kn}$ is the depth of the primitive in the $n$ -th camera, and $\mathbb{1}_n$ encodes visibility. To eliminate aliasing, each primitive is convolved with an isotropic low-pass Gaussian filter of variance $\sigma_k^2 = (s/\nû_k) \cdot I$, yielding a closed-form band-limited primitive:

$G_k^{(\mathrm{reg})}(x) = \Bigg(\frac{|\Sigma_k|}{|\Sigma_k + (s/\nû_k) I|}\Bigg)^{1/2} \exp\bigg[-\frac{1}{2} (x-p_k)^T \big(\Sigma_k + (s/\nû_k) I\big)^{-1} (x-p_k)\bigg]$

The frequency response is likewise attenuated:

$H_k(\omega) = \exp\big[-2\pi^2 (s/\nû_k)\|\omega\|^2 \big]$

This guarantees no primitive ever encodes spatial frequencies above the training-set’s observable limit; high-frequency artifacts upon zoom-in are thus impossible (Yu et al., 2023).

3. 2D Mip Filter: Analytical Pixel Footprint Integration

To render models at lower pixel densities (zoom-out), proper anti-aliasing necessitates integrating each projected Gaussian over the area of a screen pixel. This “pixel box” filter is approximated analytically via a 2D Gaussian Mip filter:

$G_k^{(\mathrm{mip})}(x) = \Bigg( \frac{|\Sigma_k^{2\mathrm{D}}|}{|\Sigma_k^{2\mathrm{D}} + \sigma_\mathrm{mip}^2 I|} \Bigg)^{1/2} \exp\bigg[ -\frac12 (x-u_k)^T (\Sigma_k^{2\mathrm{D}} + \sigma_\mathrm{mip}^2 I)^{-1} (x-u_k) \bigg]$

where $\Sigma_k^{2\mathrm{D}}$ is the projected covariance and $f_n$ 0 the mean in image space. With a typical choice $f_n$ 1, this filter accurately simulates the effect of integrating over a pixel. The filtering equation attenuates high-frequency content analogously to a true box filter:

$f_n$ 2

Collectively, the 3D and 2D filters ensure the representation is always simultaneously band-limited in both the intrinsic and rendering domains (Yu et al., 2023).

4. Algorithmic Integration and Rendering Pipeline

Mip-Splatting is structurally lightweight, requiring only simple modifications to the standard 3DGS optimization and rendering loop:

Sampling Frequency Update: At regular intervals during training, compute each primitive’s maximal observable sampling frequency $f_n$ 3.
3D Filter Fusion: Each primitive’s covariance is incremented by the band-limiting variance, fusing the low-pass filter directly into its parameterization.
Rendering with Pre-filtered Gaussians: For each ray’s sample, the now band-limited primitive is projected to screen space and further pre-filtered using the 2D Mip filter.
Alpha-Blending Composition: Rendered color and opacity are composited front-to-back by the standard line integral for splatting.

The pseudocode provided in the original work illustrates the cycle over Gaussians and samples, including per-primitive and per-pixel filter application. During inference, the required filters are already fused into the parameters, making deployment efficient and scale-robust (Yu et al., 2023).

5. Empirical Evaluation, Comparison, and Extensions

Mip-Splatting is quantitatively validated on single-scale training and multi-scale testing, e.g., Blender Objects and Mip-NeRF 360 benchmarks. When training at full resolution and rendering at lower (or higher) resolutions, classic 3DGS PSNR drops precipitously (e.g., 29.26 dB to 19.63 dB for an 8× zoom-out). In contrast, Mip-Splatting maintains consistent image quality across scales (e.g., 29.91–29.61 dB, Table 5), outperforming standard 3DGS and alternatives such as EWA filtering. Structural stability is also confirmed visually—thin features remain crisp, while aliasing- and dilation-induced “bloom” or “gaps” are absent (Yu et al., 2023).

Comparative work, such as SA-GS, demonstrates that while Mip-Splatting requires retraining with scale-aware filters, similar frequency-consistent anti-aliasing can be achieved at test-time through adaptive 2D filtering and exact pixel integration (Song et al., 2024). The primary distinction is that Mip-Splatting fuses low-pass filtering into the learned model, providing strong guarantees against both high-frequency synthesis and rendering artifacts at the cost of slightly more complex training.

Key limitations and extensions:

The Gaussian Mip filter only approximates the ideal box; at extreme downsampling there can be slight amplitude errors.
Per-primitive frequency analysis incurs a moderate (5–10%) overhead, addressable via GPU acceleration or caching.
Potential improvements include switching to separable sinc- or anisotropic EWA-based filters, handling dynamic (time-varying) frequency bounds, or learning per-primitive pixel filters (Yu et al., 2023).

6. Broader Impact and Applications

Mip-Splatting’s frequency-aware construction allows for real-time alias-free novel view synthesis under arbitrary zoom, enabling its integration in VR, AR, and dynamic SLAM frameworks. Its principled architecture has made it a baseline for recent advancements, including training accelerations (e.g., Group Training with opacity-based prioritization (Wang et al., 2024)), frequency-aware SLAM systems (e.g., MipSLAM’s EAA and SA-PGO modules (Li et al., 7 Mar 2026)), and planar 2DGS anti-aliasing strategies (e.g., AA-2DGS (Younes et al., 12 Jun 2025)). As a result, it underpins robust, scalable rendering pipelines wherever view consistency and cross-resolution fidelity are critical.

7. Summary Table of Core Mip-Splatting Components

Component	Function	Formula/Operation
3D Smoothing Filter	Limits each primitive’s world-space frequency	$f_n$ 4
2D Mip Filter	Simulates accurate pixel integration in screen space	$f_n$ 5
Band-Limited Primitive Evaluation	Ensures no frequencies above $f_n$ 6 are synthesized	$f_n$ 7

Mip-Splatting’s coupling of frequency-domain analysis with closed-form filter design ensures a uniquely robust framework for scale-invariant, artifact-free differentiable rendering (Yu et al., 2023, Li et al., 7 Mar 2026).