Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Mipmap-GS: Let Gaussians Deform with Scale-specific Mipmap for Anti-aliasing Rendering

Jiameng Li, Yue Shi, Jiezhang Cao, Bingbing Ni, Wenjun Zhang, Kai Zhang, Luc Van Gool

TL;DR

This work tackles aliasing in 3D Gaussian Splatting when rendering at unseen scales by introducing Mipmap-GS, a plug-in that deform Gaussians through scale-specific mipmap-like pseudo-ground-truth. The method constructs scale-aware supervision at test time and optimizes Gaussians with a gradient-guided loss to match the scale-specific pseudo-GT, achieving rapid convergence (≈1K iterations) and producing a more compact representation via active pruning. It reports substantial gains in PSNR for both zoom-in and zoom-out cases on NeRF Synthetic (averages of 9.25 dB and 10.40 dB, respectively) and demonstrates robust improvements across datasets, including Mip-NeRF 360. The approach offers a practical, low-overhead enhancement that can be integrated into existing 3DGS pipelines to reduce aliasing and preserve fine details across scales, facilitating interactive rendering and dynamic scenes.

Abstract

3D Gaussian Splatting (3DGS) has attracted great attention in novel view synthesis because of its superior rendering efficiency and high fidelity. However, the trained Gaussians suffer from severe zooming degradation due to non-adjustable representation derived from single-scale training. Though some methods attempt to tackle this problem via post-processing techniques such as selective rendering or filtering techniques towards primitives, the scale-specific information is not involved in Gaussians. In this paper, we propose a unified optimization method to make Gaussians adaptive for arbitrary scales by self-adjusting the primitive properties (e.g., color, shape and size) and distribution (e.g., position). Inspired by the mipmap technique, we design pseudo ground-truth for the target scale and propose a scale-consistency guidance loss to inject scale information into 3D Gaussians. Our method is a plug-in module, applicable for any 3DGS models to solve the zoom-in and zoom-out aliasing. Extensive experiments demonstrate the effectiveness of our method. Notably, our method outperforms 3DGS in PSNR by an average of 9.25 dB for zoom-in and 10.40 dB for zoom-out on the NeRF Synthetic dataset.

Mipmap-GS: Let Gaussians Deform with Scale-specific Mipmap for Anti-aliasing Rendering

TL;DR

This work tackles aliasing in 3D Gaussian Splatting when rendering at unseen scales by introducing Mipmap-GS, a plug-in that deform Gaussians through scale-specific mipmap-like pseudo-ground-truth. The method constructs scale-aware supervision at test time and optimizes Gaussians with a gradient-guided loss to match the scale-specific pseudo-GT, achieving rapid convergence (≈1K iterations) and producing a more compact representation via active pruning. It reports substantial gains in PSNR for both zoom-in and zoom-out cases on NeRF Synthetic (averages of 9.25 dB and 10.40 dB, respectively) and demonstrates robust improvements across datasets, including Mip-NeRF 360. The approach offers a practical, low-overhead enhancement that can be integrated into existing 3DGS pipelines to reduce aliasing and preserve fine details across scales, facilitating interactive rendering and dynamic scenes.

Abstract

3D Gaussian Splatting (3DGS) has attracted great attention in novel view synthesis because of its superior rendering efficiency and high fidelity. However, the trained Gaussians suffer from severe zooming degradation due to non-adjustable representation derived from single-scale training. Though some methods attempt to tackle this problem via post-processing techniques such as selective rendering or filtering techniques towards primitives, the scale-specific information is not involved in Gaussians. In this paper, we propose a unified optimization method to make Gaussians adaptive for arbitrary scales by self-adjusting the primitive properties (e.g., color, shape and size) and distribution (e.g., position). Inspired by the mipmap technique, we design pseudo ground-truth for the target scale and propose a scale-consistency guidance loss to inject scale information into 3D Gaussians. Our method is a plug-in module, applicable for any 3DGS models to solve the zoom-in and zoom-out aliasing. Extensive experiments demonstrate the effectiveness of our method. Notably, our method outperforms 3DGS in PSNR by an average of 9.25 dB for zoom-in and 10.40 dB for zoom-out on the NeRF Synthetic dataset.
Paper Structure (18 sections, 5 equations, 11 figures, 4 tables, 1 algorithm)

This paper contains 18 sections, 5 equations, 11 figures, 4 tables, 1 algorithm.

Table of Contents

  1. Introduction
  2. Related Work
  3. Novel View Synthesis
  4. 3D Gaussian Splatting
  5. Zoom-out and Zoom-in
  6. Preliminaries
  7. 3D Gaussian Splatting
  8. Motivation
  9. Proposed Method
  10. Mipmap-like Pseudo-GT
  11. Scale-Adaptive 3D Gaussians
  12. Optimization
  13. Experiments
  14. Experimental Setup
  15. Comparisons with State-of-the-Art
  16. ...and 3 more sections

Figures (11)

  • Figure 1: Aliasing at different observation distances.
  • Figure 2: Comparison of 3DGS, Mip-Splatting and ours in zoom-out and zoom-in scenarios. 3DGS (top) lacks scale sensitivity and exhibits strong degradation. Mip-Splatting (middle) introduces a 2D Mip filter and 3D smooth filter to regulate scales heuristically while leaving 3D Gaussians unaware of the varying footprint in pixel space. Ours (bottom) adopts scale-adaptive Gaussians for explicit pixel coverage optimization. The projection in previous pixel coverage is illustrated in the dashed line.
  • Figure 3: A toy 3DGS model under multi-scale rendering. The splatted Gaussians (dark) go through 2D dilation (light) for faithful rendering (a). However, the varying shrinkage bias and constant dilation cause zoom-in spikes (c) and zoom-out thickness (d).
  • Figure 4: Pipeline. We adjust single-scale trained base Gaussians for arbitrary zoom factors supervised by mipmap-like pseudo-GT. The core component is the mipmap proposal module, which constructs scale-specific mipmap by upsampling or downsampling novel views.
  • Figure 5: Visualization of Gaussians. 3DGS kerbl20233d use fixed Gaussians for multi-scale rendering, while our adaptive adjustment provides deformable Gaussians towards various zoom factors.
  • ...and 6 more figures