Alias-free 4D Gaussian Splatting

TL;DR

Alias-free-4D Gaussian Splatting addresses artifacts that arise in 4D Gaussian Splatting when rendering resolutions change by enforcing a maximum per-Gaussian sampling frequency $\hat{\nu}_k$ and introducing a 4D scale-adaptive filter that uses a scale-aware dilation factor $\rho_{adapt}$. A scale regularization term $\mathcal{L}_{\text{scales}}$ complements the filter to constrain Gaussian scales during deformation, while a 2D Mip filter mitigates residual scale mismatch. Integrated into state-of-the-art monocular (D3DGS) and multi-view (4DGaussians) reconstruction pipelines, the method reduces high-frequency artifacts, suppresses redundant Gaussians, and preserves reconstruction quality across a range of rendering frequencies. Empirical results on D-NeRF and N3DV datasets demonstrate improved anti-aliasing and more faithful dynamic reconstructions, highlighting practical impact for real-time, high-fidelity dynamic scene rendering.

Abstract

Existing dynamic scene reconstruction methods based on Gaussian Splatting enable real-time rendering and generate realistic images. However, adjusting the camera's focal length or the distance between Gaussian primitives and the camera to modify rendering resolution often introduces strong artifacts, stemming from the frequency constraints of 4D Gaussians and Gaussian scale mismatch induced by the 2D dilated filter. To address this, we derive a maximum sampling frequency formulation for 4D Gaussian Splatting and introduce a 4D scale-adaptive filter and scale loss, which flexibly regulates the sampling frequency of 4D Gaussian Splatting. Our approach eliminates high-frequency artifacts under increased rendering frequencies while effectively reducing redundant Gaussians in multi-view video reconstruction. We validate the proposed method through monocular and multi-view video reconstruction experiments.Ours project page: https://4d-alias-free.github.io/4D-Alias-free/

Figures (8)

• Figure 1: Similar to 3DGS 3dgs, 4DGS undergoes substantial dilation when the sampling rate is lowered and experiences erosion along with high-frequency artifacts when the sampling rate is increased.This occurs because changes in resolution alter the effective pixel size while the Gaussian scale remains fixed, causing a mismatch between the Gaussian scale and the filter dilation scale, compounded by the absence of a maximum sampling frequency constraint in 4DGS.
• Figure 2: We trained all models on single-scale images (full resolution) and rendered images at different resolutions by adjusting the focal length. When the sampling rate changes, 4D Gaussians4dGaussians exhibit strong artifacts, which our method effectively eliminates.
• Figure 3: The fixed dilation scale used by the 3D smoothing filter proposed in Mip-Splatting mip-splatting can inadvertently render imperceptible Gaussians with extremely small scales visible in 4DGS, and may alter the scale ratios among Gaussian dimensions, affecting Gaussian anisotropy. We present Alias-free-4DGS, which leverages a 4D Scale-adaptive Filter and Scale Regularization to jointly constrain the maximum sampling frequency of 4DGS. This reduces the filter’s minimum dilation scale and avoids significant filtering errors when Gaussian scales are much smaller than the filter’s dilation scale. Moreover, our 4D Scale-adaptive Filter masks out imperceptible Gaussians and adapts the dilation scale using the ratio of Gaussian scales before and after changes in the current time frame, thus mitigating the filter’s impact on Gaussian anisotropy. Alias-free-4DGS effectively eliminates artifacts that arise from varying sampling frequencies without compromising reconstruction quality.
• Figure 4: Single-scale Training and Multi-scale Testing on the D-NeRF dataset dnerf.All methods are trained at full resolution and evaluated at various lower resolutions to simulate zoom-out effects. D3DGS d3dgs exhibits noticeable blurring and inflation artifacts at lower resolutions. Integrating Mip-splatting mip-splatting mitigates inflation but introduces local reconstruction distortions, such as the deformation observed in the hand region (first row). In contrast, our method preserves the reconstruction quality of D3DGS while maintaining more realistic visual fidelity at lower resolutions.
• Figure 5: Single-Scale Training and Multi-Scale Testing on the N3DV Datasetn3dv. All models are trained on images downsampled by a factor of four and rendered at full resolution to simulate zoom-in and moving-closer effects. Our method effectively eliminates high-frequency artifacts and produces more complete object shapes compared to Mip-Splatting.