AA-Splat: FF-3D Gaussian Splatting Model

• AA-Splat is a feed-forward 3D Gaussian splatting model that introduces opacity-balanced band-limiting to address aliasing artifacts in sparse-view novel view synthesis.
• It combines a 3D band-limiting post-filter with opacity balancing to prevent degenerate Gaussian predictions and achieve PSNR gains of 5–8 dB over traditional methods.
• The method overcomes the limitations of screen-space dilation filters, enabling robust anti-aliasing in FF-3DGS pipelines like DepthSplat for consistent rendering quality.

Searching arXiv for AA-Splat and closely related Gaussian splatting papers to ground the article in current literature. AA-Splat is a feed-forward 3D Gaussian Splatting (FF-3DGS) model for sparse-view 3D reconstruction and novel view synthesis that targets robust anti-aliased rendering at any resolution. It is motivated by the observation that existing FF-3DGS methods are built on incorrect screen-space dilation filters, which cause severe rendering artifacts when rendering at out-of-distribution sampling rates. The method introduces an opacity-balanced band-limiting (OBBL) design that combines a 3D band-limiting post-filter with Opacity Balancing, and it is demonstrated on a DepthSplat backbone with average PSNR gains of $5.4\sim7.5$ dB over the baseline between $4\times$ and $1/4\times$ resolutions (Suh et al., 31 Mar 2026).

1. Research setting and definitional scope

AA-Splat belongs to the FF-3DGS regime, in which a single network predicts a scene’s 3D Gaussian primitives from sparse views in one forward pass. The paper explicitly situates the method alongside FF-3DGS systems such as DepthSplat, MVSplat, NoPoSplat, and SPFSplat(V2), and frames anti-aliasing as a central unresolved problem in that regime (Suh et al., 31 Mar 2026).

This positioning is important because anti-aliasing in Gaussian splatting had already been studied in a different setting by "Analytic-Splatting: Anti-Aliased 3D Gaussian Splatting via Analytic Integration" (Liang et al., 2024). That earlier method addresses aliasing in per-scene 3DGS by analytically approximating the Gaussian integral within the 2D-pixel window area and incorporating the approximated response into transmittance calculation. AA-Splat instead addresses feed-forward reconstruction pipelines and their instability under out-of-distribution sampling rates. A plausible implication is that the two works occupy different points in the Gaussian-splatting design space rather than representing minor variants of the same procedure.

2. Failure mode addressed by AA-Splat

The core diagnosis is that existing FF-3DGS methods rely on incorrect screen-space dilation filters. During training, the learned Gaussians are dilated in screen space by an ad-hoc filter to cover at least one pixel. This can preserve rendering quality near the training sampling rates, but it does not generalize when focal length, distance, or resolution changes (Suh et al., 31 Mar 2026).

The resulting artifacts differ by sampling regime. Under zoom-in, tiny or degenerate Gaussians open gaps, producing erosion and seam-like black regions. Under zoom-out, expanded Gaussians overlap excessively, producing dilation artifacts and over-thickened structures. The paper further notes that simply replacing the dilation rule with the 2D MIP filter from Mip-Splatting reduces one kind of artifact but does not prevent the network from regressing overly thin Gaussians, so aliasing remains severe at out-of-distribution sampling rates such as $\pm4\times$ resolution or more (Suh et al., 31 Mar 2026).

A common misconception is that aliasing in FF-3DGS can be solved entirely in screen space. AA-Splat’s formulation rejects that premise: the method band-limits the 3D representation itself and then compensates for the overlap induced by that regularization.

3. Opacity-balanced band-limiting (OBBL)

AA-Splat’s central mechanism is the OBBL framework, which consists of two coupled modules: a 3D band-limiting post-filter (3D-BLPF) and Opacity Balancing (OB) (Suh et al., 31 Mar 2026).

For the $j$-th Gaussian, the method estimates the maximal screen-space sampling frequency across context views by projecting its world-space center $\mu_j$ into each context view and combining visibility, focal length, and depth: $\hat\nu_j=\max_{i=1\ldots N}\left[\mathds{1}^{(i)}(\mu_j)\frac{f^{(i)}}{d_j^{(i)}}\right].$ Here, $\mathds{1}^{(i)}(\mu_j)\in\{0,1\}$ indicates whether Gaussian $j$ is visible in view $i$.

AA-Splat then defines an isotropic 3D Gaussian low-pass filter with variance

$4\times$0

Convolving the predicted Gaussian with this low-pass filter yields a regularized scale update

$4\times$1

This makes the resulting Gaussian unable to shrink below the Nyquist scale implied by the highest sampling rate at which that primitive would be observed in any context view. In the paper’s terms, the network is thereby forced to regress “well-behaved” primitives and degenerate Gaussians are eliminated (Suh et al., 31 Mar 2026).

The significance of this design is not merely smoothing. The regularization is view-aware through $4\times$2, and it is imposed in 3D rather than as a purely 2D raster-space heuristic. This suggests that AA-Splat treats anti-aliasing as a constraint on representational bandwidth, not only as a post hoc rendering correction.

4. Opacity Balancing and rendering formulation

Band-limiting alone increases mutual overlap among Gaussian primitives. AA-Splat therefore introduces Opacity Balancing to preserve rendering fidelity and photometric supervision when expanded Gaussians interact (Suh et al., 31 Mar 2026).

The first step is opacity clamping: $4\times$3 This prevents a single primitive from becoming fully opaque and hiding all primitives behind it.

The second step is color normalization. Using the band-limited Gaussians, the renderer rasterizes alpha and color in the usual back-to-front manner: $4\times$4

$4\times$5

If the accumulated opacity satisfies $4\times$6, with $4\times$7 at test time, the final color is

$4\times$8

Within the full FF-3DGS pipeline, the paper describes four stages: depth and feature extraction from context views $4\times$9; a Gaussian head that unprojects each pixel to a world-space center $1/4\times$0 and predicts $1/4\times$1 as residuals; 3D band-limiting using $1/4\times$2 and $1/4\times$3; and rendering with OB while applying a 2D MIP filter in place of vanilla dilation (Suh et al., 31 Mar 2026). The method is demonstrated on DepthSplat, but the paper presents it as a design that can be built on any FF-3DGS backbone.

5. Quantitative performance

AA-Splat is evaluated by rendering the same predicted Gaussians at resolutions ranging from $1/4\times$4 to $1/4\times$5, corresponding to zoom-out through zoom-in. The reported baseline is DepthSplat (Suh et al., 31 Mar 2026).

Dataset	Baseline avg PSNR	AA-Splat avg PSNR
RE10K	22.11 dB	27.51 dB
DL3DV	20.02 dB	27.03 dB
ACID (HRRC)	21.38 dB	28.86 dB

On RE10K, the average improvement is $1/4\times$6 dB. The paper also reports scale-specific RE10K changes: zoom-out $1/4\times$7 improves from $1/4\times$8 to $1/4\times$9 dB, native $\pm4\times$0 from $\pm4\times$1 to $\pm4\times$2 dB, and zoom-in $\pm4\times$3 from $\pm4\times$4 to $\pm4\times$5 dB. On DL3DV the average gain is $\pm4\times$6 dB, and on ACID the average gain is $\pm4\times$7 dB (Suh et al., 31 Mar 2026).

The paper states that SSIM and LPIPS likewise improve dramatically, giving as an example SSIM increasing from approximately $\pm4\times$8 to above $\pm4\times$9 under zoom-in and LPIPS being roughly halved. Qualitatively, AA-Splat removes the black seams associated with erosion and the puffy dilation artifacts associated with overlap at both zoom-in and zoom-out. In the paper’s summary, the result is alias-free novel-view synthesis over a $j$0 range of sampling rates with consistent $j$1–$j$2 dB PSNR gains over the previous state of the art (Suh et al., 31 Mar 2026).

6. Relation to Analytic-Splatting, misconceptions, and limitations

AA-Splat is distinct from "Analytic-Splatting" (Liang et al., 2024), despite the close thematic overlap in anti-aliasing for Gaussian splatting. Analytic-Splatting starts from the observation that standard 3DGS shades each pixel by sampling the projected 2D Gaussian at the pixel center,

$j$3

and attributes aliasing to point-sample rendering that is insensitive to pixel footprint. It therefore approximates the Gaussian CDF with a conditioned logistic function,

$j$4

and uses that approximation to integrate the 2D Gaussian over the pixel window area, making both shading and transmittance sensitive to pixel footprint changes across resolutions (Liang et al., 2024).

AA-Splat does not use analytic pixel-window integration as its central mechanism. Its intervention is instead a Nyquist-driven 3D band-limiting post-filter plus opacity balancing inside an FF-3DGS pipeline (Suh et al., 31 Mar 2026). This suggests that the two methods are complementary responses to aliasing: Analytic-Splatting addresses the discrete point-sampling assumption in per-scene 3DGS, whereas AA-Splat addresses degenerate Gaussian prediction and overlap management in feed-forward reconstruction.

The limitations reported for AA-Splat are specific. The method introduces additional hyperparameters $j$5, which are fixed by cross-validation. The demonstrations are limited to static scenes; extension to dynamic or deformable scenes would require time-aware band-limiting kernels, for example along trajectories. The 3D convolution is implemented as a closed-form variance update, but non-Gaussian or anisotropic filters for highly non-stationary sampling rates remain open directions. The paper also identifies integration into other FF-3DGS architectures, including pose-free models, and into faster real-time renderers as future work (Suh et al., 31 Mar 2026).