AA-Splat: Anti-Aliased Feed-forward Gaussian Splatting

Taewoo Suh; Sungpyo Kim; Jongmin Park; Munchurl Kim

cs.CVMar 31, 2026

AA-Splat: Anti-Aliased Feed-forward Gaussian Splatting

Authors: Taewoo Suh, Sungpyo Kim, Jongmin Park, Munchurl Kim

Organizations: Korea Advanced Institute of Science and Technology

Abstract

Feed-forward 3D Gaussian Splatting (FF-3DGS) emerges as a fast and robust solution for sparse-view 3D reconstruction and novel view synthesis (NVS). However, existing FF-3DGS methods are built on incorrect screen-space dilation filters, causing severe rendering artifacts when rendering at out-of-distribution sampling rates. We firstly propose an FF-3DGS model, called AA-Splat, to enable robust anti-aliased rendering at any resolution. AA-Splat utilizes an opacity-balanced band-limiting (OBBL) design, which combines two components: a 3D band-limiting post-filter integrates multi-view maximal frequency bounds into the feed-forward reconstruction pipeline, effectively band-limiting the resulting 3D scene representations and eliminating degenerate Gaussians; an Opacity Balancing (OB) to seamlessly integrate all pixel-aligned Gaussian primitives into the rendering process, compensating for the increased overlap between expanded Gaussian primitives. AA-Splat demonstrates drastic improvements with average 5.4 $\sim$ 7.5dB PSNR gains on NVS performance over a state-of-the-art (SOTA) baseline, DepthSplat, at all resolutions, between $4\times$ and $1/4\times$ . Code will be made available.

Figures & tables

Figure 2: Overview of AA-Splat. After unprojecting per-view depth maps to obtain the 3D Gaussian centers, the maximal frequency bound of each 3D Gaussian primitive is computed from context view camera poses and focal lengths, which is then converted to its 3D band-limiting Gaussian post-filter. The GS head predicts the residual Gaussian parameters which are then post-filtered by the 3D filters to obtain the band-limited 3D Gaussians. During rendering, opacity balancing (OB) compensates for the increase in overlap between pixel-aligned Gaussian primitives for improved photometric supervision and rendering fidelity.

Figure 3: Multi-scale visual comparison on RE10K [ 44 ] . $\times 4$ indicates rendering target views at higher resolution to simulate zoom-in effects, while $\times 1/4$ indicates rendering target views at lower resolution to simulate zoom-out effects. Additional visualizations will be available in supplementary materials.

Figure 4: Multi-scale visual comparison of cross-dataset generalization on ACID [ 24 ] . Models trained on RE10K are directly used to test on scenes from ACID. Additional visualizations will be available in supplementary materials.

Method	PSNR (dB) $\uparrow$						SSIM $\uparrow$						LPIPS $\downarrow$
	$1/4\times$ Res.	$1/2\times$ Res.	$1\times$ Res.	$2\times$ Res.	$4\times$ Res.	Avg.	$1/4\times$ Res.	$1/2\times$ Res.	$1\times$ Res.	$2\times$ Res.	$4\times$ Res.	Avg.	$1/4\times$ Res.	$1/2\times$ Res.	$1\times$ Res.	$2\times$ Res.	$4\times$ Res.	Avg.
MVSplat [ 7 ]	20.252	22.428	26.387	20.294	17.882	21.449	0.735	0.847	0.869	0.808	0.753	0.803	0.152	0.129	0.128	0.279	0.421	0.220
DepthSplat [ 33 ]	21.300	24.449	26.791	20.595	17.399	22.107	0.740	0.864	0.878	0.779	0.671	0.787	0.139	0.102	0.122	0.319	0.449	0.226
NoPoSplat [ 38 ]	21.367	24.786	26.755	21.808	18.809	22.705	0.739	0.871	0.878	0.821	0.755	0.813	0.126	0.095	0.126	0.266	0.362	0.195
SPFSplat [ 18 ]	21.283	24.694	27.313	21.929	18.922	22.828	0.740	0.873	0.888	0.827	0.765	0.819	0.127	0.093	0.119	0.247	0.353	0.188
SPFSplatV2 [ 19 ]	21.329	24.781	27.519	22.305	19.444	23.076	0.741	0.876	0.892	0.837	0.785	0.826	0.125	0.091	0.116	0.237	0.340	0.182

Table 1: Quantitative comparison of single-scale reconstruction and multi-scale rendering on the RE10K [ 44 ] dataset under various target resolutions. We evaluate all models by performing a single feed-forward inference under

256\times 256

input resolution and rendering the resulting 3D Gaussians at

1/4\times

1/2\times

1\times

2\times

, and

4\times

resolutions. Due to the absence of high-resolution ground truths in the RE10K dataset, we instead report the HRRC metrics which are computed against bicubic upsampled ground truth frames, for

2\times

and

4\times

resolutions. Red and Blue indicate the best and second-best performances, respectively.

Method	PSNR (dB) $\uparrow$						SSIM $\uparrow$						LPIPS $\downarrow$
	$1/4\times$ Res.	$1/2\times$ Res.	$1\times$ Res.	$2\times$ Res.	$4\times$ Res.	Avg.	$1/4\times$ Res.	$1/2\times$ Res.	$1\times$ Res.	$2\times$ Res.	$4\times$ Res.	Avg.	$1/4\times$ Res.	$1/2\times$ Res.	$1\times$ Res.	$2\times$ Res.	$4\times$ Res.	Avg.
DepthSplat [ 33 ]	19.687	22.660	27.109	17.294	13.357	20.022	0.679	0.848	0.885	0.647	0.419	0.695	0.173	0.108	0.104	0.399	0.566	0.270
Ours	28.183	29.012	27.882	25.701	24.351	27.026	0.934	0.927	0.892	0.821	0.762	0.867	0.061	0.067	0.096	0.233	0.365	0.164

Table 2: Zero-shot generalization from RE10K [ 44 ] to DL3DV [ 23 ] dataset under various target resolutions. We evaluate all models by performing a single feed-forward inference with 2 context views of

256\times 448

input resolution and rendering the resulting 3D Gaussians at

1/4\times

1/2\times

1\times

2\times

, and

4\times

resolutions.

Method	PSNR (dB) $\uparrow$						SSIM $\uparrow$						LPIPS $\downarrow$
	$1/4\times$ Res.	$1/2\times$ Res.	$1\times$ Res.	$2\times$ Res.	$4\times$ Res.	Avg.	$1/4\times$ Res.	$1/2\times$ Res.	$1\times$ Res.	$2\times$ Res.	$4\times$ Res.	Avg.	$1/4\times$ Res.	$1/2\times$ Res.	$1\times$ Res.	$2\times$ Res.	$4\times$ Res.	Avg.
MVSplat [ 7 ]	20.899	22.232	28.160	19.226	16.422	21.388	0.730	0.829	0.841	0.745	0.648	0.759	0.154	0.141	0.147	0.349	0.507	0.259
DepthSplat [ 33 ]	22.264	24.512	28.026	18.019	14.078	21.380	0.701	0.823	0.841	0.622	0.421	0.682	0.145	0.126	0.147	0.486	0.624	0.306
NoPoSplat [ 38 ]	22.843	25.184	27.452	21.188	18.060	22.945	0.723	0.832	0.824	0.755	0.674	0.761	0.130	0.120	0.161	0.345	0.452	0.242
SPFSplat [ 18 ]	22.532	24.814	28.077	21.217	18.155	22.959	0.725	0.842	0.841	0.765	0.688	0.772	0.129	0.115	0.131	0.311	0.435	0.228
SPFSplatV2 [ 19 ]	22.734	25.159	28.322	21.714	18.822	23.350	0.726	0.845	0.846	0.781	0.721	0.784	0.125	0.111	0.145	0.296	0.418	0.219

Table 3: Zero-shot generalization from RE10K [ 44 ] to ACID [ 24 ] dataset under various target resolutions. We evaluate all models by performing a single feed-forward inference under

256\times 256

input resolution and rendering the resulting 3D Gaussians at

1/4\times

1/2\times

1\times

2\times

, and

4\times

resolutions. Due to the absence of high-resolution ground truths in the ACID dataset, we instead report the HRRC metrics, which are computed against bicubic upsampled ground truth frames, for

2\times

and

4\times

resolutions. Red and Blue indicate the best and second-best performances, respectively.

Figure 5: Ablation results on the DL3DV [ 23 ] dataset. All methods are trained and inferred at full resolution and rendered at $4\times$ resolution to mimic zoom-in. Removing either OB or 3D-BLPF causes unnatural grid-like artifacts or sharp jagged edges. Our method (AA-Splat) effectively suppresses these rendering artifacts.

Variants	PSNR (dB) $\uparrow$						SSIM $\uparrow$						LPIPS $\downarrow$
	$1/4$ Res.	$1/2$ Res.	$1\times$ Res.	$2\times$ Res.	$4\times$ Res.	Avg.	$1/4$ Res.	$1/2$ Res.	$1\times$ Res.	$2\times$ Res.	$4\times$ Res.	Avg.	$1/4$ Res.	$1/2$ Res.	$1\times$ Res.	$2\times$ Res.	$4\times$ Res.	Avg.
(a) Baseline [ 33 ]	19.687	22.660	27.109	17.294	13.357	20.022	0.679	0.848	0.885	0.647	0.419	0.695	0.173	0.108	0.104	0.399	0.566	0.270
(b) w/o OB	28.140	27.936	26.623	24.868	23.715	26.257	0.934	0.915	0.875	0.806	0.752	0.856	0.057	0.076	0.113	0.248	0.375	0.174
(c) w/o 3D-BLPF	28.013	28.639	27.322	25.186	23.771	26.586	0.932	0.924	0.885	0.806	0.738	0.857	0.062	0.069	0.112	0.273	0.423	0.188
(d) Ours	28.183	29.012	27.882	25.701	24.351	27.026	0.934	0.927	0.892	0.821	0.762	0.867	0.061	0.067	0.096	0.233	0.365	0.164

Table 4: Ablation results on the DL3DV [ 23 ] dataset. We evaluate all variants by performing a single feed-forward inference under

256\times 448

input resolution and rendering the resulting 3D Gaussians at

1/4\times

1/2\times

1\times

2\times

, and

4\times

resolutions. Red and Blue indicate the best and second-best performances, respectively.

Method	PSNR (dB) $\uparrow$						SSIM $\uparrow$						LPIPS $\downarrow$
	$1/4\times$ Res.	$1/2\times$ Res.	$1\times$ Res.	$2\times$ Res.	$4\times$ Res.	Avg.	$1/4\times$ Res.	$1/2\times$ Res.	$1\times$ Res.	$2\times$ Res.	$4\times$ Res.	Avg.	$1/4\times$ Res.	$1/2\times$ Res.	$1\times$ Res.	$2\times$ Res.	$4\times$ Res.	Avg.
MVSplat [ 7 ]	18.690	19.771	21.834	18.173	16.140	18.922	0.635	0.713	0.711	0.630	0.582	0.654	0.201	0.181	0.206	0.349	0.485	0.284
DepthSplat [ 33 ]	19.687	22.660	27.109	17.294	13.357	20.022	0.679	0.848	0.885	0.647	0.419	0.695	0.173	0.108	0.104	0.399	0.566	0.270
Ours	28.183	29.012	27.882	25.701	24.351	27.026	0.934	0.927	0.892	0.821	0.762	0.867	0.061	0.067	0.096	0.233	0.365	0.164

Table 5: Zero-shot generalization from RE10K [ 44 ] to DL3DV [ 23 ] dataset under various target resolutions. We evaluate all models by performing a single feed-forward inference with 2 context views of

256\times 448

input resolution and rendering the resulting 3D Gaussians at

1/4\times

1/2\times

1\times

2\times

, and

4\times

resolutions.

Figure 6: Ablation results on the DL3DV [ 23 ] dataset. All variants perform feed-forward reconstruction with 2 context views of $256\times 448$ resolution and are rendered at $4\times$ resolution to mimic zoom-in effects. Removing either OB or 3D-BLPF causes unnatural grid-like artifacts or sharp jagged edges. Using only the 2D Mip filter also causes grid-like artifacts. Our method (AA-Splat) effectively suppresses these rendering artifacts, enabling smooth interpolation. Best viewed zoomed-in.

Variants	2D-	3D-		PSNR (dB) $\uparrow$						SSIM $\uparrow$						LPIPS $\downarrow$
	Mip	BLPF	OB	$1/4\times$	$1/2\times$	$1\times$	$2\times$	$4\times$	Avg.	$1/4\times$	$1/2\times$	$1\times$	$2\times$	$4\times$	Avg.	$1/4\times$	$1/2\times$	$1\times$	$2\times$	$4\times$	Avg.
(a) Baseline [ 33 ]	✗	✗	✗	19.687	22.660	27.109	17.294	13.357	20.022	0.679	0.848	0.885	0.647	0.419	0.695	0.173	0.108	0.104	0.399	0.566	0.270
(b) w/o OB	✓	✓	✗	28.140	27.936	26.623	24.868	23.715	26.257	0.934	0.915	0.875	0.806	0.752	0.856	0.057	0.076	0.113	0.248	0.375	0.174
(c) w/o 3D-BLPF	✓	✗	✓	28.013	28.639	27.322	25.186	23.771	26.586	0.932	0.924	0.885	0.806	0.738	0.857	0.062	0.069	0.112	0.273	0.423	0.188
(d) 2D Mip only	✓	✗	✗	28.072	28.066	27.002	25.116	23.886	26.429	0.934	0.916	0.879	0.808	0.749	0.857	0.058	0.075	0.110	0.241	0.372	0.171
(e) 3D-BLPF only	✗	✓	✗	19.812	22.698	26.967	21.554	19.162	22.038	0.708	0.861	0.884	0.795	0.717	0.793	0.160	0.102	0.106	0.256	0.400	0.205

Table 6: Additional ablation results on the DL3DV [ 23 ] dataset. We evaluate all variants by performing a single feed-forward inference under

256\times 448

input resolution and rendering the resulting 3D Gaussians at

1/4\times

1/2\times

1\times

2\times

, and

4\times

resolutions. Red and Blue indicate the best and second-best performances, respectively.

Figure 7: Additional multi-scale visual comparisons on RE10K [ 44 ] . All methods are trained and inferred at $256\times 256$ resolution and rendered at $1/4\times\sim 4\times$ resolutions to mimic zoom-in/out effects. Best viewed zoomed-in.

Figure 8: Additional multi-scale visual comparisons of cross-dataset generalization to ACID [ 24 ] . Models trained on RE10K are directly tested on scenes from ACID. When rendering novel views at decreased sampling rates on the ACID dataset where thin foreground structures are relatively rare, aliasing artifacts in existing models primarily manifest as brightening artifacts, where scenes are rendered as being unnaturally bright.

Explore similar work

ATSplat: Compact Feed-forward 3D Gaussian Splatting with Adaptive Token Expansion
Jul 22, 2026In Cho, Jeonghwan Cho, Mijin Yoo +23D Gaussian Splatting Adaptive Gaussian Densification
Yonsei University, South Korea · National University of Singapore, Singapore
SparseSplat: Towards Applicable Feed-Forward 3D Gaussian Splatting with Pixel-Unaligned Prediction
Apr 3, 2026Zicheng Zhang, Xiangting Meng, Ke Wu +13D Gaussian Splatting Feed-Forward 3D Reconstruction
Fudan University · ShanghaiTech University
RoSplat: Robust Feed-Forward Pixel-wise Gaussian Splatting for Varying Input Views and High-Resolution Rendering
May 13, 2026Hoang Chuong Nguyen, Renjie Wu, Jose M. Alvarez +13D Gaussian Splatting Novel View Synthesis
Australian National University · NVIDIA

Jul 22, 2026J/K move · Enter open · S save

ATSplat: Compact Feed-forward 3D Gaussian Splatting with Adaptive Token Expansion

In Cho, Jeonghwan Cho, Mijin Yoo, Gim Hee Lee+1

Yonsei University, South Korea · National University of Singapore, Singapore

3D Gaussian Splatting Adaptive Gaussian Densification Novel View Synthesis Feed-Forward 3D Reconstruction

Open paper AskPDF arXiv

3D Gaussian Splatting (3DGS) achieves high-quality novel-view synthesis by optimizing freely placed primitives in 3D and adaptively densifying them in under-reconstructed regions. However, this scene-adaptive capacity allocation is largely lost in existing feed-forward 3DGS methods, which commonly regress Gaussians at input pixels and lift them along camera rays. Such pixel-aligned formulations make the number and placement of primitives depend on image resolution and input viewpoints rather than scene complexity, resulting in dense and often redundant Gaussian sets. We present ATSplat, a feed-forward 3DGS framework that restores the adaptive allocation capability of 3DGS optimization through Adaptive 3D Tokens. ATSplat first lifts coarse patch-level depth and camera cues into sparse 3D anchor tokens, forming a compact scaffold of the scene. Each token is then regressed into local Gaussians with learnable 3D offsets, decoupling primitive placement from input image grids. An Adaptive Token Expansion module predicts a token-level uncertainty score, supervised by rendering error maps, and selectively expands high-uncertainty tokens through learnable expansion layers. This sparse-to-adaptive formulation enables ATSplat to concentrate primitives in challenging regions while maintaining a compact representation. Experiments on two representative datasets, RealEstate10K and DL3DV, show that ATSplat achieves state-of-the-art rendering quality while reducing the number of Gaussians by more than

5.7\times

compared with dense feed-forward 3DGS methods. From 12 input images at

512 \times 960

resolution, ATSplat completes reconstruction in less than a second using a single commercial GPU, and renders high-quality novel views at 1136 FPS (

512 \times 960

) with only 311K Gaussians.

Abstract

Explore similar work

ATSplat: Compact Feed-forward 3D Gaussian Splatting with Adaptive Token Expansion

SparseSplat: Towards Applicable Feed-Forward 3D Gaussian Splatting with Pixel-Unaligned Prediction

RoSplat: Robust Feed-Forward Pixel-wise Gaussian Splatting for Varying Input Views and High-Resolution Rendering