AnchorSplat: Feed-Forward 3D Gaussian Splatting with 3D Geometric Priors

Xiaoxue Zhang; Xiaoxu Zheng; Yixuan Yin; Tiao Zhao; Kaihua Tang; Michael Bi Mi; Zhan Xu; Dave Zhenyu Chen

cs.CVApr 8, 2026

AnchorSplat: Feed-Forward 3D Gaussian Splatting with 3D Geometric Priors

Authors: Xiaoxue Zhang, Xiaoxu Zheng, Yixuan Yin, Tiao Zhao, Kaihua Tang, Michael Bi Mi, Zhan Xu, Dave Zhenyu Chen

Abstract

Recent feed-forward Gaussian reconstruction models adopt a pixel-aligned formulation that maps each 2D pixel to a 3D Gaussian, entangling Gaussian representations tightly with the input images. In this paper, we propose AnchorSplat, a novel feed-forward 3DGS framework for scene-level reconstruction that represents the scene directly in 3D space. AnchorSplat introduces an anchor-aligned Gaussian representation guided by 3D geometric priors (e.g., sparse point clouds, voxels, or RGB-D point clouds), enabling a more geometry-aware renderable 3D Gaussians that is independent of image resolution and number of views. This design substantially reduces the number of required Gaussians, improving computational efficiency while enhancing reconstruction fidelity. Beyond the anchor-aligned design, we utilize a Gaussian Refiner to adjust the intermediate Gaussiansy via merely a few forward passes. Experiments on the ScanNet++ v2 NVS benchmark demonstrate the SOTA performance, outperforming previous methods with more view-consistent and substantially fewer Gaussian primitives.

Figures & tables

Figure 1: Novel view synthesis comparison between AnySplat [ 16 ] and our AnchorSplat. AnySplat, as a voxel-aligned feed-forward approach, often suffers from inaccurate geometry, leading to ghosting artifacts, floaters, and blurred structures, especially in regions with sparse or ambiguous depth cues. In contrast, given the same sequence of input images, our anchor-aligned approach performs feed-forward Gaussian reconstruction using nearly 20× fewer Gaussians and about half the reconstruction time, while producing sharper geometry, cleaner surfaces, artifact-free renderings, and more accurate depth estimation.

Figure 2: Comparison of pixel-aligned and anchor-aligned Gaussian representations. Pixel-aligned Gaussians exhibit inconsistent sampling across views, especially under occlusions, low-texture regions, and motion parallax, while anchor-aligned Gaussians provide a more stable and consistent 3D representation.

Figure 3: Overview of the proposed AnchorSplat pipeline. The framework consists of three components: a pretrained Multi-View stereo module that extracts 3D geometric priors and depth maps from posed images, a transformer-based Gaussian decoder that predicts anchor-aligned Gaussians guided by these priors, and a lightweight Gaussian refiner that further refines Gaussian attributes to improve reconstruction quality and view consistency.

Dataset	Model	Category	RGB			Depth		NumGS	ReconTime(s)
Dataset	Model	Category	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	$\delta_{1}\uparrow$	AbsRel $\downarrow$	NumGS	ReconTime(s)
ScanNet++	3DGS	Opt.	19.98	0.72	0.30	0.31	0.42	496,087	391.44
	Mip-Splatting	Opt.	19.92	0.75	0.34	0.35	0.38	398,212	289.95
	AnySplat	feed-forward	20.20	0.73	0.32	0.71	0.16	5,550,940	6.83
	AnchorSplat ⋆	feed-forward	20.96	0.78	0.47	0.94	0.068	247,153	3.11
	AnchorSplat	feed-forward	21.48	0.79	0.38	0.94	0.066	247,153	5.52

Table 1: Novel 4-view evaluation on ScanNet++ with 32 input views. AnchorSplat achieves competitive or superior quality with nearly 20× fewer Gaussians and lower reconstruction time compared with 3DGS, Mip-Splatting, and AnySplat.

Figure 4: Comparison of reconstructed 3D Gaussians. Compared with AnySplat, our AnchorSplat produces cleaner and more geometry-consistent structures with far fewer Gaussians, reducing floaters, ghost artifacts, and irregular density distributions.

Exp setting	Model	RGB			Depth		NumGS	ReconTime(s)
	Model	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	$\delta_{1}$ $\uparrow$	AbsRel $\downarrow$	NumGS	ReconTime(s)
32 sv / 4 nv	AnySplat	20.20	0.73	0.32	0.71	0.16	5,550,940	6.83
	AnchorSplat ⋆	20.96	0.78	0.47	0.94	0.068	247,153	3.11
	AnchorSplat	21.48	0.79	0.38	0.94	0.066	247,153	5.52
48 sv / 6 nv	AnySplat	20.66	0.74	0.31	0.67	0.17	8,197,441	7.57
	AnchorSplat ⋆	20.80	0.78	0.48	0.94	0.064	247,153	3.23

Table 2: Evaluation on ScanNet++ V2 with 32/48/64 input views and 4/6/8 novel views, respectively. More input views improve reconstruction quality, while our anchor-aligned representation remains efficient with a fixed number of Gaussians.

Figure 5: Reconstruction quality and runtime comparison between AnySplat and AnchorSplat across input-view settings. Dashed lines denote runtime, solid lines denote novel-view PSNR, and marker size with annotated numbers indicates the number of Gaussians (NumGS). AnchorSplat achieves better quality with lower runtime and fewer Gaussians.

Figure 6: Visual comparison before and after applying the Gaussian Refiner. The refiner improves boundary sharpness, fills missing regions, and enhances color consistency, leading to better overall rendering quality.

Exp setting	Method	RGB			Depth		NumGS	ReconTime(s)
sv/nv	Method	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	$\delta_{1}\uparrow$	AbsRel $\downarrow$	NumGS	ReconTime(s)
3 sv / 1 nv	AnySplat	19.51	0.68	0.38	0.70	0.18	543,987	1.34
3 sv / 1 nv	AnchorSplat ⋆	19.99	0.78	0.38	0.92	0.073	247,153	3.18
5 sv / 1 nv	AnySplat	20.21	0.77	0.36	0.71	0.17	909,132	2.25
5 sv / 1 nv	AnchorSplat ⋆	20.35	0.78	0.38	0.92	0.073	247,153	3.23
128 sv / 16 nv	AnySplat	20.47	0.74	0.34	0.75	0.18	19,767,552	31.67

Table 3: Ablation results under extremely sparse (3/5) and extremely dense (128/256) input-view settings. AnchorSplat maintains stable performance while reconstructing scenes efficiently with far fewer Gaussians.

Figure 7: Comparison visualization. AnchorSplat produces noticeably higher-quality renderings with more accurate geometry and sharper depth, while avoiding the multi-view inconsistencies that often lead to artifacts such as ghosting and floaters. For example, in the first row, the chair region rendered by AnySplat exhibits clear view-dependent artifacts caused by inconsistent multi-view alignment, whereas our method maintains clean geometry and coherent appearance. Similar improvements can be observed across various scenes in the visualization, highlighting the robustness and reliability of our anchor-aligned Gaussian representation.

Appendix figures & tables7 assets

Supplementary material from the paper’s appendix.

Appendix

Setting	RGB			Depth
	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	$\delta_{1}\uparrow$	AbsRel $\downarrow$
RGB	20.53	0.78	0.46	0.92	0.076
RGB+Depth	20.67	0.78	0.48	0.94	0.057
RGB+CamRay	20.34	0.78	0.48	0.92	0.080
RGB+CamRay+Depth	20.96	0.78	0.47	0.94	0.068

Appendix

Table 4: Ablation study on input information for the Gaussian decoder. This experiment presents the evaluation results on the novel 4 views using ScanNet++ V2 with 32 input views. In this table, RGB denotes using only RGB images as input; Depth indicates the use of depth maps; CamRay represents Plücker ray embeddings derived from camera intrinsics and extrinsics.

Setting	RGB			Depth		numGS
Setting	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	$\delta_{1}\uparrow$	AbsRel $\downarrow$	numGS
numGS=1	20.14	0.78	0.50	0.93	0.077	61,788
numGS=2	20.61	0.79	0.48	0.93	0.075	123,577
numGS=4	20.66	0.79	0.48	0.93	0.076	247153
numGS=8	20.46	0.78	0.49	0.93	0.073	494,307
numGS=16	20.79	0.79	0.49	0.94	0.072	988,613

Appendix

Table 5: Ablation on the number of Gaussians per anchor. We evaluate the effect of varying the Gaussian multiplicity in the Gaussian decoder on the ScanNet++ V2 dataset. All results are obtained using the decoder only (w/o Refiner). Using too few Gaussians restricts representational capacity, while very large counts yield diminishing returns with substantially higher computational cost. A moderate setting of 4 Gaussians per anchor achieves the best overall balance between accuracy and efficiency.

Dataset	Method	RGB			Depth		NumGS	ReconTime(s)
Dataset	Method	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	$\delta_{1}\uparrow$	AbsRel $\downarrow$	NumGS	ReconTime(s)
ARKitScenes	AnySplat	21.35	0.75	0.30	0.88	0.12	3,237,113	6.63
ARKitScenes	AnchorSplat ⋆	21.00	0.77	0.41	0.96	0.06	400,000	2.23
Replica	AnySplat	21.19	0.70	0.32	0.86	0.11	5,740,107	4.72
Replica	AnchorSplat ⋆	23.48	0.78	0.31	0.95	0.066	800,000	2.40
T&T	AnySplat	15.30	0.45	0.46	0.44	0.39	3,952,092	6.74

Appendix

Table 6: Comparison of AnchorSplat variants on different datasets (32 sampled views / 4 novel views). Experiments are conducted on indoor datasets (ARkitScenes and Replica) and an outdoor dataset (T&T), using the same number of training steps, learning rate, and backbone settings. AnchorSplat ⋆ denotes AnchorSplat without the Refiner. Performance is evaluated in terms of RGB and depth rendering quality as well as reconstruction efficiency.

Figure 8: PCA visualization of three feature aggregations.

Figure 9: Comparison of reconstructed Gaussians between AnySplat and AnchorSplat

Figure 10: Comparison of rendered RGB images and depth images between AnySplat and AnchorSplat

Backbone	RGB			Depth		NumGS	ReconTime(s)
Backbone	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	$\delta_{1}\uparrow$	AbsRel $\downarrow$	NumGS	ReconTime(s)
MapAnything	22.41	0.79	0.33	0.92	0.084	800,000	1.89
DA3	22.04	0.75	0.35	0.91	0.085	800,000	1.72

Appendix

Table 7: Ablation study of AnchorPredictor backbones on the Replica dataset (32 sampled views / 32 novel views). We compare MapAnything and DA3 as backbones under the same training configurations, including identical number of training steps, learning rate, and network settings. Performance is measured in terms of RGB and depth rendering quality as well as reconstruction efficiency.

Explore similar work

May 21, 2026

ForeSplat: Optimization-Aware Foresight for Feed-Forward 3D Gaussian Splatting

Feed-forward 3D Gaussian Splatting models offer fast single-pass reconstruction,but scaling them to match per-scene optimization quality is fundamentally hindered by the scarcity of large-scale 3D annotations. A practical compromise is predict-then-refine,where post-prediction optimization compensates for the limited capacity of the feed-forward network. However,standard feed-forward 3DGS is trained solely for zero-step rendering error,ignoring whether its output constitutes a good initialization for the downstream optimizer. We present ForeSplat,an optimization-aware training framework that equips feed-forward 3DGS models to produce initializations explicitly designed for rapid,effective refinement. By offloading part of the scene-modeling burden to the optimizer,ForeSplat substantially reduces the capacity pressure on the feed-forward model,making high-quality reconstruction feasible even with compact networks. At its core is MetaGrad,a lightweight multi-anchor meta-gradient training rule that bypasses costly higher-order differentiation through the 3DGS optimizer. MetaGrad unrolls a short inner-loop refinement trajectory,samples anchor states,and back-propagates aggregated first-order gradients to the prediction head as a surrogate optimization-aware signal. This fine-tuning adds no inference cost and enables high-quality reconstruction within seconds after a few refinement steps. We instantiate ForeSplat on diverse backbones,including AnySplat,Pi3X,and a distilled variant tailored for edge deployment. Across all tested architectures,a ForeSplat-trained initialization converges in fewer refinement steps and reaches a higher peak reconstruction quality than its vanilla counterpart,even fully converged. The framework consistently bridges the gap between amortized prediction and per-scene optimization,establishing a practical path toward lightweight,high-fidelity 3D reconstruction.

Yuke Li, Weihang Liu, Cheng Zhang +8

ShanghaiTech University · GGU Technology Co., Ltd · Stereye

3D Gaussian Splatting Meta-Learning 3D Reconstruction Feed-Forward 3D Reconstruction

Ask PDF

Apr 16, 2026

GlobalSplat: Efficient Feed-Forward 3D Gaussian Splatting via Global Scene Tokens

The efficient spatial allocation of primitives serves as the foundation of 3D Gaussian Splatting, as it directly dictates the synergy between representation compactness, reconstruction speed, and rendering fidelity. Previous solutions, whether based on iterative optimization or feed-forward inference, suffer from significant trade-offs between these goals, mainly due to the reliance on local, heuristic-driven allocation strategies that lack global scene awareness. Specifically, current feed-forward methods are largely pixel-aligned or voxel-aligned. By unprojecting pixels into dense, view-aligned primitives, they bake redundancy into the 3D asset. As more input views are added, the representation size increases and global consistency becomes fragile. To this end, we introduce GlobalSplat, a framework built on the principle of align first, decode later. Our approach learns a compact, global, latent scene representation that encodes multi-view input and resolves cross-view correspondences before decoding any explicit 3D geometry. Crucially, this formulation enables compact, globally consistent reconstructions without relying on pretrained pixel-prediction backbones or reusing latent features from dense baselines. Utilizing a coarse-to-fine training curriculum that gradually increases decoded capacity, GlobalSplat natively prevents representation bloat. On RealEstate10K and ACID, our model achieves competitive novel-view synthesis performance while utilizing as few as 16K Gaussians, significantly less than required by dense pipelines, obtaining a light 4MB footprint. Further, GlobalSplat enables significantly faster inference than the baselines, operating under 78 milliseconds in a single forward pass. Project page is available at https://r-itk.github.io/globalsplat/

Roni Itkin, Noam Issachar, Yehonatan Keypur +3

The Hebrew University of Jerusalem · Westlake University

3D Gaussian Splatting 3DGS Acceleration Novel View Synthesis 3D Scene Representation Feed-Forward 3D Reconstruction

Ask PDF

Apr 3, 2026

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

Recent progress in feed-forward 3D Gaussian Splatting (3DGS) has notably improved rendering quality. However, the spatially uniform and highly redundant 3DGS map generated by previous feed-forward 3DGS methods limits their integration into downstream reconstruction tasks. We propose SparseSplat, the first feed-forward 3DGS model that adaptively adjusts Gaussian density according to scene structure and information richness of local regions, yielding highly compact 3DGS maps. To achieve this, we propose entropy-based probabilistic sampling, generating large, sparse Gaussians in textureless areas and assigning small, dense Gaussians to regions with rich information. Additionally, we designed a specialized point cloud network that efficiently encodes local context and decodes it into 3DGS attributes, addressing the receptive field mismatch between the general 3DGS optimization pipeline and feed-forward models. Extensive experimental results demonstrate that SparseSplat can achieve state-of-the-art rendering quality with only 22% of the Gaussians and maintain reasonable rendering quality with only 1.5% of the Gaussians. Project page: https://victkk.github.io/SparseSplat-page/.

Zicheng Zhang, Xiangting Meng, Ke Wu +1

Fudan University · ShanghaiTech University

3D Gaussian Splatting Feed-Forward 3D Reconstruction Adaptive Gaussian Densification

Ask PDF