OpenNovelty.org

SurfSplat: Conquering Feedforward 2D Gaussian Splatting with Surface Continuity Priors

ICLR 2026 Conference SubmissionAnonymous Authors
OpenReview Score: 6.5Download Report PDF
3D Gaussian splatting2DGSfeed forward reconstruction
Abstract:

Reconstructing 3D scenes from sparse images remains a challenging task due to the difficulty of recovering accurate geometry and texture without optimization. Recent approaches leverage generalizable models to generate 3D scenes using 3D Gaussian Splatting (3DGS) primitive. However, they often fail to produce continuous surfaces and instead yield discrete, color-biased point clouds that appear plausible at normal resolution but reveal severe artifacts under close-up views. To address this issue, we present SurfSplat, a feedforward framework based on 2D Gaussian Splatting (2DGS) primitive, which provides stronger anisotropy and higher geometric precision. By incorporating a surface continuity prior and a forced alpha blending strategy, SurfSplat reconstructs coherent geometry together with faithful textures. Furthermore, we introduce High-Resolution Rendering Consistency (HRRC), a new evaluation metric designed to evaluate high-resolution reconstruction quality. Extensive experiments on RealEstate10K, DL3DV, and ScanNet demonstrate that SurfSplat consistently outperforms prior methods on both standard metrics and HRRC, establishing a robust solution for high-fidelity 3D reconstruction from sparse inputs.

Disclaimer
This report is AI-GENERATED using Large Language Models and WisPaper (A scholar search engine). It analyzes academic papers' tasks and contributions against retrieved prior work. While this system identifies POTENTIAL overlaps and novel directions, ITS COVERAGE IS NOT EXHAUSTIVE AND JUDGMENTS ARE APPROXIMATE. These results are intended to assist human reviewers and SHOULD NOT be relied upon as a definitive verdict on novelty.
NOTE that some papers exist in multiple, slightly different versions (e.g., with different titles or URLs). The system may retrieve several versions of the same underlying work. The current automated pipeline does not reliably align or distinguish these cases, so human reviewers will need to disambiguate them manually.
If you have any questions, please contact: mingzhang23@m.fudan.edu.cn

Overview

Overall Novelty Assessment

SurfSplat introduces a feedforward framework using 2D Gaussian primitives with surface continuity priors and forced alpha blending to reconstruct coherent geometry and textures from sparse views. The paper resides in the '2D Gaussian Splatting Approaches' leaf, which currently contains only this single work within the broader Gaussian Splatting-Based Feedforward Reconstruction branch. This positioning suggests the paper occupies a relatively sparse research direction, distinguishing itself from the more populated 3D Gaussian methods that dominate sibling categories like Cost Volume-Guided and Transformer-Based approaches.

The taxonomy reveals substantial activity in neighboring 3D Gaussian Splatting categories, with multiple papers exploring cost volumes, transformers, and generative priors. SurfSplat diverges by prioritizing anisotropic 2D primitives over isotropic 3D Gaussians, aiming for stronger surface alignment and geometric precision. This choice connects conceptually to Pointmap and Depth-Based methods that emphasize explicit geometry, yet differs by retaining the splatting rendering paradigm. The exclude_note clarifies that 2D Gaussian approaches focus on surface continuity, separating them from 3D methods that may produce discrete point clouds.

Among the three contributions analyzed, the core SurfSplat framework examined six candidates with zero refutations, the HRRC metric examined ten candidates with zero refutations, and the performance claims examined ten candidates with zero refutations. The limited search scope (26 total candidates examined) means these statistics reflect top-K semantic matches rather than exhaustive coverage. The absence of refutable prior work across all contributions suggests that within this bounded search, the specific combination of 2D Gaussians, surface continuity priors, and high-resolution evaluation appears relatively unexplored, though the small candidate pool limits definitive conclusions.

Based on the top-26 semantic matches examined, SurfSplat appears to introduce a distinct approach within the Gaussian splatting paradigm, occupying a sparsely populated taxonomy leaf. The analysis does not cover the full breadth of related work in implicit representations or alternative geometric methods, and the limited candidate pool means potentially relevant papers outside the top-K may exist. The novelty assessment reflects what is visible within this constrained search scope rather than an exhaustive field survey.

Taxonomy

Core-task Taxonomy Papers
50
3
Claimed Contributions
26
Contribution Candidate Papers Compared
0
Refutable Paper

Research Landscape Overview

Core task: feedforward 3D scene reconstruction from sparse multi-view images. The field has evolved around several complementary representation paradigms, each offering distinct trade-offs between speed, quality, and geometric fidelity. Gaussian Splatting-Based Feedforward Reconstruction has emerged as a prominent branch, leveraging efficient point-based primitives for real-time rendering; within this branch, works like MVSplat[1] and MVSplat360[8] demonstrate how 3D Gaussians can be predicted directly from sparse views. Neural Radiance Field-Based approaches (e.g., ReConFusion[3], DistillNeRF[28]) continue to explore implicit volumetric representations, while Pointmap and Depth-Based methods (MV-DUST3R+[2], Surf3R[5]) emphasize explicit geometric priors. Hybrid and Multi-Stage frameworks (UniForward[7], Forge4D[6]) combine multiple representations or refine predictions iteratively, and Pose Estimation integration (No Pose Problem[9], Light3R-SfM[24]) addresses the challenge of unknown camera parameters. Specialized domains (DrivingForward[4], CAD-NeRF[17]) and theoretical surveys (Feed-forward Review[11], Sparse-view Survey[14]) round out the taxonomy, alongside alternative geometric representations such as line-based methods (PluckeRF[46]). Recent activity highlights a tension between representation expressiveness and computational efficiency. Gaussian splatting methods have attracted considerable attention for their balance of quality and speed, yet debates persist over whether 3D or 2D Gaussian primitives better capture surface detail and view-dependent effects. SurfSplat[0] sits within the 2D Gaussian Splatting subgroup, emphasizing surface-aligned representations that may offer improved geometric accuracy compared to volumetric 3D Gaussians like those in MVSplat[1] or GS-LRM[16]. This focus on surface fidelity contrasts with works such as Surf3R[5], which prioritizes dense pointmap prediction, and ReConFusion[3], which relies on diffusion-based NeRF refinement. The interplay between feedforward efficiency and reconstruction quality remains a central open question, with SurfSplat[0] contributing to the ongoing exploration of how geometric priors and representation choices shape the sparse-view reconstruction landscape.

Claimed Contributions

SurfSplat feedforward framework with surface continuity prior and forced alpha blending

The authors introduce SurfSplat, a feedforward model that uses 2D Gaussian Splatting primitives to reconstruct 3D scenes from sparse images. The method incorporates a surface continuity prior that binds rotation and scale attributes to spatial positions, and a forced alpha blending strategy to prevent opacity collapse and maintain 3D alignment.

6 retrieved papers
High-Resolution Rendering Consistency (HRRC) evaluation metric

The authors propose HRRC, a novel evaluation metric that assesses 3D scene quality by rendering at higher resolutions to expose geometric artifacts like spatial voids and discontinuities that are hidden at standard resolutions. This metric can be computed from standard datasets without requiring new annotations.

10 retrieved papers
State-of-the-art performance on multiple benchmarks

The authors demonstrate through comprehensive experiments that SurfSplat outperforms prior methods on standard novel view synthesis metrics and the proposed HRRC metric across RealEstate10K, DL3DV, and ScanNet datasets, establishing a new performance benchmark for sparse-view 3D reconstruction.

10 retrieved papers

Core Task Comparisons

Comparisons with papers in the same taxonomy category

Within the taxonomy built over the current TopK core-task papers, the original paper is assigned to a leaf with no direct siblings and no cousin branches under the same grandparent topic. In this retrieved landscape, it appears structurally isolated, which is one partial signal of novelty, but still constrained by search coverage and taxonomy granularity.

Contribution Analysis

Detailed comparisons for each claimed contribution

Contribution

SurfSplat feedforward framework with surface continuity prior and forced alpha blending

The authors introduce SurfSplat, a feedforward model that uses 2D Gaussian Splatting primitives to reconstruct 3D scenes from sparse images. The method incorporates a surface continuity prior that binds rotation and scale attributes to spatial positions, and a forced alpha blending strategy to prevent opacity collapse and maintain 3D alignment.

Contribution

High-Resolution Rendering Consistency (HRRC) evaluation metric

The authors propose HRRC, a novel evaluation metric that assesses 3D scene quality by rendering at higher resolutions to expose geometric artifacts like spatial voids and discontinuities that are hidden at standard resolutions. This metric can be computed from standard datasets without requiring new annotations.

Contribution

State-of-the-art performance on multiple benchmarks

The authors demonstrate through comprehensive experiments that SurfSplat outperforms prior methods on standard novel view synthesis metrics and the proposed HRRC metric across RealEstate10K, DL3DV, and ScanNet datasets, establishing a new performance benchmark for sparse-view 3D reconstruction.

Table of Contents

SurfSplat: Conquering Feedforward 2D Gaussian Splatting with Surface Continuity Priors | Novelty Validation