Skip to content

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

Conference: ICLR 2026 arXiv: 2602.02000 Code: https://hebing-sjtu.github.io/SurfSplat-website/ Area: 3D Vision Keywords: 2D Gaussian Splatting, feedforward 3D reconstruction, surface continuity, high-resolution rendering consistency, sparse-view

TL;DR

SurfSplat proposes a feedforward 3D reconstruction framework based on 2DGS that binds Gaussian rotation and scale to local neighborhood positions via surface continuity priors, resolves color bias through a forced alpha blending strategy, and introduces the HRRC metric to reveal reconstruction quality discrepancies at high resolutions.

Background & Motivation

Existing feedforward 3DGS methods appear to achieve strong NVS metrics at standard resolution, yet suffer from a critical overlooked problem:

Degenerate 3D scenes: Reconstructions are effectively discrete, color-biased point clouds rather than continuous surfaces, exposing severe artifacts (holes, color bias, surface discontinuities) under close-up or off-axis viewpoints.

Underutilized anisotropy: Learnable Gaussians struggle to disentangle geometry and texture through gradient supervision alone, causing Gaussians to degenerate toward near-spherical shapes.

Metric failure: Standard PSNR/SSIM/LPIPS at native resolution cannot capture geometric inaccuracies, masking true reconstruction quality.

The authors observe that directly training 2DGS is more challenging than 3DGS — the planar nature of 2D Gaussians causes small geometric perturbations to produce large deviations in rendered output, a problem exacerbated under limited supervision.

Method

Overall Architecture

SurfSplat employs a dual-path encoder (monocular Depth Anything V2 + multi-view cross-attention), fuses the resulting features through a U-Net to regress intermediate attributes (depth, scale multipliers, appearance components), and finally converts them to standard 2DGS attributes via surface continuity priors and forced alpha blending.

Key Designs

  1. Surface Continuity Prior

Core assumption: visible geometry in real scenes consists primarily of smooth, continuous surfaces, and spatially adjacent surface elements correspond to adjacent pixels in the image. This constrains Gaussian rotation and scale:

Rotation: For the 3D position \(\mathbf{p}_0\) of pixel \((h,w)\) and its neighborhood, two tangent vectors \(\mathbf{t}_1, \mathbf{t}_2\) are computed via Sobel filtering; the local normal \(\mathbf{n}\) is obtained by their cross product, and the rotation matrix is derived via the Rodrigues formula:

\(\mathbf{R} = \mathbf{I} + [\mathbf{v}]_\times + \frac{1-c}{\|\mathbf{v}\|^2}[\mathbf{v}]_\times^2\)

where \(\mathbf{v} = \mathbf{n}_0 \times \mathbf{n}\) and \(c = \mathbf{n}_0^\top \mathbf{n}\).

Scale: Coarse estimates \(\bar{\sigma}_u, \bar{\sigma}_v\) are derived from image-space neighborhood distances; the network predicts scale multipliers \(\hat{\sigma}_u, \hat{\sigma}_v\) in the range \([1/3, 3]\), yielding final scales \(\sigma_u = \bar{\sigma}_u \hat{\sigma}_u\). The depth-axis scale \(\sigma_w\) of 2DGS is fixed to zero.

Gaussian attributes are thus derived from predicted 3D positions rather than regressed independently, ensuring spatial consistency.

  1. Forced Alpha Blending

Under the surface continuity prior, the model tends to learn high-opacity Gaussians, causing occluded Gaussians to contribute negligibly to alpha blending and preventing learning of 3D structure. The solution:

  • Clip opacity with an upper bound \(\tau_{\text{opa}} < 1\) (set to 0.6), ensuring all Gaussians participate in rendering.
  • Initialize RGB colors to the DC component of the spherical harmonic basis.

  • Apply opacity-normalized compensation to the rendered output: \(C = C/\alpha\) when \(\alpha \geq \tau_\alpha\).

  • HRRC: High-Resolution Rendering Consistency Metric

The reconstructed scene is rendered at \(2\times\) and \(4\times\) resolution and compared against bicubic-upsampled ground truth:

\(\text{HRRC}_{\text{metric}} = \text{metric}(\hat{I}^{HR}, \hat{I}^{GT\uparrow})\)

This effectively exposes sparsity holes, degenerate Gaussian shapes, and surface discontinuities, distinguishing models that truly recover 3D geometry from those that merely memorize sparse viewpoints.

Loss & Training

\[L_{\text{gs}} = \sum_{m=1}^M \left(\text{MSE}(I_{\text{render}}^m, I_{\text{gt}}^m) + \lambda \cdot \text{LPIPS}(I_{\text{render}}^m, I_{\text{gt}}^m)\right)\]

\(\lambda = 0.05\); training is conducted at \(256\times256\) resolution. The Depth Anything V2 backbone uses a learning rate of \(2 \times 10^{-6}\); all other layers use \(2 \times 10^{-4}\).

Key Experimental Results

Main Results

Method RE10K 256 PSNR↑ RE10K 512 PSNR↑ RE10K 1024 PSNR↑ RE10K Avg PSNR↑
DepthSplat 27.504 20.031 16.385 21.307
MVSplat 26.359 20.408 17.966 21.578
Ours-L 27.537 26.331 24.897 26.255

Ablation Study

Configuration Standard PSNR HRRC 2× PSNR HRRC 4× PSNR Notes
3DGS baseline (DepthSplat) 27.504 20.031 16.385 Severe HRRC degradation
2DGS (SurfSplat-L) 27.537 26.331 24.897 HRRC remains stable
w/o surface prior Degraded Largely degraded Largely degraded Surface discontinuity
w/o forced blending Degraded Degraded Degraded Color bias

Key Findings

  • HRRC reveals the truth: DepthSplat achieves the best standard-resolution score (27.5) but collapses to 16.4 at 1024 resolution; SurfSplat drops only from 27.5 to 24.9, demonstrating genuine 3D structure recovery.
  • MVSplat and TranSplat similarly degrade substantially under HRRC (around 18 at 1024), indicating that surface degeneration is pervasive among feedforward 3DGS methods.
  • Cross-dataset evaluation on DL3DV and ScanNet confirms generalization.
  • pixelSplat (multiple Gaussians per pixel) performs comparatively better on HRRC (24.9), as redundant Gaussians partially compensate for surface holes.

Highlights & Insights

  1. High value in problem identification: The work exposes a widely overlooked surface degeneration issue in feedforward 3DGS; the HRRC metric deserves broader adoption.
  2. Geometry-driven attribute prediction: Deriving rotation from positions via Sobel filtering and the Rodrigues formula is an elegant and physically intuitive design.
  3. First successful application of 2DGS in feedforward settings: Demonstrates that 2DGS (planar primitives) is more suitable than 3DGS (ellipsoids) for feedforward reconstruction, offering stronger anisotropy and geometric precision.
  4. Clever design of forced blending: Constraining the opacity upper bound resolves local optima and preserves multi-layer expressiveness.

Limitations & Future Work

  • At standard resolution, the method only marginally outperforms DepthSplat; advantages are primarily reflected in HRRC.
  • The single-Gaussian-per-pixel design may provide insufficient coverage for complex scenes.
  • The HRRC metric relies on bicubic-upsampled ground truth rather than true high-resolution references.
  • Evaluation is limited to static scenes; extension to dynamic scenes remains future work.
  • Huang et al. (2024) introduced 2DGS for per-scene optimization; SurfSplat is the first to bring it into a feedforward framework.
  • The depth interaction design of DepthSplat (Xu et al., 2024b) is partially inherited, but SurfSplat replaces independent attribute regression with surface priors.
  • The design philosophy of the HRRC metric generalizes to evaluation of other 3D generation tasks.

Rating

  • Novelty: ⭐⭐⭐⭐ The surface continuity prior and HRRC metric are original contributions, though individual components are combinations of established techniques.
  • Experimental Thoroughness: ⭐⭐⭐⭐⭐ Three datasets, multi-resolution HRRC, multiple backbone variants, and cross-dataset evaluation — highly comprehensive.
  • Writing Quality: ⭐⭐⭐⭐ Problem formulation is clear, visual comparisons are compelling, and mathematical derivations are complete.
  • Value: ⭐⭐⭐⭐ The HRRC metric and the exposure of surface degeneration constitute important contributions to the community.