PaveSplat: Resolution-Robust Feed-Forward Gaussian Splatting with Residually Corrected Surface-Aligned Covariances

ETRI, Electronics and Telecommunications Research Institute ACCV 2026, Osaka, Japan, December 14–18, 2026

ACCV 2026 · Osaka, Japan Dec 14–18, 2026

Supplementary video (96.7 s): the re-rendering idea (illustration), real re-renders at 1×–8×, coverage holes, test-set results, covariance ellipsoids.

TL;DR

Feed-forward 3DGS predicts covariances (scale, rotation) that work well at the training resolution — but re-render the same prediction at a higher resolution and holes open.

PaveSplat keeps the analytic surface-aligned frame and learns one identity-initialized rotation residual Δq, so surfels keep paving the surface.

Abstract

We present PaveSplat, a feed-forward Gaussian splatting method that adds a learned residual to the analytic surface-aligned covariance, keeping surfels paving the visible surface across rendering resolutions. Pixel-aligned feed-forward Gaussian splatting predicts one primitive per pixel from a few images in a single pass, and its rendering quality depends on how well these primitives cover the surface. Recent surfel-based methods compute the surface-aligned covariance analytically from predicted depth. However, this covariance has no way to correct the surfel orientation where the depth-derived frame is ill-posed. The surfels then cover the surface loosely; at the training resolution each surfel’s one-pixel minimum footprint covers the gaps, but they show up as holes when the same prediction is re-rendered at a higher resolution. To address this, we keep the analytic covariance as an initialization, with its axes aligned to the surface tangents, and add an identity-initialized unit-quaternion residual that reorients a surfel only where this improves rendering, turning surfels that the analytic frame mis-aims back onto the surface. On RealEstate10K, PaveSplat achieves the best PSNR, SSIM, and LPIPS among single-Gaussian-per-pixel methods at every scale (1×–8×). The native-resolution gain is small; the benefit appears when the same prediction is rendered at higher display resolutions without re-running the network. On zero-shot DL3DV, it leads PSNR within this group at every scale and ranks first or second on SSIM/LPIPS. A coverage-hole analysis shows fewer uncovered pixels than both baselines under high-resolution re-rendering.

The problem: holes at high resolution

SurfSplat‡

SurfSplat‡ at 8×: rows of black holes open across the wall and the doorway
SurfSplat‡ at 1×: the same prediction looks covered

PaveSplat (ours)

PaveSplat at 8×: the wall and the doorway stay covered
PaveSplat at 1×
Same prediction re-rendered at 8× (2048²). Insets: the 1× (256²) frame, dashed box = the 8× crop. SurfSplat‡ = same-recipe reproduction.
  • Most feed-forward 3DGS methods predict pixel-aligned Gaussians — typically one per pixel — from a few input views in a single pass.
  • At 1×, each surfel’s one-pixel minimum footprint hides the gaps between loosely placed surfels.
  • Re-render the same prediction on a 2×–8× denser grid and the gaps show up as holes.
  • Root cause: depth sets each surfel’s size from its pixel footprint, but that size only covers the surface if the orientation is right — and analytic surfels have no learned orientation to fix mis-aimed ones at depth edges, thin structures and grazing surfaces.

Method: a correctable surface-aligned covariance

Three ways to cover a surface with surfels

(a) Fully learned

Fully learned covariances drawn as ellipsoids from a novel view: scale is uncontrolled, so oversized ellipsoids over-spreadSchematic: surfels of uncontrolled size along a surface
scale uncontrolled

(b) Analytic

Analytic surface-aligned covariances from a novel view: the frame is set from depth, leaving anisotropic surfels mis-orientedSchematic: analytic surfels mis-oriented at a depth edge
frame fixed by depth

(c) PaveSplat

PaveSplat covariances from a novel view: the analytic frame is corrected per pixel, so surfels reorient to pave the surfaceSchematic: surfels reoriented to pave the surface across the edge
analytic frame + learned Δq
  1. Predicted depth z(u)
  2. Surface-aligned prior Ra=[e^1e^2n^]
    tangent-aligned frame
  3. In-plane scale multiplierssi=s¯imi Residual reorientation Δqidentity-initialized
  4. Final covariance R=RaRΔ(Δq), Σ=Rdiag(s1,s2,sϵ)2R⊤
  5. Rasterizer (unchanged)
q=qa⊗ΔqΔq= identity at initialization
  • Surface-aligned prior: in-plane axes follow the measured tangents, so the anisotropic radii act where they were measured.
  • Residual reorientation Δq: the only new learned covariance parameter — 9,512 parameters (<0.05% of the model), +2% encoder time.
  • Identity-initialized: training starts exactly at the analytic frame and rotates a surfel only where rendering asks for it.

Results: robust high-resolution re-rendering

RealEstate10K — PSNR vs. render scale

PSNR from main Tab. 1 at 1×–8× (256² to 2048²). ‡ same-recipe reproduction.
best PSNR, SSIM & LPIPS among single-Gaussian-per-pixel methods at every scale
1×–8×

Zero-shot DL3DV (trained on RealEstate10K)

PSNR from main Tab. 2.

PaveSplat leads PSNR among single-Gaussian-per-pixel methods at every scale. Zero-shot ScanNet: slightly behind at 1× (−0.10 dB), ahead at 8× (+2.511 dB).

Full Table 1
More qualitative results (main Fig. 3, supp Fig. 2)
RealEstate10K scenes with two inputs, ground truth, DepthSplat, SurfSplat and PaveSplat; native renders above boxed 8× crops.
Fig. 3: Qualitative comparison on RealEstate10K. Per scene: inputs, ground truth, DepthSplat [18], SurfSplat [7], and PaveSplat; top row = full-frame native render (crop boxed), bottom row = 8× crop. In these high-margin cases, PaveSplat keeps more support along thin structures.
Additional RealEstate10K scenes comparing native renders and 8× crops across ground truth and the three methods.
supp Fig. 2: Extended qualitative comparison on RealEstate10K. Additional scenes beyond the main-paper qualitative figure, same layout: per scene, the full-frame native render (1×, crop boxed) above a crop of the 8× (2048²) re-render; columns are the two input views, ground truth, DepthSplat [8], SurfSplat‡ [4] (our same-recipe reproduction), and PaveSplat (ours).

Source tiles in these composites are 178 px wide; they are not upscaled beyond their native size.

Why it works: mechanism evidence

Fewer coverage holes when re-rendered

Hole = pixel rendered near-black (max RGB < 0.02) where the ground truth has content; mean over 6,474 test scenes.

Design path (identical recipe per row)

The last step changes only Δq, so it isolates the residual (step 1 also adopts our camera-space formulation).
Configurationsmin/smax ↓holes 8× ↓PSNR 8× ↑
SurfSplat‡ (same recipe, normal-only frame)0.6671.00%23.342
+ tangent-aligned frame0.5801.11%23.850
+ residual Δq (ours)0.5850.78%24.266

Δq is used — it does not collapse to identity

Input view: red room with a glazed door Input view: hallway with an open door Residual angle map of the red room: bright along the door frame and object edges Residual angle map of the hallway: bright along the door edges and the floor at grazing angles

Mean residual angle 0.160 rad (≈9.1°), p99 1.073 rad (≈61.5°), concentrated at depth discontinuities and grazing surfaces.

top: input · bottom: residual angle |Δq|

Per-scene medians (main Fig. 4a)
main Fig. 4a — PaveSplat leaves fewer coverage holes under high-resolution re-rendering (RealEstate10K). The coverage-hole rate (black-hole fraction) = share of pixels rendering near-black (max RGB channel <0.02) where the GT has content (max channel >0.02).

Coverage data: medians and means

BibTeX

@inproceedings{kim2026pavesplat,
  title     = {PaveSplat: Resolution-Robust Feed-Forward Gaussian Splatting with Residually Corrected Surface-Aligned Covariances},
  author    = {Janghun Kim and Hyun-Cheol Kim and Eunbin An and Myeonghun Oh and Hyon-Gon Choo},
  booktitle = {Proceedings of the Asian Conference on Computer Vision (ACCV)},
  year      = {2026},
  eprint    = {2611.26045},
  archivePrefix = {arXiv},
  primaryClass = {cs.CV}
}