Skip to content

title: >- [Paper Note] SubSplat: High-Resolution Pixel-aligned 3DGS via Sub-pixel Gaussian Reparameterization description: >- [ECCV 2026][3D Vision][Pixel-aligned 3DGS] Introduces Sub-pixel Gaussian Reparameterizer (SPGR) to subdivide primary Gaussians into high-density 3D primitives from low-resolution inputs, decoupling grid resolution from primitive density and achieving real-time high-fidelity novel view synthesis. tags: - ECCV 2026 - 3D Vision - 3D Gaussian Splatting - Pixel-aligned 3DGS - High-Resolution Rendering date: 2026-09-19 content_hash: d5d4fb80edb3b433

SubSplat: High-Resolution Pixel-aligned 3DGS via Sub-pixel Gaussian Reparameterization

Conference: ECCV 2026
Paper: ECCV Official Page
Area: 3D Vision
Keywords: 3D Gaussian Splatting, pixel-aligned 3DGS, generalizable novel-view synthesis, sub-pixel reparameterization, high-resolution rendering

TL;DR

SubSplat introduces a Sub-pixel Gaussian Reparameterizer (SPGR) coupled with deformable multi-view feature aggregation, which subdivides each grid-anchored primary Gaussian into fine-grained 3D primitives with conserved opacity redistribution directly from low-resolution inputs, eliminating quadratic backbone computation while achieving real-time, high-fidelity rendering.

Background & Motivation

Pixel-aligned feed-forward 3D Gaussian Splatting (3DGS) models have emerged as an effective paradigm for generalizable novel-view synthesis from sparse views, circumventing the slow and compute-heavy per-scene optimization of vanilla 3DGS. However, existing pixel-aligned frameworks predict Gaussian primitives on a rigid image grid, inherently tying the total count of 3D primitives to the input spatial resolution. When scaling to high-resolution outputs required by modern interactive AR/VR displays and high-density screens (e.g., \(1024 \times 1024\)), this coupled grid design encounters a severe quality-versus-efficiency dilemma.

A naive strategy to satisfy high-resolution rendering is scaling up the input image resolution to match the target canvas. Nevertheless, this incurs quadratic computational and memory growth in the backbone network (doubling resolution quadruples computation), leading to prohibitive inference latency and recurrent out-of-memory (OOM) failures. Conversely, keeping the input resolution low while forcing the coarse, grid-anchored Gaussians to render at high output resolutions results in severe under-parameterization relative to target pixels, producing prominent blur, jagged edges, and halo artifacts. Furthermore, applying 2D post-processing upsamplers (such as bilinear interpolation or super-resolution networks) in image space fails to recover underlying 3D spatial geometry and often induces flickering and cross-view inconsistencies.

The key insight of this paper is to decouple the computational grid resolution of the backbone from the primitive density in 3D space, keeping the heavy backbone network operating purely at a low resolution while lightweightly scaling primitive density during geometric reparameterization. Core idea: introduce the Sub-pixel Gaussian Reparameterizer (SPGR) driven by three-stage deformable attention multi-view feature aggregation, which subdivides each primary Gaussian into \(K\) fine-grained sub-pixel primitives with footprint-aware opacity conservation and color modulation, recovering sharp structural details with negligible computational overhead.

Method

Overall Architecture

SubSplat takes sparse, low-resolution input views (e.g., \(256 \times 256\)) and synthesizes high-resolution novel views (e.g., \(512 \times 512\) or \(1024 \times 1024\)) end-to-end. The pipeline consists of three coordinated stages: first, a pixel-aligned backbone predicts initial primary Gaussians \(\mathcal{A}\) and spherical harmonic coefficients \(\mathcal{H}\); second, a multi-stage deformable attention module aggregates multi-view feature maps to construct geometry and appearance descriptors; third, the Sub-pixel Gaussian Reparameterizer (SPGR) employs dedicated prediction heads to forecast primitive-level 3D spatial perturbations (sub-pixel offsets, depth residuals, anisotropic scale updates, and axis-angle rotations) as well as appearance adjustments (footprint-aware opacity redistribution and color modulation), producing dense sub-pixel primitives that are rendered in real time via a differentiable rasterizer.

%%{init: {'flowchart': {'rankSpacing': 24, 'nodeSpacing': 28, 'padding': 6, 'wrappingWidth': 400}}}%%
flowchart TD
    In["Sparse Low-Resolution Inputs<br/>(e.g., 256×256)"] --> Init["Pixel-Aligned Backbone Initialization<br/>Predict primary Gaussians A and SH colors H"]
    Init --> FA["Multi-View Feature Aggregation<br/>Three-stage deformable attention for geometry & appearance"]
    FA --> Geom["Geometry Reparameterization (SPGR Geom Head)<br/>Predict sub-pixel 3D offsets, depth residuals, scale & rotation"]
    FA --> App["Appearance Reparameterization (SPGR App Head)<br/>Footprint-aware opacity redistribution & color modulation"]
    Geom --> Prims["K Fine-Grained Sub-Pixel 3D Primitives<br/>(K = 4 or 16)"]
    App --> Prims
    Prims --> Rast["Differentiable Gaussian Rasterizer<br/>Real-time high-resolution novel view output"]

Key Designs

1. Multi-View Deformable Attention Feature Aggregation: Decoupled Geometry Guidance and Appearance Conditioning To supply rich, view-consistent context for sub-pixel primitive subdivision, SubSplat aggregates cross-view multi-scale features instead of relying solely on coarse single-view representations. For each primary Gaussian \(p\), attribute-specific MLPs project its center, scale, quaternion, and opacity into a unified geometry query \(\mathbf{z}^g_p\). The anchor position is projected onto reference views via camera matrices, and a three-stage deformable attention module progressively samples multi-view feature maps from local to global scales, yielding a compact geometry descriptor \(\mathbf{f}^g_p\). For appearance, the flattened spherical harmonic coefficients form the query \(\mathbf{z}^a_p\), and deformable attention is conditioned directly on the geometry descriptor \(\mathbf{f}^g_p\). This cascading design ensures that geometry determines spatial alignment while appearance amplifies high-frequency color variations, preserving cross-view multi-perspective consistency.

2. Geometry Reparameterization: Ray-Aligned Sub-Pixel 3D Offsets and Covariance Evolution The geometry head \(g_\theta\) maps each geometry descriptor \(\mathbf{f}^g_p\) to geometric perturbation vectors for \(K\) sub-pixel primitives, where \(K\) corresponds to the ratio of output-to-input pixel count (\(K=4\) for \(2\times\) scale, \(K=16\) for \(4\times\) scale). To maintain 3D geometric stability across varying viewpoints, the reference camera with the smallest positive depth \(z_{\text{sel}}\) is selected as the primary coordinate frame. In this camera plane, 2D sub-pixel offsets \((\Delta u_k, \Delta v_k)\) anchored around a uniform sub-grid are back-projected through intrinsics and combined with a bounded depth residual \(\Delta \hat{z}_k = \tanh(\Delta z_k)\) along the viewing ray: $$ \mathbf{\mu}{p,k} = \mathbf{\mu}_p + R $$ Furthermore, each sub-pixel Gaussian inherits and refines the primary orientation and scale via log-scale residuals } \begin{bmatrix} z_{\text{sel}} \frac{\Delta u_k}{f_x} \ z_{\text{sel}} \frac{\Delta v_k}{f_y} \ \Delta \hat{z}_k \end{bmatrix\(\Delta \log \mathbf{s}_k\) and axis-angle rotation \(\Delta \mathbf{r}_k\): scale is updated element-wise via \(\mathbf{s}_{p,k} = \mathbf{s}_p \odot \exp(\Delta \log \mathbf{s}_k)\), and rotation is updated via normalized quaternion composition \(q_{p,k} = \text{norm}(q_p \otimes q(\Delta \mathbf{r}_k))\). This allows the subdivided primitives to stretch, compress, and rotate adaptively to align with fine surface contours.

3. Footprint-Aware Opacity Redistribution and Color Modulation: Balancing Sharp Edges with Anti-Aliasing Subdividing Gaussians without careful opacity control leads to density explosion and blending artifacts. The appearance head \(h_\phi\) predicts unnormalized distribution logits \(\ell_k\) and a bounded color gain \(\gamma_k\). To sharpen structural boundaries, the model computes the screen-space projected footprint area of each sub-primitive via the projection Jacobian \(J_p\) as \(\tilde{A}_k \approx \sqrt{\det(J_p \Sigma_k J_p^\top)}\), and derives inverse-area weights \(\tilde{w}_k = \frac{\text{softmax}(\ell_k)}{\tilde{A}_k + \varepsilon}\). The final opacity \(\alpha_k\) is assigned as: $$ \alpha_k = \alpha_p \cdot \frac{\tilde{w}_k}{\sum_j \tilde{w}_j} $$ This footprint-aware distribution concentrates density onto smaller primitives that delineate sharp edges while strictly preserving total opacity conservation (\(\sum_{k=1}^K \alpha_k = \alpha_p\)). Color coefficients are modulated via \(\mathbf{h}_{p,k} = \gamma_k \mathbf{h}_p\), introducing subtle appearance variance without drifting from the primary color. To prevent sub-pixel primitives from flickering or vanishing at high rendering scales, a lower bound \(A_{\min}\) clamps the projected area to guarantee robust anti-aliasing.

Loss & Training

The framework is trained end-to-end using a weighted objective combining \(\ell_2\) photometric reconstruction loss and LPIPS perceptual loss: $$ \mathcal{L} = \mathcal{L}2 + \lambda \mathcal{L} $$ where }\(\lambda\) is set to 0.05. The model is trained on a single NVIDIA A100 (80GB) GPU for 300K iterations with batch size 12, utilizing the Adam optimizer and a OneCycleLR scheduler with an initial peak learning rate of \(2 \times 10^{-4}\) and a 2K-step warm-up. For \(4\times\) inference (\(1024 \times 1024\)), a base model trained on \(128 \times 128 \to 512 \times 512\) directly generalizes to \(256 \times 256 \to 1024 \times 1024\) without retraining.

Key Experimental Results

Main Results

On the large-scale RealEstate10K and ACID benchmarks, all methods take \(256 \times 256\) inputs and are evaluated at \(2\times\) (\(512 \times 512\)) and \(4\times\) (\(1024 \times 1024\)) output scales:

Dataset Output Scale Metric PixelSplat MVSplat DepthSplat HiSplat Ours (SubSplat) Gain / Improvement
RealEstate10K \(512 \times 512\) (\(2\times\)) PSNR↑ 23.44 19.46 19.54 23.26 25.52 +2.08 dB vs PixelSplat
RealEstate10K \(512 \times 512\) (\(2\times\)) SSIM↑ 0.789 0.751 0.720 0.803 0.850 +0.047 vs HiSplat
RealEstate10K \(512 \times 512\) (\(2\times\)) LPIPS↓ 0.227 0.269 0.309 0.207 0.167 -0.040 vs HiSplat
RealEstate10K \(1024 \times 1024\) (\(4\times\)) PSNR↑ 21.44 16.98 16.29 22.05 22.65 +0.60 dB vs HiSplat
RealEstate10K \(1024 \times 1024\) (\(4\times\)) SSIM↑ 0.767 0.671 0.587 0.779 0.781 +0.002 vs HiSplat
RealEstate10K \(1024 \times 1024\) (\(4\times\)) LPIPS↓ 0.323 0.424 0.461 0.309 0.268 -0.041 vs HiSplat
ACID \(512 \times 512\) (\(2\times\)) PSNR↑ 24.41 18.38 17.19 23.59 26.03 +1.62 dB vs PixelSplat
ACID \(512 \times 512\) (\(2\times\)) SSIM↑ 0.709 0.645 0.535 0.724 0.775 +0.051 vs HiSplat
ACID \(512 \times 512\) (\(2\times\)) LPIPS↓ 0.292 0.349 0.469 0.273 0.216 -0.057 vs HiSplat
ACID \(1024 \times 1024\) (\(4\times\)) PSNR↑ 22.95 16.57 16.32 22.03 23.35 +0.40 dB vs PixelSplat
ACID \(1024 \times 1024\) (\(4\times\)) SSIM↑ 0.661 0.616 0.630 0.674 0.674 On par with best baseline
ACID \(1024 \times 1024\) (\(4\times\)) LPIPS↓ 0.409 0.489 0.476 0.388 0.330 -0.058 vs HiSplat

Cost-performance trade-off at target \(512 \times 512\) resolution on RealEstate10K (tested on an NVIDIA A100 GPU):

Method Input Res. Output Res. Latency (ms)↓ Peak Memory (GB)↓ PSNR↑ SSIM↑ LPIPS↓
PixelSplat \(512 \times 512\) \(512 \times 512\) 561 9.56 24.80 0.839 0.190
MVSplat \(512 \times 512\) \(512 \times 512\) 131 4.27 24.98 0.842 0.172
TranSplat \(512 \times 512\) \(512 \times 512\) 192 5.14 23.50 0.804 0.197
DepthSplat \(512 \times 512\) \(512 \times 512\) 123 4.58 24.70 0.851 0.170
HiSplat \(512 \times 512\) \(512 \times 512\) 1960 4.87 25.34 0.850 0.168
Ours (SubSplat) \(256 \times 256\) \(512 \times 512\) 42 1.57 25.52 0.850 0.167

SubSplat achieves higher quality with only 42 ms latency (24 FPS, real-time) and 1.57 GB peak memory, reducing latency by over \(3\times\) compared to full-resolution MVSplat and \(46\times\) compared to multi-stage HiSplat.

Ablation Study

Ablation on RealEstate10K across scaling factors \(K=4\) and \(K=16\), and Deformable Attention (DA) stages:

Configuration / Stage \(K=4\) (\(512\times512\)) PSNR↑ \(K=4\) SSIM↑ \(K=4\) LPIPS↓ \(K=16\) (\(1024\times1024\)) PSNR↑ \(K=16\) SSIM↑ \(K=16\) LPIPS↓
Baseline (MVSplat) 19.46 0.751 0.269 16.98 0.671 0.424
+ Sub-pixel Reparam. (SPGR) 25.15 (+5.69) 0.843 0.173 19.49 (+2.51) 0.736 0.415
+ Feature Aggregation (Full) 25.52 (+0.37) 0.850 0.167 22.65 (+3.16) 0.781 0.268
Deformable Attention (DA) Stages PSNR↑ SSIM↑ LPIPS↓ Note
One-stage DA 23.94 0.833 0.194 Baseline cross-view sampling
Two-stage DA 24.52 0.839 0.183 Progressive contextual refinement
Three-stage DA (Full) 25.52 0.850 0.167 +1.58 dB gain over single-stage

Key Findings

  • SPGR drives the primary fidelity recovery: At \(K=4\), introducing SPGR alone boosts PSNR from 19.46 dB to 25.15 dB (+5.69 dB), confirming that expanding primitives in 3D space resolves grid under-parameterization.
  • Feature aggregation becomes vital at extreme scales: At \(K=16\) (\(1024 \times 1024\)), relying solely on SPGR yields 19.49 dB due to geometric drift; adding multi-view deformable feature aggregation lifts PSNR by +3.16 dB to 22.65 dB, illustrating that high-ratio subdivision requires multi-view contextual guidance.
  • Superiority over 2D image-space super-resolution: Coupling MVSplat with the 2D super-resolution model HiT-SR achieves only 23.88 dB PSNR with 215 ms latency and noticeable cross-view ringing; SubSplat reaches 25.52 dB at 42 ms, confirming that 3D primitive densification is fundamentally superior to 2D image-space post-processing.

Highlights & Insights

  • Decoupling Backbone Grid Resolution from 3D Primitive Density: Conventional pixel-aligned 3DGS is forced to scale inputs to scale primitive density, triggering \(\mathcal{O}(HW)\) or quadratic latency growth. SubSplat demonstrates that dense 3D primitives can be effectively derived from compact low-resolution features via lightweight geometric reparameterization.
  • Footprint-Aware Opacity Allocation: By penalizing opacity on larger projected primitives and concentrating weight on smaller sub-pixel primitives, the network sharpens high-frequency geometric contours while mathematically preserving total integrated opacity.
  • Real-Time Interactive Deployment: With 42 ms latency (24 FPS) and 1.57 GB memory footprint on an A100 GPU, SubSplat is the first method among feed-forward pixel-aligned baselines to enable high-resolution generalizable novel-view synthesis at interactive frame rates.

Limitations & Future Work

  • Uniform vs. Content-Aware Primitive Subdivision: Currently, every primary Gaussian is uniformly subdivided into a fixed count of \(K\) primitives, leading to redundant primitives in planar, textureless regions. Future work could explore content-aware dynamic density allocation to optimize rendering throughput.
  • Extreme Scale Hallucination Boundaries: Beyond \(4\times\) scaling (e.g., \(8\times\) or \(16\times\)), deterministic feature aggregation from low-resolution inputs reaches its information-theoretic limits; integrating 3D generative diffusion priors could help hallucinate plausible microscopic textures.
  • vs MVSplat [Chen et al., ECCV 2024]: MVSplat predicts one Gaussian per pixel on a fixed grid, resulting in severe blur or cubic memory overhead at high resolutions; SubSplat builds upon its cost-volume backbone but introduces SPGR, achieving over \(3\times\) latency reduction and up to +6 dB PSNR improvement at high resolutions.
  • vs HiSplat [Tang et al., 2024]: HiSplat improves quality via multi-stage feature extraction but incurs 1960 ms latency per frame; SubSplat uses multi-stage features to guide a lightweight reparameterizer, achieving equal or superior fidelity at 42 ms (a \(46\times\) speedup).
  • vs PixelSplat [Charatan et al., CVPR 2024]: PixelSplat allocates three Gaussians per pixel, which offers partial coverage but cannot flexibly scale density; SubSplat dynamically defines \(K\) based on target canvas resolution and guarantees footprint-aware opacity conservation.

Rating

  • Novelty: ⭐⭐⭐⭐ [Reconceptualizes high-resolution feed-forward 3DGS through 3D sub-pixel primitive reparameterization]
  • Experimental Thoroughness: ⭐⭐⭐⭐⭐ [Extensive evaluations on RealEstate10K and ACID across \(2\times\) and \(4\times\) scales, detailed latency/memory profiling, and solid ablations]
  • Writing Quality: ⭐⭐⭐⭐⭐ [Clear mathematical formulations, cohesive prose, and well-structured comparative analysis]
  • Value: ⭐⭐⭐⭐⭐ [Resolves the critical quality-efficiency bottleneck of generalizable 3DGS, providing an actionable blueprint for real-time mobile and VR novel view synthesis]