SubSplat: High-Resolution Pixel-aligned 3DGS via Sub-pixel Gaussian Reparameterization
SubSplat uses SPGR to render high-res views from 256×256 inputs, reaching 25.52 dB on RealEstate10K.
Key Findings
Methodology
SubSplat extends an MVSplat-style pixel-aligned pipeline with a Sub-pixel Gaussian Reparameterizer (SPGR). It first aggregates multi-view geometry and appearance via three-stage Deformable Attention, then subdivides each primary Gaussian ap=[μp,sp,qp,αp] into K sub-pixel primitives. A geometry head predicts offsets, depth residuals, log-scale residuals, and axis-angle rotation residuals; an appearance head predicts opacity logits and bounded color modulation. Footprint-aware redistribution preserves total opacity.
Key Results
- On RealEstate10K, using 256×256 inputs and rendering at 512×512, SubSplat achieves 25.52 PSNR / 0.850 SSIM / 0.167 LPIPS, clearly outperforming HiSplat (23.26 / 0.803 / 0.207) and PixelSplat (23.44 / 0.789 / 0.227).
- At 1024×1024 output on RealEstate10K ×4, SubSplat still reaches 22.65 / 0.781 / 0.268, ahead of HiSplat (22.05 / 0.779 / 0.309) and PixelSplat (21.44 / 0.767 / 0.323). On ACID ×2, it records 26.03 / 0.775 / 0.216.
- On ACID ×4, SubSplat attains 23.35 / 0.674 / 0.330, with substantially lower LPIPS than HiSplat’s 0.388 while matching its SSIM at 0.674. This indicates that the method preserves perceptual detail and structure as the output scale grows.
Significance
The paper addresses a long-standing bottleneck in pixel-aligned 3DGS: higher input resolution improves fidelity but makes backbone cost grow quadratically, whereas low-resolution input keeps latency stable but starves the renderer of Gaussian density. SubSplat breaks this coupling by restoring output density in 3D primitive space rather than inflating the input grid. This is highly relevant for interactive view synthesis on high-resolution displays, AR/VR telepresence, and deployment on resource-constrained devices, where latency and memory budgets are non-negotiable. Importantly, the method preserves the feed-forward, generalizable nature of pixel-aligned models, avoiding per-scene optimization while delivering much sharper results.
Technical Contribution
Technically, SubSplat contributes a new reparameterization paradigm rather than a conventional image-space upsampler. SPGR maps each primary Gaussian into K sub-pixel primitives and predicts per-primitive geometry and appearance refinements, while a footprint-aware opacity allocation enforces Σk αk = αp. The use of tanh-bounded depth residuals, exp-parameterized log-scale updates, and quaternion composition keeps the subdivision numerically stable. The three-stage Deformable Attention aggregator is equally important: geometry features decide where to sample across views, and appearance features decide what to amplify. Together, these pieces create a differentiable, scale-decoupled rendering pipeline that can increase primitive density without scaling backbone cost with image area.
Novelty
The novelty is not merely adding more Gaussians, but making primitive density a post-backbone, sub-pixel reparameterization problem. Compared with PixelSplat/MVSplat/HiSplat/TranSplat/DepthSplat, which remain tied to a fixed feature grid, SubSplat explicitly decouples output resolution from the network’s computational resolution. Unlike scene-specific densification or 2D super-resolution, it modifies 3D Gaussian geometry and opacity directly, using multi-view evidence to recover sharp edges and fine structures at high output scales.
Limitations
- The method depends on calibrated multi-view inputs and strong cross-view consistency. With very sparse views, severe occlusions, or inaccurate camera parameters, Deformable Attention may aggregate ambiguous evidence and SPGR may place sub-pixel primitives incorrectly, especially around thin structures and repetitive textures.
- The evaluation focuses on integer scale factors, mainly 2× and 4×, and on mostly static scenes from RealEstate10K and ACID. The paper does not fully demonstrate robustness to arbitrary scaling ratios, dynamic scenes, heavy motion blur, or highly reflective/transparent surfaces.
- Although more efficient than full-resolution baselines, SPGR still adds extra attention and subdivision overhead. The paper reports relative latency gains, but not a complete device-diverse profile, so the practical ceiling on very small edge devices remains open.
Future Work
Promising follow-ups include arbitrary-scale reparameterization, temporal SPGR for dynamic scenes, and lighter aggregation modules for mobile deployment. A useful next step would also be stronger depth priors or uncertainty-aware subdivision to better handle sparse-view ambiguity, reflective materials, and thin geometry. Finally, the community would benefit from more complete end-to-end latency and memory reports on different hardware classes, not only GPU benchmarks.
AI Executive Summary
Deep Dive
Abstract
Pixel-aligned Gaussian splatting enables efficient and generalizable novel-view synthesis. However, high-resolution rendering faces a critical trade-off where increasing input resolution improves detail at the expense of quadratically rising network computational cost. Conversely, maintaining low-resolution inputs stabilizes this cost but results in insufficient Gaussian density and artifacts. To address this, we propose SubSplat, which introduces Sub-pixel Gaussian Reparameterizer(SPGR) to subdivide primary Gaussians into fine-grained primitives, restoring structural density directly from low-resolution features. We further enhance the reparameterization quality through feature aggregation, which effectively captures high-frequency details across multiple views. Experiments on RealEstate10K and ACID demonstrate that SubSplat achieves high-fidelity rendering with superior efficiency. Our results validate that the proposed framework successfully resolves the trade-off between reparameterization fidelity and network computational cost inherent in pixel-aligned Gaussian Splatting.