HybridSplat: Hybrid 3D Gaussian Splatting

• HybridSplat is a family of hybrid rendering techniques that blend Gaussian splatting with reflection baking, NeRF modulation, mesh integration, and more to overcome traditional limitations.
• These methods retain real-time performance while enhancing photorealism by addressing view-dependent effects, redundancy, and geometry–appearance disentanglement issues.
• HybridSplat achieves significant speed, memory efficiency, and improved rendering quality in reflective, dynamic, and complex scenes using GPU-optimized pipelines.

HybridSplat refers to a family of hybrid techniques that combine Gaussian splatting primitives with additional mechanisms (reflection baking, learned latents, triangle meshes, or neural modulation) to overcome the fidelity, efficiency, and generalization limits of conventional splatting. These methods retain the inherent real-time performance and high-quality view synthesis of 3D Gaussian splatting (3DGS), but solve crucial weaknesses in handling specularities, complex reflections, dynamic scenes, geometric sparsity, and disentanglement between geometry and appearance. The term “HybridSplat” is applied both to the specific state-of-the-art method for reflection-baked Gaussian tracing (Liu et al., 9 Dec 2025) and to a broader class of approaches that hybridize 3DGS with other representations or compositing schemes (Malarz et al., 2023, Oh et al., 19 May 2025, Jiang et al., 26 May 2025, Kelkar et al., 16 Apr 2026).

1. Motivation and Limitations of Classic 3DGS

Classic 3D Gaussian splatting models a scene as a set of millions of anisotropic Gaussian primitives, rendered via tile-based screen-space splatting. This paradigm excels for static, diffuse, or mildly specular scenes and achieves real-time rates due to GPU-friendly compositing. However, certain scene attributes pose intrinsic challenges:

• View-dependent effects such as mirror-like reflections and complex lighting require tracing secondary rays or encoding view-dependent reflectance beyond the capacity of spherical harmonic (SH) coefficients.
• Redundancy: High-fidelity reconstruction of reflective or dynamic phenomena forces the model to utilize excessive numbers of Gaussians, inflating memory storage and decreasing speed.
• Conditioning limitations: Purely local primitives often cannot capture long-range effects or separate geometry from appearance, resulting in artifacts with high-frequency textures hiding geometric errors.

HybridSplat approaches address these by “hybridizing” splatting. Architectures include combining splatting with reflection baking (Liu et al., 9 Dec 2025), with NeRF-style latent MLPs (Malarz et al., 2023), with explicit mesh/triangle cores (Jiang et al., 26 May 2025), or with frequency-aware representation separation (Kelkar et al., 16 Apr 2026).

2. Reflection-Baked Hybrid Splatting

The HybridSplat method (“Fast Reflection-baked Gaussian Tracing using Hybrid Splatting” (Liu et al., 9 Dec 2025)) introduces a tractable solution for photorealistic rendering of scenes with complex specular reflections, side-stepping the computational burden of frame-wise per-pixel ray-tracing. The principal innovations are:

• Reflection-baked Gaussian tracing: Each reflective Gaussian $g_r$ contains a learned coefficient $\mathbf{r}$ and, at render time, executes a single reflection-ray trace to precompute a baked reflection $\mathbf{r}_{REF}$ stored within the primitive.
• Mathematical formulation: The process computes the incident view vector $v$ and mirror-reflection direction $d_r$. Tracing through a sorted list $N_r$ of reflective Gaussians along $d_r$, the model accumulates

$$\mathbf{r}_{REF} = \sum_{k\in N_r} \phi(\mathbf{r}^k, \mathbf{d}_r)\,\alpha_k\,\prod_{j<k}(1 - \alpha_j)$$

where $\phi$ encodes Fresnel modulation or a learned view factor and $\alpha_k$ is the projected screen-space opacity.

• Splat pipeline integration: The reflection term is composited via a unified tile-based pipeline, enabling high speed and shared memory access for both base and reflective contributions.

This mechanism replaces millions of per-frame intersection tests with a single per-Gaussian bake, enabling tiled rasterization and amortizing the cost of view-dependent reflection.

3. Unified Rendering and Data Pipeline

HybridSplat introduces a dual set of Gaussian primitive types:

• Base Gaussians $\mathbf{r}$0: Standard 3DGS primitives with SH color.
• Reflective Gaussians $\mathbf{r}$1: Carry view-dependent reflection coefficients.

Rendering proceeds as follows:

• For each screen-space tile, the system collects all intersecting Gaussians.
• For each pixel, compositing proceeds front-to-back, accumulating both base color and reflection contributions, then blends with a per-pixel mixing weight $\mathbf{r}$2:

$\mathbf{r}$3

with $\mathbf{r}$4. All stages—preprocessing, sorting, rendering—are fully parallelized and implemented for GPU wavefront execution, minimizing latency and maximizing data locality.

4. Pipeline Acceleration and Pruning

Efficiency and compactness are achieved through several strategies:

• Wavefront pipeline: Preprocessing, Gaussian attribute gathering, and compositing are staged in flight so GPU compute is saturating for all steps.
• Memory layout: All Gaussian attributes are packed into contiguous arrays, and tile-to-Gaussian index lists enable coalesced memory access.
• Reflection-sensitive pruning: The second-order pruning score $\mathbf{r}$5 for each Gaussian incorporates both base and reflective gradient contributions:

$\mathbf{r}$6

Gaussians with the lowest scores are periodically pruned, yielding models with $\mathbf{r}$7 fewer active primitives and $\mathbf{r}$8 improved frame rates versus tracing-based baselines, with $\mathbf{r}$9\,dB PSNR loss on Ref-NeRF and NeRF-Casting scenes.

5. Extensions in Hybrid Splatting

Other HybridSplat strategies adapt this hybridization for broader challenges:

• NeRF-modulated splatting (VDGS): VDGS (Malarz et al., 2023) augments each 3D Gaussian with a NeRF-style MLP modulating per-Gaussian color/opacity as a function of viewing direction, spatial parameters, and Fourier-encoded ray direction. This captures dynamic appearance (shadows, reflections, transparency) beyond fixed SH or baked coefficients, with little cost increase.
• Hybrid 3D–4D Splatting: Hybrid 3D-4DGS (Oh et al., 19 May 2025) adaptively collapses the full 4D temporal representation to 3D for primitives that are temporally invariant, based on a learned log-scale threshold on the time axis. This approach achieves $\mathbf{r}_{REF}$0–$\mathbf{r}_{REF}$1 faster convergence and $\mathbf{r}_{REF}$2–$\mathbf{r}_{REF}$3 memory compression relative to pure 4DGS while matching dynamic scene fidelity.
• Frequency-aware hybrid surfel splatting: Hybrid Latents (Kelkar et al., 16 Apr 2026) introduces per-surfel latents for low-frequency geometry and appearance, combined with high-frequency residuals from a hash grid, fused by late alpha blending and decoded for color. This reduces the entanglement whereby fine texture compensates for poor geometry, and enforces sparsity via probabilistic pruning driven by a binary cross-entropy loss.
• Hybrid triangle-Gaussian scene representation: HaloGS (Jiang et al., 26 May 2025) maintains two loosely coupled components: explicit triangle soups for structured geometry and neural Gaussians for texture/appearance, optimized alternately with monocular priors and image supervision. This enables compact, mesh-extractable geometric models with photorealistic renderings, even in large-scale and planar-dominant scenes.

6. Quantitative and Qualitative Performance

HybridSplat (reflection-baked variant) outperforms both non-tracing and traced splatting baselines, especially for reflective scenes:

Method	PSNR	SSIM	LPIPS	FPS	Speedup	# Gaussians
3DGS (no trace)	23.9	0.64	0.265	228	1×	3.03 M
Speedy-Splat	24.3	0.66	0.324	2305	10×	0.23 M
MetaSapiens	24.3	0.63	0.330	1359	6×	0.21 M
EnvGS (trace)	24.6	0.67	0.282	15	1×	1.41 M
Ref-Gaussian	24.6	0.69	0.252	66	4.4×	0.59 M
HybridSplat (Ours)	24.4	0.66	0.280	107	7.1×	0.39 M

Key experimental outcomes:

• HybridSplat achieves sharp specular highlights and physically plausible reflections, closely matching ground truth for mirror-like and environment-reflective objects.
• Qualitatively, pure non-tracing approaches yield blurry or ghosted reflections, while ray-tracing approaches are slow.
• Across benchmarks (Ref-NeRF, NeRF-Casting, Mip-NeRF360, Tanks & Temples, DeepBlending), HybridSplat maintains or exceeds top-tier PSNR/SSIM/LPIPS with order-of-magnitude speed/memory benefit.

7. Limitations and Prospects

HybridSplat methods, while addressing several core deficiencies, inherit notable constraints:

• The reflection baking is inherently “one-bounce”; almost all methods under-represent deeper inter-reflections without an iterative or more global scheme.
• Accurately modeling very near-field mirror surfaces or rapidly varying normals can introduce artifacts due to the single-ray reflection approximation or normal estimation errors (Liu et al., 9 Dec 2025).
• Hybrid 3D–4DGS relies on heuristic thresholding for static-dynamic separation, and densification remains ad hoc (Oh et al., 19 May 2025).
• Frequency-aware approaches depend critically on effective separation of geometry vs. appearance; inadequate tuning can induce leakage or over-smoothing (Kelkar et al., 16 Apr 2026).
• Dual representations (triangle + Gaussian) are well suited for strong planar structure but are less effective on highly curved/organic geometry (Jiang et al., 26 May 2025).

Proposed research extensions include multi-bounce reflection baking, learnable Fresnel/falloff functions, hardware-accelerated intersection during baking, adaptive tile sizing, and unified densification/pruning strategies. The implementation of continuous, learnable staticness scores and integration of higher-order geometric primitives are noted directions for future work across hybrid splatting research.

• "HybridSplat: Fast Reflection-baked Gaussian Tracing using Hybrid Splatting" (Liu et al., 9 Dec 2025)
• "Hybrid 3D-4D Gaussian Splatting for Fast Dynamic Scene Representation" (Oh et al., 19 May 2025)
• "Hybrid Latents -- Geometry-Appearance-Aware Surfel Splatting" (Kelkar et al., 16 Apr 2026)
• "Gaussian Splatting with NeRF-based Color and Opacity" (Malarz et al., 2023)
• "HaloGS: Loose Coupling of Compact Geometry and Gaussian Splats for 3D Scenes" (Jiang et al., 26 May 2025)