SSGaussian: 3D Style Transfer via Gaussian Splatting

• SSGaussian is a 3D style transfer framework that integrates diffusion priors with Gaussian splatting to enhance semantic style fidelity and preserve instance-level structure.
• It employs a two-stage pipeline: first, key views are stylized using cross-view aligned diffusion, then instance-aware Gaussian grouping transfers the style onto the 3D scene.
• The method demonstrates superior performance on metrics like LPIPS and RMSE, delivering real-time rendering at 118 FPS and improved multi-view consistency.

SSGaussian is a label applied to several distinct Gaussian-based constructions. In an explicit titled usage, it denotes a two-stage 3D style transfer pipeline for 3D Gaussian Splatting that integrates diffusion priors through Cross-View Style Alignment and Instance-level Style Transfer in order to improve semantic style fidelity, instance-level consistency, and structural separation (Xu et al., 4 Sep 2025). The same label is also used more loosely or informally for sparse/selective 3D Gaussian splatting, semantic-spectral Gaussian scene representation, sparse-view super-resolution 3DGS, spherical Gaussian edge reconstruction, shading-aware Gaussian rendering, and Split Generalized Gaussian models in ICA (Liu et al., 2024).

1. Terminology and scope

In the most direct sense, SSGaussian is the name of "SSGaussian: Semantic-Aware and Structure-Preserving 3D Style Transfer," a method for stylizing 3D scenes represented by 3D Gaussian Splatting while preserving instance-level structure across views (Xu et al., 4 Sep 2025). Its target problem is not sparse reconstruction or semantic completion, but style transfer under multi-view consistency constraints.

At the same time, adjacent papers attach closely related meanings to the same or nearly the same label. In the MaskGaussian line, the term is used for the broader goal of sparse/selective 3D Gaussian splatting, where Gaussian existence is modeled probabilistically and pruning is treated as an adaptive differentiable selection process rather than deterministic removal (Liu et al., 2024). In "SpectralGaussians," the framework is presented as a direct instantiation of Semantic + Spectral Gaussian Splatting, again abbreviated conceptually as an SSGaussian approach, but այստեղ the emphasis is multi-spectral reflectance, illumination, and semantics rather than stylization (Sinha et al., 2024). "S2Gaussian" states that SSGaussian is an informal shorthand sometimes used in discussions for Sparse-view Super-resolution 3D Gaussian Splatting (Wan et al., 6 Mar 2025). "SGCR" states that if SSGaussian is intended to mean a spherical or isotropic splat primitive for edges, it corresponds directly to the Spherical Gaussian representation used there (Yang et al., 7 May 2025). The GaussianShader exposition uses SSGaussian as an abbreviation for a shading-aware 3D Gaussian Splatting formulation on reflective surfaces (Jiang et al., 2023). Outside graphics, the ICA paper uses SSGaussian for the Split Generalized Gaussian distribution (Spurek et al., 2018).

This distribution of meanings indicates that SSGaussian is not a stable universal acronym. In graphics, its most explicit named usage is the 3D style transfer method (Xu et al., 4 Sep 2025), but neighboring work reuses the label for semantically different Gaussian parameterizations and objectives.

2. Core representation in the named SSGaussian method

The named SSGaussian method builds on standard 3D Gaussian Splatting and Gaussian Grouping. A scene is represented by colored 3D Gaussians with mean $\mu \in \mathbb{R}^3$, covariance $\Sigma \in \mathbb{R}^{3 \times 3}$, opacity $\alpha \in \mathbb{R}$, and color $c$. Rendering follows front-to-back alpha compositing:

$$C = \sum_{i \in N} c_i \alpha_i \prod_{j=1}^{i-1}(1-\alpha_j).$$

Gaussian Grouping augments each Gaussian with a 16-dimensional learnable identity encoding $e_i$, which is rendered analogously:

$$E_{id} = \sum_{i \in N} e_i \alpha_i \prod_{j=1}^{i-1}(1-\alpha_j).$$

A linear layer $f$ and softmax then produce per-pixel identity classification over $K$ groups, written as $\mathrm{softmax}(f(E_{id}))$ (Xu et al., 4 Sep 2025).

This representation is central because the method treats stylization as a problem of preserving both 3D scene content and instance-level grouping. The paper identifies three difficulties in prior 3D style transfer: ineffective transfer of high-level style semantics from the reference image, insufficient structural clarity and instance separation, and view-dependent inconsistency when purely 2D diffusion priors are applied to multi-view stylization. The response is a two-stage pipeline. First, selected key viewpoints are stylized with a diffusion model modified for cross-view consistency. Second, those stylized key views are transferred onto the 3D Gaussian scene using instance identities inherited from Gaussian Grouping (Xu et al., 4 Sep 2025).

The method therefore couples appearance transfer to a pre-existing instance-aware 3D scene representation. A plausible implication is that the grouping mechanism is not an auxiliary convenience but the mechanism that prevents style patterns from bleeding across object boundaries during optimization.

3. Diffusion-guided multi-view stylization

The first stage stylizes selected key viewpoints with a pretrained latent diffusion model equipped with three components: IP-Adapter for style image prompting, depth-conditioned ControlNet for preserving content structure and improving cross-view consistency, and Cross-View Style Alignment (CVSA) inserted into the last upsampling block of the UNet (Xu et al., 4 Sep 2025). DDIM inversion is adopted to initialize consistent noise across views.

Within a single view, self-attention is written as

$\Sigma \in \mathbb{R}^{3 \times 3}$0

where $\Sigma \in \mathbb{R}^{3 \times 3}$1. CVSA extends this to cross-view attention over $\Sigma \in \mathbb{R}^{3 \times 3}$2 key views $\Sigma \in \mathbb{R}^{3 \times 3}$3:

$\Sigma \in \mathbb{R}^{3 \times 3}$4

with $\Sigma \in \mathbb{R}^{3 \times 3}$5 denoting the concatenation of tokens from the key views. In practice, multiple key views are processed jointly by stacking view tokens so that queries from one view may attend to keys and values from the others (Xu et al., 4 Sep 2025).

The paper places CVSA in the last upsampling block rather than earlier layers. Earlier-block insertion is reported to under-align semantics, whereas the last upsampling stage is described as semantically rich and spatially refined, giving the best trade-off between preserving fine-grained style details and enforcing instance-level consistency. The exact key-view selection strategy, the number of key views, guidance scales, diffusion steps, UNet head dimensions, and view positional encodings are not specified. Depth conditioning via ControlNet and consistent noise via DDIM inversion provide the explicit consistency mechanisms that the manuscript does specify (Xu et al., 4 Sep 2025).

The output of this stage is a set of stylized key views that preserve both style fidelity and instance-level coherence across viewpoints. These key views become the supervisory signal for the subsequent 3D optimization stage rather than the final result.

4. Instance-level transfer onto 3D Gaussian Splatting

The second stage transfers the stylized key views onto the 3D Gaussian scene by aligning Gaussian Grouping identities across the original training views and the stylized key views. For a sampled training view $\Sigma \in \mathbb{R}^{3 \times 3}$6 with identity map $\Sigma \in \mathbb{R}^{3 \times 3}$7 and stylized key views $\Sigma \in \mathbb{R}^{3 \times 3}$8 with identity map $\Sigma \in \mathbb{R}^{3 \times 3}$9, the method aggregates pixel locations sharing the same group identity into sets $\alpha \in \mathbb{R}$0 and $\alpha \in \mathbb{R}$1. It then defines the group mapping

$\alpha \in \mathbb{R}$2

This associates each training-view instance with its counterpart in the stylized key views, while falling back to global style pooling when a corresponding instance is absent (Xu et al., 4 Sep 2025).

Style transfer is then enforced by localized nearest-neighbor matching in VGG feature space. If $\alpha \in \mathbb{R}$3 and $\alpha \in \mathbb{R}$4 are high-level VGG features for $\alpha \in \mathbb{R}$5 and $\alpha \in \mathbb{R}$6, and $\alpha \in \mathbb{R}$7 is cosine distance, the Instance-level Style Transfer (IST) loss is

$\alpha \in \mathbb{R}$8

where $\alpha \in \mathbb{R}$9 is the number of pixels in $c$0. Optimization updates the 3DGS parameters $c$1 to minimize $c$2 (Xu et al., 4 Sep 2025).

This construction makes the transfer explicitly instance-aware. The manuscript states that direct fine-tuning of 3DGS on stylized key views, which do not provide pixel-level 3D consistency, produces blurriness and boundary artifacts such as degraded background edges and distorted truck bucket contours. By contrast, IST’s group-wise nearest-neighbor matching reduces blurriness and artifacts while preserving sharp contours and hierarchical instance separation (Xu et al., 4 Sep 2025).

The training objective is notably narrow. The paper defines $c$3 as the principal supervisory objective in Stage 2 and states that additional reconstruction losses, Gram-matrix style losses, instance-consistency losses, or edge-preservation losses are discussed only for evaluation purposes, not as training losses in the proposed method. A plausible implication is that the method relies heavily on the quality of the Stage 1 stylized key views and the identity alignment supplied by Gaussian Grouping.

5. Empirical profile, runtime, and ablations

The method is evaluated on LLFF and Tanks and Temples, using styles that include abstract, sketches, cartoons, fantasy art, oil paintings, and ink wash (Xu et al., 4 Sep 2025). Qualitatively, the paper reports superior large-scale semantic transfer and fine-grained brushstroke fidelity relative to ARF, StyleGaussian, and G-Style, as well as improved instance-level consistency across distant views.

Evaluation axis	SSGaussian	Reported comparisons
Short-range consistency	LPIPS 0.031, RMSE 0.028	ARF 0.049/0.041, StyleGaussian 0.036/0.030, G-Style 0.035/0.035
Long-range consistency	LPIPS 0.073, RMSE 0.068	ARF 0.128/0.082, StyleGaussian 0.077/0.071, G-Style 0.089/0.072
Content Loss	2.298	ARF 2.490, StyleGaussian 2.300, G-Style 2.467
Style Loss	3.091	ARF 3.184, StyleGaussian 5.297, G-Style 3.303
Runtime / FPS	20 mins total, 118 FPS	ARF 24 mins, 10 FPS; G-Style 31 mins, 110 FPS; StyleGaussian ~5 hours, 3 FPS

The reported runtime is approximately 1 minute for Stage 1 and 19 minutes for Stage 2 on a single NVIDIA RTX 3090 GPU, with real-time rendering at 118 FPS. The user study with 30 participants reports preference percentages of 36.7% for Structural Integrity, 53.3% for Style Similarity, and 56.7% for Visual Quality, each above the compared methods. The method is also compared to video style transfer approaches such as AnyV2V and UniVST, with better temporal or multi-view consistency and structure preservation when multi-view sequences are treated as videos (Xu et al., 4 Sep 2025).

The ablations isolate two decisions. First, CVSA is most effective when inserted in the last upsampling block. Second, IST is preferable to direct fine-tuning on stylized key views because direct pixel-space supervision from diffusion outputs introduces blur and boundary artifacts. The paper does not explicitly enumerate a limitations section, but it does state that strict pixel-level 3D consistency from 2D diffusion outputs remains difficult, that the method targets instance-level rather than pixel-level consistency, and that rare styles, extreme occlusions, exact key-view counts, and diffusion hyperparameters are not analyzed or prescribed (Xu et al., 4 Sep 2025).

6. Other uses of the label in Gaussian research

The broader Gaussian literature attaches several other meanings to SSGaussian or near-equivalent expansions.

In sparse/selective 3D Gaussian splatting, MaskGaussian treats each Gaussian as a probabilistic entity with a learnable existence variable, uses differentiable Gumbel-Softmax sampling, and applies masks only during blending so that even currently absent Gaussians continue to receive gradients. This formulation is explicitly described as turning sparse/selective 3D Gaussian splatting into a differentiable selection process and reports average pruning ratios of 62.4%, 67.7%, and 75.3% with only approximately 0.02 PSNR decline on Mip-NeRF360 and Tanks&Temples, together with speedups of 2.05×, 2.19×, and 3.16× relative to 3DGS (Liu et al., 2024).

In semantic-spectral Gaussian splatting, "SpectralGaussians" is presented as a semantic, multi-spectral extension of 3DGS and is described as directly instantiating an SSGaussian (Semantic + Spectral Gaussian Splatting) approach. Each Gaussian carries per-spectrum material parameters, normals, and per-spectrum identity encodings, with a physically based spectral shading model and semantic grouping used for rendering and editing. The framework reports average PSNR/SSIM/LPIPS of 38.46 / 0.980 / 0.044 on the SpectralNeRF synthetic dataset and uses a cross-spectral warm-start that improves average PSNR from 40.17 to 41.63 in ablation (Sinha et al., 2024).

In sparse-view super-resolution 3DGS, "S2Gaussian" states that SSGaussian is an informal shorthand sometimes used in the community for the same method. That framework reconstructs high-resolution scenes from sparse low-resolution views using a two-stage LR-to-HR pipeline, Gaussian Shuffle Split densification, blur-free inconsistency modeling, and 3D robust optimization. It reports PSNR 24.19 on Blender, 20.45 on LLFF, and 22.05 on Mip-NeRF 360 under the stated sparse-view $c$4 super-resolution settings (Wan et al., 6 Mar 2025).

In edge reconstruction, "SGCR" states that if SSGaussian is intended to mean a spherical or isotropic splat primitive, it corresponds directly to the paper’s Spherical Gaussian representation: a 3D isotropic Gaussian with fixed radius and minimal attributes. SGCR then extracts rational Bézier curves from aligned Spherical Gaussians and reports on ABC-NEF a Chamfer Distance of 0.0280, F-score 0.9260, and IoU 0.8615 (Yang et al., 7 May 2025).

In shading-aware 3DGS, the GaussianShader exposition abbreviates a shading-enhanced 3D Gaussian Splatting formulation as SSGaussian, with per-Gaussian diffuse color, specular tint, roughness, residual color, and normals estimated from the shortest axis of each Gaussian. It reports an average PSNR of 31.94 dB on Shiny Blender versus 30.37 dB for standard 3DGS, with training time 0.58 h versus 23 h for Ref-NeRF and rendering at 97 FPS (Jiang et al., 2023).

Outside 3D graphics, the ICA paper uses SSGaussian for the Split Generalized Gaussian distribution, a family with different left and right scales and a shared shape parameter for heavy tails. In that setting the term does not refer to Gaussian splatting at all, but to a probability distribution used for maximum-likelihood independent component analysis (Spurek et al., 2018).

Taken together, these usages show that SSGaussian names either a specific stylization method (Xu et al., 4 Sep 2025) or a family of Gaussian-centered ideas whose common thread is not a single algorithm but the adaptation of Gaussian representations to a particular inductive bias: semantic structure, spectral physics, sparsity, super-resolution, isotropy, shading, or statistical asymmetry.