Constrained Gaussian Splatting

Updated 25 April 2026

Constrained Gaussian Splatting is a method that regularizes 3D/4D Gaussian representations by embedding explicit constraints to improve surface fidelity, efficiency, and memory usage.
It employs geometric constraints such as SDF projections, covariance bounding, and curvature alignment to achieve improved rendering quality, with metrics showing over 1 dB PSNR gains and significant memory reductions.
The approach integrates capacity constraints through budget optimization and dynamic Gaussian management, enabling scalable and resource-aware deployment in both static and dynamic scenes.

Constrained Gaussian Splatting is a class of methods that seek to regularize, structure, or otherwise constrain the parameters, geometry, or capacity of 3D (and 4D) Gaussian splatting representations to address fundamental limitations in standard, unconstrained approaches. By embedding explicit constraints—geometric, capacity, or optimization-based—within the Gaussian splatting pipeline, these methods enforce priors such as surface fidelity, regularity, compactness, and resource awareness, enabling improved rendering quality, accurate geometry recovery, and tractable deployment in practical scenarios. Below, key families of constraint mechanisms, representative algorithms, and their implications are synthesized from recent research.

1. Geometric Constraints via Signed Distance Fields

A recent advance introduces geometric constraints by linking each Gaussian to a sampled, discretized signed distance field (SDF), which encodes the (signed) proximity of every Gaussian center $\mu_i$ to the implicit surface. This alleviates the historical incompatibility between pure splat-based methods and strong surface-based geometry priors. In this paradigm, each primitive carries an additional scalar $s_i$ denoting $f(\mu_i)$ , where $f$ is the unknown continuous SDF. The SDF sample is tied directly to the Gaussian's opacity via a learnable SDF-to-opacity transformation,

$o_i = \mathcal{T}_\gamma(s_i) = \frac{4\exp(-\gamma s_i)}{(1 + \exp(-\gamma s_i))^2},$

with $\gamma$ as an annealed sharpness parameter. This "bell-shaped" mapping concentrates opacity near the learned surface (the $s_i=0$ set), and decays smoothly away, thus anchoring the radiance field to the zero level set of the SDF while preserving differentiability.

To enforce that the discrete SDF approximates a legitimate signed distance field, a projection-based consistency loss is used: each Gaussian is orthogonally projected to the estimated surface,

$\mu^i_{\rm proj} = \mu_i - s_i n_i,$

with $n_i$ the local normal direction, and the depth of this projected point is compared to the composited depth by splatting. The per-primitive loss,

$\mathcal{L}_{\rm p} = \frac{1}{N}\sum_{i:\,\epsilon_i \le \varepsilon} |\hat D^i_{\rm agg} - \hat D^i_{\rm proj}|,$

aligns the aggregate field with the discrete SDF projections, acting as a surrogate for Eikonal regularization in the absence of a continuous field.

A further stabilization employs a median loss on $s_i$ 0 to ensure the opacity mapping dynamically tracks the evolving SDF distribution. Empirically, this discretized SDF constraint improves relighting quality, surface recovery, and memory usage compared to methods relying on external SDF networks, with mean PSNR increasing by over 1 dB and memory reduced by $s_i$ 1 on standard benchmarks (Zhu et al., 21 Jul 2025).

2. Explicit Geometric and Statistical Constraints on Gaussian Parameters

Several works regularize the geometry and statistics of Gaussians directly, to ensure plausible, efficient, and physically meaningful representations.

Covariance Bounding (Isotropy): Micro-splatting introduces a loss penalizing the trace of each Gaussian covariance,

$s_i$ 2

where $s_i$ 3 is a user-defined threshold. This constrains the spatial support, prevents excessive blur, and enforces near-isotropy while maintaining fidelity (Lee et al., 8 Apr 2025).

Principal Curvature and Normal Alignment: GeoSplat initializes and regularizes Gaussians based on differential geometric quantities—principal curvatures $s_i$ 4 and normal vectors $s_i$ 5—estimated via robust, local manifold or varifold analysis. The scale and rotation matrices are aligned with local geometry, and shape regularizers prevent "needle" artifacts while promoting surface coverage. The optimization objective is augmented as

$s_i$ 6

enforcing second-order smoothness and orientation consistency (Li et al., 5 Sep 2025).

Flatness and Symmetry Constraints: In driving scenes, certain background classes (e.g., road or sky) are constrained with flatness priors by penalizing roll, pitch, and vertical scale components, while a reflected Gaussian consistency loss supervises unseen sides of symmetric objects via parity transformations of positions, rotations, and SH coefficients (Khan et al., 2024).
Patch and Gradient-based Attention: Adaptive regularization focuses computation and supervision on regions under-reconstructed or with high-frequency content, for example using dynamic pixel-wise weighting and multi-sample anti-aliasing to counteract aliasing and oversmoothing (Zhou et al., 14 Aug 2025).

3. Capacity, Memory, and Budget Constraints

With practical deployments in resource-limited environments, constraining the number and memory usage of Gaussians during both training and inference is critical.

Budget-Constrained Optimization: Constrained Dynamic Gaussian Splatting (CDGS) embeds Gaussian count directly into the optimization via a differentiable proxy $s_i$ 7, with gates $s_i$ 8, and penalizes deviation from a user-specified $s_i$ 9:

$f(\mu_i)$ 0

the overall loss includes rendering, budget, and regularization terms. A unified importance score fuses geometric, motion, and perceptual metrics for adaptive pruning/densification and static/dynamic allocation (Zheng et al., 3 Feb 2026).

Memory-Bounded Growth and Pruning: "Gaussians on a Diet" interleaves Gaussian growth (by splitting under-fit primitives based on hybrid gradients) and pruning (removing Gaussians with lowest ray-based importance) on a schedule, always enforcing $f(\mu_i)$ 1 for a fixed Gaussian budget matched to device memory. Sub-pixel compensation adds new Gaussians only where reconstruction error warrants. This approach attains state-of-the-art rendering quality at $f(\mu_i)$ 2 lower Gaussian count and 80% lower memory usage compared to unconstrained 3DGS (Zhang et al., 21 Apr 2026).
Densification and Simplification: Mini-Splatting achieves competitive rendering with orders of magnitude fewer Gaussians by alternating refinement (e.g., blur split, depth reinitialization) and pruning (e.g., intersection preserving, importance-weighted sampling) within a specified cardinality constraint (Fang et al., 2024).

4. Subdivision, Decomposition, and Hierarchical Constraints

For scalability to large or complex scenes, constrained Gaussian splatting strategies embrace spatial subdivision or hierarchical refinement:

Block-wise Training and Merging: GaussianFocus splits the input space into axis-aligned blocks, trains Gaussian sets for each independently with hard scale constraints, and merges for the final scene. This approach realizes efficient large-scene reconstruction with reduced redundancy and GPU memory load (Huang et al., 22 Mar 2025).
Tetrahedral and Mesh Constraints: In avatar reconstruction, Tetrahedron-constrained Gaussian Splatting initializes Gaussians on mesh surfaces derived from Marching Tetrahedra on SDF-predicted tetrahedral grids. Selective freezing and adaptation of tetrahedra enable localized deformation and strict region preservation for editable assets (Liu et al., 29 Apr 2025).
Progressive Refinement for Surgery: SurGSplat employs sequential global and local optimization, geometric feature matching, and patch correlation to enforce frame-to-frame geometric consistency and maintain depth accuracy, targeting fine vascular structures while integrating new data progressively (Zheng et al., 6 Jun 2025).

5. Multiview and Pose-Constrained Optimization

Scene geometry can be regularized by incorporating constraints from multiple views and leveraging noisy external estimates:

Multiview Regularization: Explicit constraints on rendered/expected depth and normal consistency are induced using prior depth maps from MVS algorithms, triplet-based multiview sampling, and uncertainty-weighted losses. MVS-guided initialization aligns Gaussian positions for better geometric coverage (Kim et al., 16 Jun 2025).
Pose and Epipolar Constraints: Simultaneous camera pose, intrinsic, and Gaussian refinement from noisy LiDAR and coarsely-posed images is attained via log-barrier bounds on parameters, sensitivity-conditioned gradient preconditioning, epipolar and reprojection errors, without traditional SfM or bundle adjustment (Peng et al., 12 Apr 2025).

6. Frequency and Sampling Constraints

To mitigate aliasing and preserve stability under scale changes or viewpoint variation:

Smoothing Filters and Mipmaps: Mip-Splatting applies an analytic 3D low-pass filter to ensure each Gaussian's spatial support is band-limited by the maximal camera sampling rate (determined from depth and focal length). At render time, a 2D mip filter simulates box-filter pixel integration to suppress aliasing, preserving crisp geometry and colors across arbitrary zoom factors and camera scales (Yu et al., 2023).

7. Dynamic and Temporal Constraints

In dynamic or 4D splatting, temporal coherence and capacity constraints are achieved by:

State-Space and Wasserstein Constraints: Per-Gaussian dynamics are modeled with state-space filters, with parameters evolved via log/exp maps in Wasserstein space. A Kalman-like state consistency filter fuses predictions and observations, while Wasserstein regularization penalizes rapid parameter drift, yielding smooth, physically plausible motion (Deng et al., 2024).
Adaptive Static–Dynamic Split: CDGS adaptively allocates budget between static and dynamic Gaussians based on empirical motion statistics, rebudgeting throughout training to prioritize time-varying regions (Zheng et al., 3 Feb 2026).

These numerous constraint types reflect the versatility of constrained Gaussian splatting as a research program: leveraging geometry, statistics, sampling theory, optimization, and resource awareness to enforce structure and efficiency within a fundamentally discrete, splat-based representation paradigm. Recent ablation studies across multiple methods consistently show that well-designed constraint mechanisms yield substantial improvements in rendering quality, geometric accuracy, robustness to noise or underconstrained inputs, and ability to operate at scale or under strict hardware limits (Zhu et al., 21 Jul 2025, Huang et al., 22 Mar 2025, Lee et al., 8 Apr 2025, Zheng et al., 3 Feb 2026, Zhang et al., 21 Apr 2026, Li et al., 5 Sep 2025, Kim et al., 16 Jun 2025, Yu et al., 2023).