Mirror-3DGS: Reflection-Aware 3D Gaussian Splatting

Updated 3 March 2026

Mirror-3DGS is a reflection-aware technique that integrates planar mirror geometry into 3D Gaussian Splatting to prevent artifacts and ensure physical consistency.
It employs dual rendering with explicit mirror mask prediction to fuse direct and reflective views, enabling novel view synthesis and accurate scene reconstruction.
The framework optimizes rendering performance and scene compression using dedicated loss functions and geometric reasoning, achieving state-of-the-art metrics.

Mirror-3DGS refers to a family of methods and frameworks that explicitly incorporate planar mirror geometry and reflection modeling into 3D Gaussian Splatting (3DGS) pipelines for the purposes of novel view synthesis, scene reconstruction, or compression. Unlike conventional 3DGS, which yields incorrect geometry and photometry in reflective environments by misinterpreting mirror signals as direct scene content, Mirror-3DGS constructs mathematically explicit virtual mirror images—either by reflecting the scene’s Gaussian primitives or the camera itself—restoring physical fidelity and multi-view consistency in mirror regions. These reflection-aware techniques have enabled state-of-the-art results for real-time rendering, accurate scene editing, single- and multi-view geometry recovery, and substantial compression gains.

1. Mirror Geometry: Parameterization and Virtual Camera Construction

A planar mirror is defined in 3D by its surface normal $n \in \mathbb{R}^3$ and signed offset $d \in \mathbb{R}$ , yielding a plane equation

$P(x):\ n^\top x + d = 0.$

Given a point $x$ , its reflection $x'$ across $P$ is computed as

$x' = x - 2(n^\top x + d) n.$

The reflection can be expressed in homogeneous coordinates via a $4\times4$ matrix, enabling the transformation of both scene primitives and camera extrinsics.

For camera extrinsics $[R_\text{c}\ |\ t_\text{c}]$ , the corresponding mirrored camera (“virtual camera”) has \begin{align*} R_\text{mirrored} &= (I - 2n n^\top) R_\text{c}, \ t_\text{mirrored} &= (I - 2n n^\top) t_\text{c} + 2d n. \end{align*} This construction guarantees that rendering from the virtual camera is physically equivalent to observing the real scene via the mirror, thus creating multi-view consistency between direct and reflected rays (Meng et al., 2024, Wang et al., 2024, Wu et al., 24 Sep 2025, Liu et al., 2024).

2. Reflection-Aware Gaussian Splatting: Rendering and Fusion

Standard 3DGS represents a scene with $N$ anisotropic Gaussians $d \in \mathbb{R}$ 0. In the presence of mirrors, direct mapping of all image evidence to the scene cloud results in “ghost” Gaussians (“floaters”) behind the mirror and blurred or inconsistent mirror regions.

Mirror-3DGS overcomes this using reflection-aware dual rendering:

Each real Gaussian is reflected about the mirror plane to create a virtual Gaussian with parameters

$d \in \mathbb{R}$ 1

where $d \in \mathbb{R}$ 2.

The real and mirrored Gaussian clouds are rendered independently to produce front-view and mirror-view images via EWA splatting and alpha blending.
A per-pixel or per-Gaussian mirror mask $d \in \mathbb{R}$ 3 distinguishes mirror regions; the final output is fused as

$d \in \mathbb{R}$ 4

This framework achieves physically correct reflection synthesis while preserving the real-time differentiability and high performance of 3DGS (Meng et al., 2024, Liu et al., 2024).

3. Mirror Plane Estimation and Mask Prediction

Mirror-3DGS requires accurate estimation of the mirror plane and associated segmentation mask. Multiple strategies are employed:

Learnable per-Gaussian mirror probabilities $d \in \mathbb{R}$ 5 (derived from rendering residuals or explicit mask loss) are aggregated to produce a soft spatial mirror mask.
Centers of high $d \in \mathbb{R}$ 6 Gaussians are used to fit the mirror plane via RANSAC or least-squares (optionally refining with point normals back-projected from CNN/3DGS-inferred depths).
Mask prediction loss $d \in \mathbb{R}$ 7 is imposed, either supervised via ground-truth masks or with consistency to predicted mirror content.
In fully unsupervised or weakly annotated domains, geometric priors or self-supervision on spatial symmetry can be combined with photometric objectives (Meng et al., 2024, Wang et al., 2024, Wu et al., 24 Sep 2025, Liu et al., 2024, Guo et al., 23 Sep 2025).

4. Optimization and Loss Functions

Mirror-3DGS models are optimized through joint objectives that combine color, mask, geometry, and symmetry losses:

Photometric loss:

$d \in \mathbb{R}$ 8

ensuring both perceptual and structural alignment.

Mirror mask loss and, when applicable, depth supervision or mirror-plane regularization:

$d \in \mathbb{R}$ 9

Symmetry consistency losses:
- For matched real and mirrored Gaussians, enforce $P(x):\ n^\top x + d = 0.$ 0 (Guo et al., 23 Sep 2025).
- For virtual camera alignment, penalize pose differences: $P(x):\ n^\top x + d = 0.$ 1 (Wu et al., 24 Sep 2025).
Composite loss weights $P(x):\ n^\top x + d = 0.$ 2 are set by cross-validation or ablation.

Optimization proceeds in multiple phases: initial geometry/mask fitting, mirror plane estimation, dual-view rendering and fusion, and final refinement of Gaussian and pose parameters, typically utilizing Adam or L-BFGS for 15–60 minutes on a modern GPU, yielding real-time inference performance (Meng et al., 2024, Liu et al., 2024, Wang et al., 2024, Wu et al., 24 Sep 2025, Guo et al., 23 Sep 2025).

5. Empirical Results and Benchmarks

Mirror-3DGS achieves substantial qualitative and quantitative improvements in mirror-rich scenes:

Method	PSNR↑ (Synth)	SSIM↑ (Synth)	LPIPS↓ (Synth)	FPS↑	PSNR↑ (Real)	SSIM↑ (Real)	LPIPS↓ (Real)	FPS↑
Mirror-3DGS	37.89–39.87	0.97–0.979	0.01–0.038	120–172	23.01–29.14	0.77–0.874	0.08–0.291	96–128
Mirror-NeRF	27.33–38.08	0.94–0.958	0.05–0.028	0.55–0.71	25.04–25.25	0.86	0.06	0.07–0.55
3DGS vanilla	22.15–37.00	0.84–0.97	0.17–0.04	208–480	22.13–26.88	0.72–0.884	0.24–0.332	209–480

Mirror-3DGS reconstructs physically valid mirror reflections, eliminates “floaters,” and recovers occluded geometry visible only via reflection.
Rendering rates are 1–2 orders of magnitude higher than NeRF-based approaches.
On synthetic datasets, Mirror-3DGS outperforms all prior methods in F1, PSNR, SSIM, and LPIPS (Meng et al., 2024, Wang et al., 2024, Wu et al., 24 Sep 2025, Liu et al., 2024, Guo et al., 23 Sep 2025).
Real-world results show both visual quality and accurate mirror mask prediction, with convergence in 1–15 minutes.

6. Compression via Local Symmetry Exploitation

Mirror symmetry can also be exploited for compression of 3DGS scenes (Gupta et al., 17 Nov 2025):

SymGS introduces learnable, scene-specific mirror planes and reflects the set of surviving Gaussians about these planes, eliminating redundant primitives.
The pipeline iteratively detects strong planar symmetries by voting in a discretized normal–offset space, partitions and reflects Gaussian clusters, and then re-optimizes over plane parameters and primitives with standard photometric losses.
SymGS achieves up to $P(x):\ n^\top x + d = 0.$ 3 compression relative to the raw 3DGS baseline (and $P(x):\ n^\top x + d = 0.$ 4 over previous SOTA compression HAC) while preserving PSNR within $P(x):\ n^\top x + d = 0.$ 5.

This symmetry-driven approach demonstrates the value of explicit geometric reasoning in large-scale 3DGS memory reduction and provides a plug-and-play module compatible with generic 3DGS compression methods.

7. Limitations, Extensions, and Open Problems

The current Mirror-3DGS paradigm exhibits several important limitations and opportunities for future advancement:

Restriction to single, planar, and static mirrors: handling of curved, multi-planar, dynamic, or partially specular surfaces remains unresolved (Meng et al., 2024, Liu et al., 2024, Wang et al., 2024, Wu et al., 24 Sep 2025, Guo et al., 23 Sep 2025).
Dependence on ground-truth mirror masks or accurate per-Gaussian mirror attributes during training; robust, automated mirror detection in general scenes is an open challenge.
Incomplete handling of multi-bounce, occlusions, and volumetric/specular effects; mirror-to-mirror reflections and non-lambertian BRDFs are not modeled (Guo et al., 23 Sep 2025).
Potential quadratic computational cost in symmetry detection and voting (mitigated by heuristic clustering and CUDA acceleration) (Gupta et al., 17 Nov 2025).
Extensions of the framework to learnable multi-symmetry families (including rotation or translation symmetries), end-to-end hypothesis generation, and dynamic or video scenes have been suggested but not yet fully realized.

Future directions include integrating robust self-supervised or CNN-based mirror detection, leveraging higher-order symmetry groups (SO(3), translation), adding multi-modal inputs (e.g., LiDAR and RGB via 3DRef (Zhao et al., 2024)), hybrid neuron-explicit models for speed and memory, and joint modeling of specular/glossy/diffuse interactions.

References:

(Meng et al., 2024) Mirror-3DGS: Incorporating Mirror Reflections into 3D Gaussian Splatting
(Liu et al., 2024) MirrorGaussian: Reflecting 3D Gaussians for Reconstructing Mirror Reflections
(Wang et al., 2024) Gaussian Splatting in Mirrors: Reflection-Aware Rendering via Virtual Camera Optimization
(Guo et al., 23 Sep 2025) Seeing Through Reflections: Advancing 3D Scene Reconstruction in Mirror-Containing Environments with Gaussian Splatting
(Wu et al., 24 Sep 2025) Reflect3r: Single-View 3D Stereo Reconstruction Aided by Mirror Reflections
(Gupta et al., 17 Nov 2025) SymGS : Leveraging Local Symmetries for 3D Gaussian Splatting Compression
(Zhao et al., 2024) 3DRef: 3D Dataset and Benchmark for Reflection Detection in RGB and Lidar Data