Gaussian Splatting Holography (GSH)

• Gaussian Splatting Holography is a reconstruction method that models the complex object field using a low-dimensional parametric Gaussian mixture.
• It suppresses twin-image artifacts by reducing the degrees of freedom, reformulating an underdetermined phase retrieval into a well-posed problem.
• The approach employs efficient analytical propagation and scalable numerical optimization, achieving high fidelity metrics without heavy priors.

Searching arXiv for the cited papers to ground the article in current literature. Gaussian Splatting Holography (GSH) is a holographic reconstruction framework for single-shot in-line holography that represents the complex object field as a compressed set of parametric Gaussians rather than a pixel-wise complex image. In the formulation reported in "Gaussian splatting holography" (Zhang et al., 25 Sep 2025), this compression reduces the number of unknown parameters by a maximum of 15 folds, reframes phase retrieval from an underdetermined problem to a well-posed one, and suppresses the twin-image artifacts that arise from the Hermitian symmetry of the forward model. The method is positioned as an alternative to approaches that rely on additional measurements, explicit physical constraints, or deep priors, and it is reported to achieve constraint-free recovery with an average peak signal-to-noise ratio equal to 26 dB and structure similarity equal to 0.8, with further gains when combined with total variation (Zhang et al., 25 Sep 2025).

1. Problem setting in in-line holography

In-line holography records the interference between an object wave and a reference wave in a lens-free geometry, yielding a single intensity image

$I(x,y)=|E(x,y;z)|^2.$

The reconstruction task is to recover the complex field $E(x,y;0)$, including both amplitude and phase, from that single 2D intensity measurement. In the formulation described for GSH, this is intrinsically ill-posed because a pixel-wise complex field on an $M\times N$ grid contains $2\cdot M\cdot N$ real unknowns, whereas the measurement supplies only $M\cdot N$ intensity values (Zhang et al., 25 Sep 2025).

A central difficulty is the twin image. Because the forward model exhibits Hermitian symmetry, phase retrieval admits a conjugate solution that appears as a spurious out-of-focus image superimposed on the desired reconstruction. Classical phase-recovery procedures such as Gerchberg–Saxton and Alternating Projections are therefore troubled by twin-image artifacts. Existing remedies typically introduce additional measurements, including phase diversity, multi-distance holograms, or reference beams, or impose strong priors such as total variation or deep networks. GSH departs from those strategies by reducing the dimensionality of the unknown wavefield itself.

This reframing is significant because it changes the mathematical character of the inverse problem. Rather than attempting to infer a dense complex field with more unknowns than data, GSH fits a low-dimensional parametric representation whose degrees of freedom can be made smaller than the number of measurements. This suggests that twin-image suppression is not treated primarily as a post hoc denoising problem, but as a consequence of model design.

2. Gaussian parameterization of the complex field

GSH represents the object-plane field

$E(x,y;0)=u(x,y)=a(x,y)+j\,b(x,y)$

as two independent Gaussian mixtures, one for the real part $a$ and one for the imaginary part $b$ (Zhang et al., 25 Sep 2025). For the real part, the representation is

$a(x,y)=\sum_{l=1}^L v_l\,\sigma_l\,G_l(x,y),$

with Gaussian basis functions

$$G_l(x,y)=\exp\!\left[-\frac{1}{2}\left((x-x_{c,l},y-y_{c,l})^T\Sigma_l^{-1}(x-x_{c,l},y-y_{c,l})\right)\right].$$

Each Gaussian basis has 7 parameters: center $E(x,y;0)$0, covariance $E(x,y;0)$1 parameterized by scale $E(x,y;0)$2 and rotation, and weight $E(x,y;0)$3. The imaginary part uses a separate but identical set of $E(x,y;0)$4 Gaussians with its own parameters. The total number of real unknowns is therefore $E(x,y;0)$5, in contrast to the $E(x,y;0)$6 parameters of a pixel-wise model. In practice, $E(x,y;0)$7, and the reported implementation attains up to 15-fold compression with negligible loss of detail (Zhang et al., 25 Sep 2025).

The representation has two direct consequences. First, it is a compressive model of the optical field rather than merely an acceleration device. Second, because each Gaussian has a spatially slow-varying profile, the basis itself disfavors noisy twin-image backgrounds. The paper explicitly associates the sharpness of the recovered patterns with the smooth envelope of the Gaussian atoms. A plausible implication is that the basis serves simultaneously as a dimensionality reduction mechanism and as an implicit regularizer.

3. Forward model and optimization objective

From the Gaussian-parameterized field $E(x,y;0)$8, GSH predicts the hologram through angular-spectrum propagation from the object plane to the sensor plane:

$E(x,y;0)$9

with

$M\times N$0

followed by inverse Fourier transformation to obtain $M\times N$1 and the predicted intensity

$M\times N$2

(Zhang et al., 25 Sep 2025).

The efficiency claim in GSH is tied to the observation that each Gaussian propagates analytically as a possibly rotated elliptical Gaussian beam, making the forward operations differentiable and efficient. Optimization is performed over the Gaussian parameters

$M\times N$3

by minimizing a data-fidelity term in the amplitude domain,

$M\times N$4

An optional total-variation regularizer may be added,

$M\times N$5

yielding

$M\times N$6

This objective is noteworthy for what it omits as well as for what it includes. The core formulation is described as constraint-free recovery for in-line holography, and total variation is explicitly optional rather than fundamental. That distinguishes GSH from reconstruction pipelines in which external constraints are indispensable for suppressing phase ambiguity.

4. Numerical optimization and the mechanism of twin-image suppression

The numerical procedure reported for GSH initializes Gaussian centers uniformly or according to a coarse estimate, starts covariances $M\times N$7 large, and initializes weights $M\times N$8 with small random values. Optimization then proceeds with Adam or Lion using an initial learning rate, exemplified as $M\times N$9, decayed by 0.5 every 400 iterations, for 3,000–5,000 steps. Convergence is monitored through stabilization of $2\cdot M\cdot N$0, PSNR, or SSIM on a validation split of hologram pixels. For 256×256 holograms with a few thousand Gaussians, training is reported to converge within seconds on a modern GPU (Zhang et al., 25 Sep 2025).

The twin-image suppression mechanism is framed explicitly in terms of null-space reduction. By enforcing $2\cdot M\cdot N$1, the degrees of freedom that ordinarily support the conjugate twin solution are said to vanish. The reported argument is that conjugate solutions would require rapid oscillations or high-frequency artifacts among Gaussians, which the low-dimensional basis cannot represent. Quantitatively, once $2\cdot M\cdot N$2, the number of unknown parameters falls below the number of measurements, ensuring a unique least-squares solution up to common global phase (Zhang et al., 25 Sep 2025).

This account clarifies an important conceptual point. GSH does not primarily remove twin images after reconstruction; instead, it seeks a parameterization in which twin-image-bearing solutions are excluded from the admissible model class. A common misconception is to equate this with generic sparsity regularization. The description provided for GSH is more specific: the suppression is tied to a Gaussian-mixture wavefield model whose smooth basis functions and reduced parameter count jointly alter the identifiability of the inverse problem.

5. Empirical results and operating regimes

The reported simulation study uses a 256×256 pure-phase “cameraman” target. A pixel-wise Alternating Projections baseline with pure-phase constraint has $2\cdot M\cdot N$3 unknowns and exhibits heavy twin artifacts, with PSNR approximately 18 dB and SSIM approximately 0.45. GSH with $2\cdot M\cdot N$4 Gaussians, corresponding to $2\cdot M\cdot N$5 unknowns and 2.6× compression, yields a twin-free reconstruction with PSNR approximately 26 dB and SSIM approximately 0.80 without any regularization. The GSH+TV variant reaches PSNR approximately 31 dB and SSIM approximately 0.91 at the same number of Gaussians (Zhang et al., 25 Sep 2025).

The paper further reports a PSNR-versus-compression analysis across hologram sizes, comparing AP, AP+TV, GSH, and GSH+TV. In that analysis, GSH dominates over a 4×–15× compression range. For a quantitative phase target of size 945×517 pixels, a reconstruction with 53,000 Gaussians, corresponding to 1.3× compression, is compared against differentiable holography ($2\cdot M\cdot N$6); GSH is reported to yield sharper feature edges and higher contrast, while line traces across L1/L2 verify quantitative phase accuracy with error below 0.05 rad. In a USAF resolution test, the method resolves group 7 element 4, corresponding to 2.76 $2\cdot M\cdot N$7m lines and 181 lp/mm. In whole-slide lens-free imaging at 3840×2160 resolution with 450,000 Gaussians, the reconstruction shows clear histological features of a Ficus slice with twin-free background (Zhang et al., 25 Sep 2025).

These results position GSH as a method that spans both moderate-scale simulation and large-format experimental imaging. The progression from 256×256 synthetic phase targets to 3840×2160 whole-slide imaging suggests that the Gaussian parameterization is intended not only for inverse-problem conditioning but also for practical scaling. At the same time, the reported compression factor is not uniform across settings; it varies with scene complexity and target resolution.

6. Relation to Gaussian-based computer-generated holography

The phrase “Gaussian splatting holography” can be confused with Gaussian-based computer-generated holography for display. The literature cited here includes a separate line of work, "Gaussian Wave Splatting for Computer-Generated Holography" (Choi et al., 10 May 2025), which starts from Gaussian scene representations optimized from photographs for novel-view synthesis and derives a 2D Gaussian-to-hologram transform that supports occlusions and alpha blending. That framework also includes an efficient approximation in the Fourier domain that is easily parallelizable and implemented using custom CUDA kernels. Its objective is hologram synthesis for photorealistic scenes and next-generation holographic displays rather than twin-image-suppressed reconstruction from a single in-line hologram.

A further extension, "Random-phase Gaussian Wave Splatting for Computer-generated Holography" (Chao et al., 24 Aug 2025), addresses holographic near-eye displays. Choi et al. describe Gaussian Wave Splatting (GWS) as a forward rendering paradigm based on 3D Gaussian primitives and then introduce random-phase GWS (GWS-RP) to improve bandwidth utilization, increase eyebox size, reconstruct accurate defocus blur and parallax, and support time-multiplexed rendering for speckle suppression. In that setting, the key technical issues are wavefront compositing, alpha-blending under random phase, and perceptually faithful 3D display rather than ill-posed single-shot phase retrieval.

Method	Problem setting	Core characterization
GSH	Single-shot in-line holographic reconstruction	Compresses the unknown complex field into Gaussian mixtures; targets twin-image suppression (Zhang et al., 25 Sep 2025)
GWS	Computer-generated holography from Gaussian scene representations	Derives a 2D Gaussian-to-hologram transform with occlusions and alpha blending (Choi et al., 10 May 2025)
GWS-RP	Near-eye display CGH with random phase	Uses random-phase Gaussian primitives, intensity-domain blending, and time-multiplexing (Chao et al., 24 Aug 2025)

This distinction matters because the shared vocabulary of “Gaussian splatting” does not imply a shared inverse problem. In GSH, Gaussians parameterize the unknown wavefield to make phase retrieval well-posed. In GWS and GWS-RP, Gaussians parameterize scene content for hologram rendering. The methods are therefore related by representation strategy, but they operate on different physical tasks and optimization targets.

7. Limitations, extensions, and research directions

The reported limitations of GSH are primarily tied to the number of Gaussian components. The choice of $2\cdot M\cdot N$8 trades compression against expressivity: too few Gaussians under-fit, whereas too many may over-fit and slow convergence. Adaptive densification strategies are identified as a way to grow Gaussians where needed. The current formulation also assumes single-shot in-line geometry, although extension to multi-distance or multi-wavelength setups is described as straightforward by adding additional measurement terms in $2\cdot M\cdot N$9. GPU memory and compute scale roughly linearly with $M\cdot N$0, and very large-scale 3D holography may require multi-GPU or sparsity-aware implementations (Zhang et al., 25 Sep 2025).

The same source identifies several extensions. One is to combine GSH with learned deep priors or plug-and-play regularizers for extreme low-light or highly sparse sampling. Another is to extend the compression principle to Fourier ptychographic or diffraction-tomography settings by treating the object volume as a 3D Gaussian cloud, with the stated goal of compressive 3D phase recovery without mechanical scanning. The method is also proposed as a template for other ill-posed inverse problems, including coherent diffractive imaging and phase-contrast microscopy, wherever a large-dimensional wavefield must be inferred from limited data (Zhang et al., 25 Sep 2025).

These proposed directions indicate that GSH is best understood as a broader modeling principle: replace dense wavefield unknowns with a structured Gaussian cloud whose parameter count is commensurate with the available data. This suggests a family resemblance to recent Gaussian-based holographic rendering methods, but the principal contribution of GSH remains its use of Gaussian splatting as an identifiability mechanism for in-line phase retrieval.