---
title: 3D Radiative Gaussians (RadGS)
url: https://www.emergentmind.com/topics/3d-radiative-gaussians-radgs
type: topic
---

# 3D Radiative Gaussians (RadGS)

to=arxiv_search.search ＿俺去也json  天天中彩票可以{"query":"\"Wideband RF Radiance Field Modeling Using Frequency-embedded 3D Gaussian Splatting\"","max_results":5,"sort_by":"submittedDate"}
to=arxiv_search  彩神争霸如何json  红鼎{"query":"\"Wideband RF Radiance Field Modeling Using Frequency-embedded 3D Gaussian Splatting\"","max_results":5}
to=arxiv.search  彩神争霸快json ിക്കും{"query":"\"Wideband RF Radiance Field Modeling Using Frequency-embedded 3D Gaussian Splatting\"","max_results":5}
to=search_arxiv 񹚊ppjson ിക്കും{"query":"3D Radiative Gaussians","max_results":10}
3D Radiative Gaussians (RadGS) is best understood as a family of 3D Gaussian-based scene representations in which each anisotropic Gaussian is treated as a carrier of radiative properties rather than merely a colored splat. Depending on the sensing or rendering regime, a Gaussian may encode RF signal intensity and attenuation, X-ray absorption density, opacity and view-dependent color, radiance-transfer coefficients, or BRDF-related quantities, and image formation proceeds by ordered compositing, Beer–Lambert accumulation, exact ray integration, or stochastic Monte Carlo transport through Gaussian volumes. Recent work instantiates this idea in wideband RF radiance-field modeling, inverse rendering with radiance transfer and reflection, sparse-view tomography, stochastic path tracing of transparent Gaussian clouds, exact volumetric rendering for generic cameras, and vertebral CT/X-ray registration [2505.20714, 2507.07733, 2405.20693, 2504.06598, 2505.24053, 2508.21154].

## 1. Conceptual scope and defining attributes

Taken together, the cited works suggest that RadGS is not a single canonical algorithm but a recurring design pattern: retain the explicit anisotropic 3D Gaussian parameterization of 3D Gaussian Splatting (3DGS), then replace purely heuristic image-space appearance with radiative quantities tied to a physical forward model.

| Instantiation | Domain | Per-Gaussian radiative attributes |
|---|---|---|
| Wideband RF 3DGS [2505.20714] | RF radiance fields | attenuation factor $\delta(G)$ and RF signal intensity $\mathrm{Sig}(G)$ |
| RTR-GS [2507.07733] | inverse rendering | transfer coefficients, transfer features, reflection attributes, BRDF parameters |
| R$^2$-Gaussian / RadGS-Reg [2405.20693], [2508.21154] | X-ray CT and registration | attenuation density $\rho_i$ or absorption/opacity $\sigma_i$, mixture weight $w_i$ |
| 3DGEER / stochastic ray tracing [2505.24053], [2504.06598] | exact volumetric rendering and path tracing | density or opacity scale, color or radiative response along rays |

This perspective differs from standard 3DGS in a precise way. Standard 3DGS represents a scene by anisotropic 3D Gaussian kernels with mean, covariance, color, and opacity, then projects them to 2D Gaussians and alpha-composites them in front-to-back order. RadGS formulations keep the Gaussian primitive but reinterpret what the primitive stores and how it is accumulated. In X-ray settings, the Gaussian carries attenuation density and must obey Beer–Lambert line integrals rather than RGB alpha blending. In RF settings, the Gaussian acts as a secondary EM wave source whose contribution is attenuated by earlier Gaussians along the line of sight. In inverse rendering, the Gaussian stores enough directional and material information to evaluate radiance transfer under shared illumination.

A useful consequence is domain portability. The same primitive type can be adapted to hemispherical RF power-angular spectra, detector-plane X-ray measurements, screen-space G-buffers for deferred reflection, or exact ray-space integration under pinhole and fisheye cameras. This suggests that the central abstraction in RadGS is not the rasterizer alone, but the pairing of explicit Gaussian geometry with a transport law appropriate to the sensing modality.

## 2. Mathematical structure of radiative Gaussian rendering

A common starting point is the anisotropic 3D Gaussian kernel. In tomographic form, one explicit parameterization is
$$
G_i^3(x \mid \rho_i, p_i, \Sigma_i)=\rho_i \exp\!\left(-\frac{1}{2}(x-p_i)^T \Sigma_i^{-1}(x-p_i)\right),
$$
with a global attenuation field
$$
\sigma(x)=\sum_{i=1}^{M} G_i^3(x \mid \rho_i, p_i, \Sigma_i).
$$
For X-ray CT, the physically relevant forward model is Beer–Lambert:
$$
I'(r)=I_0 \exp\!\left(-\int \sigma(r(t))\,dt\right), \qquad
I(r)=\log I_0-\log I'(r)=\int \sigma(r(t))\,dt.
$$
R$^2$-Gaussian emphasizes that this makes the log-domain measurement additive in line integrals, not alpha composited [2405.20693].

Exact ray integration can also be written directly. For a ray $r(t)=o+t\,d$, 3DGEER derives the finite-segment integral
$$
I(t_n,t_f)=\rho \sqrt{\frac{\pi}{2A}} \exp\!\left(-\frac{1}{2}\left(c-\frac{b^2}{A}\right)\right)
\left[
\operatorname{erf}\!\left(\frac{A t_f-b}{\sqrt{2A}}\right)-
\operatorname{erf}\!\left(\frac{A t_n-b}{\sqrt{2A}}\right)
\right],
$$
with
$$
A=d^T\Sigma^{-1}d,\qquad
b=d^T\Sigma^{-1}(o-\mu),\qquad
c=(o-\mu)^T\Sigma^{-1}(o-\mu).
$$
In its canonical infinite-range form, the same paper uses
$$
T(o,d)=\sigma \exp\!\left(-\frac{1}{2}D^2\right),
$$
where $D^2$ is the squared perpendicular Mahalanobis distance from the transformed ray to the Gaussian center [2505.24053].

Surface-style RadGS uses front-to-back compositing. A standard form appearing in transparent Gaussian ray tracing is
$$
T_0=1,\qquad
T_i=\prod_{j=1}^{i-1}(1-\alpha_j),\qquad
C=\sum_{i=1}^{M} T_{i-1}\alpha_i c_i.
$$
The same algebra underlies stochastic Bernoulli thinning, where each fractional opacity is replaced by a binary random event but the estimator remains unbiased in expectation [2504.06598].

RF RadGS introduces an analogous but modality-specific operator. After projection to the receiver hemisphere, wideband RF rendering uses
$$
I(p)=\sum_{i=1}^{N}\left(\prod_{j=1}^{i-1}\big(\delta_o(G_j)+\delta_f(G_j)\big)\right)\,\mathrm{Sig}(G_i),
$$
where $\delta$ plays the role of attenuation contributed by earlier Gaussians and $\mathrm{Sig}$ is the per-Gaussian RF emission or intensity [2505.20714].

A common misconception is to treat all RadGS variants as ordinary alpha splatting with renamed attributes. The literature instead shows several distinct transport regimes: additive accumulation in X-ray log space, multiplicative transmittance with emission for optical rendering, attenuation-modulated RF composition for PAS synthesis, and exact line integration in ray space. The shared idea is the Gaussian primitive; the rendering law is domain dependent.

## 3. Wideband RF radiance fields

In "Wideband RF Radiance Field Modeling Using Frequency-embedded 3D Gaussian Splatting," 3D Gaussians represent secondary EM wave sources created when incident waves interact with scene surfaces and edges under the Huygens-Fresnel principle. Each Gaussian carries spatial parameters as in standard 3DGS together with an attenuation factor $\delta(G)$ and an RF signal intensity $\mathrm{Sig}(G)$, both dependent on frequency and transmitter position. The paper formalizes this with an EM feature network
$$
F_\Theta:\; (P(G),P(TX),freq)\rightarrow (\delta(G),\mathrm{Sig}(G)),
$$
split into an attenuation module
$$
F_{\Theta1}:\; (P(G),P(TX),freq)\rightarrow (h(G),\delta(G))
$$
and a radiance module
$$
F_{\Theta2}:\; (h(G),freq)\rightarrow \mathrm{Sig}(G),
$$
with positional encoding applied to $P(G)$, $P(TX)$, and $freq$ [2505.20714].

The sensing target is the Power Angular Spectrum (PAS), defined as a hemispherical angular power image around the receiver with azimuth $\phi \in [-\pi,\pi]$ and zenith $\theta \in [0,\pi/2]$, discretized to $H\times W = 90\times 360$ bins. A 3D point $P=(x,y,z)$ with $z\ge 0$ is mapped by
$$
\phi=\operatorname{atan2}(x,y),\qquad
\theta=\arccos\!\left(\frac{z}{\sqrt{x^2+y^2+z^2}}\right),
$$
$$
p_x=\frac{\phi+\pi}{\pi}W,\qquad
p_y=\frac{\theta}{\pi/2}H.
$$
After tile-based differentiable rasterization onto this hemisphere, PAS intensity is rendered by ordered accumulation.

The paper explicitly motivates frequency awareness from EM propagation. It relates propagation to frequency through path loss,
$$
pathloss = 20 \log_{10}\!\left(\frac{4\pi d f}{c}\right),
$$
and through reflection, transmission, absorption, diffraction, and frequency-dependent material parameters $\epsilon(f)$ and $\mu(f)$. This is used structurally rather than through an explicit physics loss: $\delta(G)$ and $\mathrm{Sig}(G)$ are functions of frequency and transmitter position, and accumulation is multiplicative along the line of sight. A plausible implication is that the architecture encodes cross-band regularities without constraining the model to a fixed closed-form EM solver.

Training uses
$$
\mathcal{L}=(1-\lambda)\mathcal{L}_1+\lambda \mathcal{L}_{SSIM},\qquad \lambda=0.2,
$$
with end-to-end SGD over Gaussian spatial parameters, base attenuation $\delta_o$, and EM feature-network parameters. Adaptive density control clones or splits Gaussians with large position gradients and removes Gaussians whose attenuation falls below a threshold $\epsilon_\delta$. The residual formulation $\delta_o+\delta_f$ is retained to stabilize co-optimization with density control.

The experimental setup comprises 6 indoor environments, 50,000 PAS samples, and frequencies from 1 to 100 GHz. Ground-truth PAS is generated with NVIDIA Sionna using conventional beamforming on a $4\times 4$ receiver array, with an omnidirectional single-antenna transmitter and approximately 900 transmitter positions per scene after removing inaccessible ones. Across 6 scenes and 10 frequencies, the proposed wideband model achieves average SSIM 0.66 versus 0.56 for the single-frequency SOTA WRF-GS, corresponding to up to 17.8% improvement. In Scene 1 with 21 frequencies, average median SSIM reaches 0.72 versus 0.63. For cross-frequency evaluation, training on half of the 21 frequencies yields zero-shot average SSIM 0.70, described as only a 2.8% drop relative to using all frequencies. The authors note that the dataset is simulated rather than captured from real RF measurements, so closing the sim-to-real gap remains future work.

## 4. Radiance transfer, reflection, and inverse rendering

RTR-GS instantiates RadGS for inverse rendering by treating each Gaussian as a radiative carrier that stores geometric attributes and enough radiometric information to locally evaluate radiance transfer. The target quantity is the rendering equation
$$
L_o(x,\omega_o)=\int_\Omega f_r(x,\omega_i,\omega_o)\,L_i(x,\omega_i)\,(n(x)\cdot \omega_i)^+\,d\omega_i.
$$
Its forward branch approximates this integral through shared spherical-harmonic lighting and per-Gaussian transfer coefficients. Diffuse outgoing radiance is modeled as
$$
C_d \approx \rho_d \sum_{j=0}^{n^2} c_j c_j^t,
$$
while specular outgoing radiance uses a direction-dependent neural transfer
$$
C_s(\omega_o)\approx \rho_s \sum_{j=0}^{n^2} c_j c_j^t(\omega_o),\qquad
c_j^t(\omega_o)=G(f_t,o),
$$
where $G$ is a lightweight 3-layer MLP with 64 hidden units. The per-Gaussian radiance is $C_r=C_d+C_s(\omega_o)$, which is then splatted and alpha blended front-to-back as
$$
I_f=\sum_{i=1}^{N} f_i \alpha_i T_i,\qquad
T_i=\prod_{j=1}^{i-1}(1-\alpha_j)
$$
[2507.07733].

A central design choice is the separation of low-frequency and high-frequency appearance. RTR-GS keeps low-frequency radiance transfer in the forward branch, but performs high-frequency reflections after splatting in a deferred branch using a microfacet BRDF and image-based lighting with split-sum pre-integration. Reflection and radiance are combined by
$$
I_{rgb}=C_r\cdot (1-R_i)+C_{ref}\cdot R_i.
$$
The deferred branch operates on blended per-pixel G-buffers such as normal, albedo, roughness, and metallic, which the paper argues avoids the "lobe broadening" caused by evaluating sharp specular shading before Gaussian blending. This separation is presented as a remedy for floating artifacts and rippled normals caused by spherical-harmonic overfitting in vanilla 3DGS.

Per-Gaussian attributes include mean $\mu$, covariance $\Sigma$, opacity $\alpha$, normal $n$, diffuse and specular base colors $(\rho_d,\rho_s)$, diffuse PRT coefficients $c_j^t$, view-dependent transfer features $f_t$, reflection intensity $R_i$, reflection tint $R_t$, reflection roughness $R_r$, PBR parameters $(c,m,r)$, and diffuse-indirect light $L_{ind}$. Visibility is pre-baked into a $128^3$ voxel grid. Training uses a hybrid stage of 30k iterations followed by a decomposition stage of 10k iterations, with auxiliary priors including a light white prior, a metal–reflection consistency prior, a pseudo-normal consistency loss, and bilateral smoothness terms.

Empirically, RTR-GS reports real-time rendering at approximately 96 FPS for the hybrid branch and approximately 131 FPS for the PBR branch. On Shiny Blender, novel-view synthesis reaches 33.99 dB with PBR and 35.24 dB with the hybrid branch; on TensoIR, it reaches 39.17 dB and 41.39 dB respectively. The paper also reports competitive or improved normal estimation, including 3.09° mean angular error on Shiny Blender versus 3DGS-DR at 3.63°. Limitations noted by the authors include the infinite-distant lighting assumption, omission of higher-order indirect effects such as interreflections and caustics, exclusion of extremely complex transparent or participating media, and residual artifacts when reflections dominate and geometry is weakly observed.

## 5. Tomographic reconstruction and medical registration

For tomography, RadGS is reformulated around attenuation rather than emitted color. R$^2$-Gaussian models each kernel as a local density field and derives the correct detector-plane footprint by integrating the 3D Gaussian along the viewing axis in ray space. After projective mapping and local affine approximation, the 3D covariance $\Sigma'_i$ is reduced to a 2D covariance $\Sigma'_{i,2D}$, but with a crucial amplitude factor
$$
\mu_i=\sqrt{2\pi\cdot \frac{|\Sigma'_i|}{|\Sigma'_{i,2D}|}},
\qquad
\hat{\rho}_i=\mu_i \rho_i.
$$
The resulting detector measurement is a direct sum of 2D Gaussian footprints in log space, not alpha compositing. The paper identifies a previously unknown integration bias in standard 3DGS-style projection: keeping the same opacity while dropping the viewing dimension implicitly assumes $\mu_i=1$, so the recovered attenuation becomes view dependent. In experiments, the biased "X-3DGS" variant renders projections slightly better but reconstructs volumes about 7.6 dB worse on average and exhibits clear view-dependent inconsistencies in slices [2405.20693].

R$^2$-Gaussian couples this rectified rasterizer with a differentiable CUDA voxelizer for querying
$$
V(x)=\sum_i \rho_i \exp\!\left(-\frac{1}{2}(x-p_i)^T\Sigma_i^{-1}(x-p_i)\right)
$$
on a 3D grid, enabling total-variation regularization and dense volume output. Initialization uses a fast FDK reconstruction because SfM points are unavailable in CT; about 50k points are sampled after thresholding non-empty voxels. On 15 cases spanning human organs, animals and plants, and artificial objects, with sparse-view settings of 75, 50, and 25 views, the method achieves 38.88 dB / 0.959 SSIM at 75 views in about 8.3 minutes, 37.98 dB / 0.952 at 50 views, and 35.19 dB / 0.923 at 25 views. Compared with SAX-NeRF, average PSNR improves by 0.93 dB and SSIM by 0.014 while being about 12$\times$ faster.

RadGS-Reg extends the radiative Gaussian idea from reconstruction to vertebral-level CT/X-ray registration. Each Gaussian carries center $\mu_i$, anisotropic covariance $\Sigma_i$, absorption or opacity $\sigma_i$, emission or color term $c_i$, and mixture weight $w_i$, giving an attenuation field
$$
\rho(x)=\sum_{i=1}^{N} w_i\, N(x\mid \mu_i,\Sigma_i).
$$
For a ray $r(t)=o+t\,d$, the single-Gaussian optical thickness is approximated by
$$
\tau_i(o,d)\approx w_i \sqrt{\frac{2\pi}{a}}\,
\exp\!\left(-\frac{1}{2}\left(c-\frac{b^2}{a}\right)\right),
$$
with opacity
$$
\alpha_i=1-\exp(-\sigma_i \tau_i).
$$
The reconstruction module predicts vertebral Gaussians from two calibrated X-rays using an encoder, a Counterfactual Attention Learning (CAL) block, a Gaussian head, an X-ray rasterizer, and a density voxelizer. CAL is formulated through
$$
G^3_{\text{effect}}=E_{A\sim \gamma}[Y(A=A,X)-Y(do(A=\bar{A}),X)],
$$
and the reconstruction objective is
$$
L_{\text{CAL-rec}}=L_1(I_{\text{effect}},I_{\text{meas}})
+\lambda_1 SSIM(I_{\text{effect}},I_{\text{meas}})
+\lambda_2 TV(V_{\text{effect}})
+L_{\text{rec}}.
$$
A separate registration module predicts one rigid $SE(3)$ pose per vertebra with
$$
L_{\text{reg}}=NCC(V_{CT}\circ \hat{T},V_{rec})+SSIM(V_{CT}\circ \hat{T},V_{rec})+\lambda\, d_{geo}(T_{gt},\hat{T}).
$$

The training strategy has three stages: generic vertebral priors from VERSE’20 DRRs, pre-training on in-house real X-rays, and patient-specific DRRs generated from the target CT. On the reported in-house dataset, reconstruction reaches 94.51 $\pm$ 0.39% SSIM and 28.80 $\pm$ 0.33 dB PSNR, while registration reaches mean TRE 1.14 $\pm$ 1.01 mm, capture range 20–25 mm, success rate 93.33% for $mTRE<2$ mm, and runtime 0.82 s per case. The paper attributes part of the robustness to CAL focusing attention on vertebral regions in noisy biplanar X-rays [2508.21154].

## 6. Exact ray-space rendering, stochastic transport, and open limitations

Two recent directions broaden the RadGS rendering core beyond conventional screen-space splatting. First, 3DGEER replaces 2D projection approximations with exact volumetric rendering in ray space. It uses the exact Gaussian line integral and a fast canonical form
$$
T=\sigma \exp\!\left(-\frac{1}{2}D^2\right),\qquad
D^2=\frac{m_u^T m_u}{d_u^T d_u},
$$
with whitened ray variables $o_u=\Sigma^{-1/2}(o-\mu)$, $d_u=\Sigma^{-1/2}d$, and $m_u=o_u\times d_u$. To make exact rendering efficient, the method introduces a tight Particle Bounding Frustum (PBF) for each Gaussian and a Bipolar Equiangular Projection (BEAP) for equiangular ray sampling and supervision under arbitrary camera models. Reported performance includes about 327 FPS on MipNeRF360 and about 251 FPS on ScanNet++ fisheye using PBF association, while achieving 31.50 PSNR / 0.953 SSIM / 0.126 LPIPS on full-FoV ScanNet++ and 27.76 PSNR / 0.821 SSIM / 0.210 LPIPS at real-time speed on MipNeRF360 [2505.24053].

Second, "Stochastic Ray Tracing of Transparent 3D Gaussians" casts transparent Gaussian rendering as unbiased Monte Carlo transport. Along a ray, each Gaussian yields a 1D restriction with
$$
t_0=-\frac{d^T\Sigma_i^{-1}(x_0-\mu_i)}{d^T\Sigma_i^{-1}d},
\qquad
\sigma_t^2=\frac{1}{d^T\Sigma_i^{-1}d},
$$
and optical thickness
$$
\tau_i=\kappa_i \sqrt{2\pi}\,\sigma_t
\exp\!\left(-\frac{1}{2}\left(c-\frac{b^2}{a}\right)\right),
\qquad
\alpha_i=1-\exp(-\tau_i).
$$
Instead of sorting all intersections, the algorithm performs a single BVH traversal and accepts a Gaussian hit with Bernoulli probability $\alpha_i$. The nearest accepted hit is then distributed according to the exact front-to-back mixture weights $w_i=T_{i-1}\alpha_i$, so returning its color is unbiased in expectation. A simple $N$-sample extension averages multiple accepted hits from one traversal. The paper reports that the cost of rays through Gaussian assets is comparable to standard mesh-intersection rays, that the method minimizes shading time and avoids primitive sorting, and that most noise is gone by 64 spp while interactive rates are achieved on CPUs and GPUs [2504.06598].

Taken together, these results show that RadGS is not restricted to approximate screen-space rasterization. It can be exact in ray space, or unbiased in a path tracer, while still exploiting the analytic structure of anisotropic Gaussians. At the same time, the literature remains explicit about current boundaries. Wideband RF modeling is still evaluated on simulated Sionna data rather than real RF measurements; RTR-GS does not model higher-order indirect effects or complex participating media; R$^2$-Gaussian is weaker under severely limited-angle CT and does not model calibration errors such as geometric drift or reference $I_0$ nonuniformity; stochastic ray tracing can incur higher variance under heavy overlap or extreme opacity ranges. This suggests that future RadGS work will likely continue in three directions: tighter physical modeling, stronger robustness under sparse or shifted acquisition conditions, and transport formulations that preserve Gaussian efficiency while broadening the class of radiative phenomena that can be represented.

Source: https://www.emergentmind.com/topics/3d-radiative-gaussians-radgs