EWA Splatting Explained
- EWA Splatting is/are a framework that transforms singular geometric primitives into differentiable, elliptical Gaussian footprints for rendering, used in 3D reconstruction and volumetric rendering.
- Key applications include Gaussian-based volumetric rendering, 3D Gaussian Splatting, surface splatting, and scene reconstruction, as well as extensions to curvilinear and non-Gaussian primitives and analytic techniques like clipping.
- The method is widely adaptable for various rendering pipelines and includes extensions like the use of differentiable filtering, gamma-based opacity, and the application of kernel types beyond Gaussians.
EWA Splatting, or Elliptical Weighted Average splatting, is a reconstruction and rendering framework in which singular geometric primitives are convolved with low-frequency kernels, projected into an image or angular domain, and accumulated using finite-support or effectively finite-support footprints. Its central operation is the transformation of a primitive’s spatial covariance into a screen-space covariance, producing an elliptical Gaussian footprint evaluated during rasterization. EWA formulations are used in surface splatting, Gaussian-based volumetric rendering, 3D Gaussian Splatting (3DGS), differentiable inverse rendering, event-based reconstruction, camera-generalized rendering, analytic clipping, and extensions to curvilinear and non-Gaussian primitives. The term also denotes a broader recipe—convolving a geometric measure with a reconstruction kernel—rather than a single universally fixed renderer.
1. Mathematical foundation and rendering model
A point, curve, or surface represented as an infinitesimally thin geometric object can be modeled by a measure involving Dirac distributions. Directly rendering such a measure is problematic because sampling it at a pixel or along a ray generally produces zero almost everywhere. EWA Splatting regularizes the representation by convolving the measure with a smooth, low-frequency kernel before rendering. In this interpretation, the rendering objective can be written schematically as
where is the geometric measure of primitive , is its reconstruction kernel, and is volume rendering from camera (Shramko et al., 9 Nov 2026).
For a point primitive at , convolution with an anisotropic Gaussian kernel gives
The result is a Gaussian centered at with covariance . The primitive specifies the support location, while the kernel specifies how the support extends through space. Standard 3DGS is the point-plus-Gaussian instance of this construction; a zero-length segment yields the same limiting representation (Shramko et al., 9 Nov 2026).
A Gaussian primitive is commonly parameterized by a mean and covariance,
with
where is a rotation and is a diagonal scale matrix. The rotation is often represented by a quaternion and the scale by a three-vector (Yura et al., 2024).
Under a locally affine camera projection, a world-space point is transformed into image coordinates according to
where , is the world-to-camera transformation, and is the projection Jacobian evaluated at the primitive center. The projected covariance is
Equivalently, in the conventional 3DGS notation,
For a pixel displacement from the projected mean, the EWA footprint has the form
The inverse projected covariance is often stored as a two-dimensional conic. Rasterization evaluates the conic over the primitive’s screen-space support, commonly restricted by an axis-aligned bounding region derived from the projected covariance (Wu et al., 2024).
The footprint is then incorporated into a compositing model. In opacity-based 3DGS, the contribution of primitive is typically
and the rendered color is
where primitives are ordered approximately front-to-back. This rendering model preserves the efficiency of tile-based rasterization, but its is a learned opacity-like quantity rather than necessarily an integrated volumetric extinction (Wu et al., 2024).
2. EWA projection, filtering, and volumetric interpretations
EWA projection is based on a local first-order approximation of the camera mapping. For a perspective camera, the projected covariance is computed from the Jacobian at the Gaussian mean. This approximation is effective when the projection is close to affine across the primitive, but it becomes less accurate for large or deep Gaussians, wide fields of view, strong distortion, and particles close to the camera (Wu et al., 2024).
The distinction between EWA extinction and opacity-based 3DGS is substantial. In an extinction-based formulation, a normalized three-dimensional Gaussian with integrated mass is projected to a normalized two-dimensional Gaussian while preserving its total integral. The projected peak amplitude changes with the projected ellipse’s area. By contrast, original 3DGS directly learns a projected amplitude or opacity and preserves that amplitude during projection. Consequently, standard 3DGS does not correspond to a single view-independent three-dimensional extinction field under the usual volumetric interpretation (Celarek et al., 26 Feb 2025).
The volumetric reference model is the emission–absorption equation
where is the extinction coefficient and the exponential factor is ray transmittance. For a Gaussian mixture, is the sum of the individual normalized Gaussian densities. EWA avoids evaluating this full mixture along every ray by projecting each Gaussian to a two-dimensional footprint and compositing the projected contributions (Celarek et al., 26 Feb 2025).
This introduces several approximations:
- Local projection linearization: perspective projection is approximated by a first-order Taylor expansion around each primitive center.
- Center-depth ordering: overlapping primitives are globally ordered using their centers, even when their volumes intersect differently across pixels.
- Independent primitive treatment: Gaussian overlap is not integrated jointly along each ray.
- Missing self-attenuation: attenuation internal to an individual Gaussian may be omitted.
- Linearized attenuation: the rasterizer commonly uses $1-g$ rather than exact exponential attenuation .
- Mean-based appearance evaluation: view-dependent color may be evaluated using the primitive center rather than continuously through its volume.
The paper “Does 3D Gaussian Splatting Need Accurate Volumetric Rendering?” distinguishes three regimes: EWA splatting, original opacity-based 3DGS, and ray-marched Gaussian rendering. Ray marching handles overlap, per-pixel visibility, exponential transmittance, and more accurate perspective geometry, but it is reported to be one to two orders of magnitude slower than splatting. Accurate volumetric rendering generally helps when there are few large primitives, whereas conventional opacity-based 3DGS often catches up or surpasses it when many Gaussians are available (Celarek et al., 26 Feb 2025).
An extinction-oriented alternative called opacity-thin-side, or OTS, retains a rasterization pipeline while parameterizing amplitudes to preserve view-dependent projected extinction. OTS and OTS with self-attenuation generally improve small models, while their advantage decreases as the number of primitives increases. Extinction-based methods remain advantageous for genuinely volumetric scenes, reflective materials, grazing-angle appearance, and situations in which view-dependent apparent thickness is important (Celarek et al., 26 Feb 2025).
3. Surface EWA splatting and coverage-based visibility
Classical EWA surface splatting reconstructs signals from irregularly sampled surface points using overlapping tangent-plane kernels and Shepard normalization:
Here is the reconstruction weight and is an attribute or radiance associated with surfel . Shepard normalization compensates for variable sample density and overlapping kernels, but it removes the absolute magnitude of the accumulated weight. A pixel with a single small contribution can therefore receive the same normalized color as a fully covered pixel.
The 3D Surface Splatting method, 3DSS, addresses this distinction by treating primitives as oriented surface samples rather than volumetric radiance primitives. A surfel has a center , tangent vectors and , a normal
and material properties including albedo, metallicity, and roughness (Younes et al., 7 May 2026).
Instead of relying on a planar projected conic alone, 3DSS evaluates the kernel through a perspective-correct object-space ray–surfel intersection. A surfel is represented as a degenerate ellipsoid, and a transformation
maps local surfel coordinates to clip space. This transformation supports projected bounding-box computation, ray–surfel intersection, depth-extent computation, and differentiable backward propagation.
For local tangent coordinates , the unfiltered EWA weight is
with contributions restricted to a cutoff radius. 3DSS further introduces object-space MIP filtering. If is the Jacobian of the ray–surfel intersection mapping and is the screen-space pixel-filter covariance, the local covariance becomes
The filtered kernel is evaluated using the inverse covariance of . MIP filtering addresses sampling aliasing, whereas coverage-derived opacity addresses visibility and silhouette aliasing (Younes et al., 7 May 2026).
3DSS separates depth-connected surface layers using depth intervals. If is a surfel’s camera-space depth and its depth extent, its interval is
Intervals are sorted by their starting depth and merged whenever they overlap, including through chains of overlapping intervals. A new layer begins only when a gap appears. This differs from center-depth sorting because it explicitly models surfaces that arise from connected kernel overlap.
For layer , the accumulated unnormalized weight and color are
The normalized layer color is
The layer opacity is derived from coverage:
With front-to-back compositing,
This opacity represents fractional surface coverage, not material transparency or learned volumetric density. The resulting model produces differentiable silhouettes and visibility gradients at sparsely covered edges. The 3DSS results reported in the supplied material include a mean depth SI-MSE of $0.06$, normal error of $0.00$, and Chamfer value of $0.01$ for a fixed-geometry configuration (Younes et al., 7 May 2026).
4. Extensions beyond planar Gaussian footprints
The local affine assumption of EWA is a central source of both efficiency and limitation. 3DGUT replaces Jacobian-based projection with an Unscented Transform (UT). Instead of linearizing the camera function, it represents a three-dimensional Gaussian by seven sigma points, projects each point through the actual nonlinear camera model, and reconstructs a projected mean and covariance from the transformed points (Wu et al., 2024).
For a Gaussian with mean and covariance , the sigma points are propagated through
where may include perspective, fisheye distortion, arbitrary camera intrinsics, and time-dependent extrinsics. The projected mean and covariance are reconstructed as weighted moments. This retains a two-dimensional rasterization footprint while supporting distorted cameras and rolling shutter without manually deriving a camera-specific Jacobian.
3DGUT also introduces a three-dimensional particle response evaluated along each ray. For a ray , the maximum-response parameter is
The renderer can use UT for candidate-particle selection while evaluating the Gaussian response in three dimensions. This makes the rasterized representation more compatible with secondary-ray tracing for reflections and refractions. The supplied timing comparison reports approximately $2.88$ ms per image for conventional 3DGS, $3.77$ ms for an unsorted 3DGUT variant, $4.98$ ms for a sorted variant, and $19.24$ ms for 3DGRT (Wu et al., 2024).
UniTriSplat generalizes the splatting domain from a camera-specific plane to the unit sphere. A Gaussian mean is mapped to longitude and latitude,
and its covariance is propagated into angular coordinates. Because longitude does not represent uniform physical distance at different latitudes, UniTriSplat introduces the local metric
with arc-length covariance
The spherical footprint is rasterized on a HEALPix grid, whose cells have equal solid angle. The per-cell contribution is an inverse-quadratic Gaussian weight evaluated in local arc-length coordinates. Although structurally analogous to an EWA footprint, UniTriSplat does not derive or perform exact EWA integration over curved HEALPix cells; it is more precisely a HEALPix-aware discrete Gaussian splatter (Zhu et al., 29 Jun 2026).
EventSplat applies projected Gaussian splatting to event-camera reconstruction. Its scene is represented by three-dimensional Gaussian ellipsoids with centers, covariances, appearance, and opacity. The renderer projects the Gaussians to two-dimensional ellipses, depth sorts them, and alpha composites them. The event-specific contribution is the supervisory signal: two rendered images at different poses are remosaiced, converted to logarithmic intensity, and differenced,
This is compared with the accumulated signed event image. EventSplat therefore uses an EWA-like projected-Gaussian rasterizer, but its principal novelty lies in temporal log-intensity supervision, event-to-video initialization, and high-rate camera-pose interpolation rather than in a new EWA kernel (Yura et al., 2024).
5. Analytic modification by half-space clipping
XClipGS extends EWA splatting by analytically restricting a Gaussian to one side of a clipping plane. The plane is
Under the local affine projection used by EWA, a pixel corresponds to an affine ray. The unrestricted ray integral of a three-dimensional Gaussian produces the ordinary two-dimensional EWA footprint. The restricted ray integral factorizes into that footprint multiplied by a conditional Gaussian probability:
Let
and
The conditional variance of the plane coordinate given the image position is
For pixel displacement , the conditional retained probability is
Thus the clipped contribution is
The conditional CDF varies across the projected footprint, unlike multiplying the entire splat by the globally retained mass . Pixels on the retained side approach a factor of one, pixels on the removed side approach zero, and pixels near the intersection receive intermediate values. The depth ordering and front-to-back compositing remain unchanged; only the per-pixel splat weight is modified (Gao et al., 7 Aug 2026).
The exactness claim is restricted to the half-space ray integral within the affine EWA model. The complete renderer still includes local projection linearization, center-based depth ordering, finite-precision rasterization, and approximation of the underlying CT or MRI volume by Gaussians. The paper reports $33.56$ dB PSNR, $0.960$ SSIM, $0.075$ LPIPS, and $655$ FPS for XClipGS in its principal comparison. In cut-face evaluation, it reports band SSIM of $0.860$ and leakage of $0.40$ in the stated units (Gao et al., 7 Aug 2026).
6. Generalizations of the kernel and primitive
Spaghetti Splatting interprets EWA as a general construction for converting singular geometric supports into differentiable, anti-aliased volumetric primitives. It extends the framework along two axes: the geometric primitive and the reconstruction kernel (Shramko et al., 9 Nov 2026).
For a finite segment with endpoints and , define
The segment measure is
Convolving it with a Gaussian gives
which can be interpreted as the accumulated Gaussian response of every infinitesimal point along the segment. A polyline is the sum of connected segment measures.
For a ray , the optical depth of a convolved segment reduces to an integral over the segment parameter. The paper obtains a one-dimensional Gaussian in and a finite-segment integral involving the error function:
The associated ray opacity is
Connected segments are volume-rendered and composited with transmittance rather than simply added as independent opaque strokes. This avoids overcounting at joints and preserves consistency under segment subdivision.
The kernel can also be changed. In addition to Gaussian kernels, Spaghetti Splatting supports compact Beta kernels. The geometric construction remains the same, while the profile, support, projected footprint, and integral evaluation change. The paper therefore treats point-plus-Gaussian, point-plus-Beta, segment-plus-Gaussian, and segment-plus-Beta representations as instances of the same primitive–kernel framework.
A point is a zero-length curve: setting collapses the segment measure to a point measure and recovers the Gaussian primitive of 3DGS. Nonzero-length segments provide connected elongated structures that ordinary point splatting represents only through many unrelated points. The method uses per-segment camera linearization, midpoint EWA computation, inverse-depth scaling, and spatial filtering. Reported results include PSNR values of $28.78$ on Mip-NeRF 360 for Beta Spaghetti Splatting, $37.38$ on Shelly for Gaussian Spaghetti Splatting, and $35.03$ on Fuzzy for Gaussian Spaghetti Splatting. On hair reconstruction, Gaussian Spaghetti Splatting is reported with PSNR $32.25$, SSIM $0.940$, and LPIPS $0.043$ (Shramko et al., 9 Nov 2026).
Other work replaces the Gaussian kernel entirely when the sensing physics requires a different point-spread function. Dirichlet Splatting models finite-window coherent measurements using the complex Dirichlet kernel rather than an elliptical Gaussian. For a rectangular window,
Its oscillatory sidelobes, periodic nulls, and complex phase encode the finite-window Fourier response. The paper argues that anisotropic Gaussian covariance cannot reproduce this response because stretching or rotating a monotonic Gaussian envelope does not produce the required sidelobes or phase law. Dirichlet Splatting consequently uses additive complex phasor summation rather than alpha compositing and employs specialized optimization based on variable projection, residual certificates, hard surfel replacement, and periodic Levenberg–Marquardt correction (Chen et al., 30 Sep 2026).
These extensions clarify the scope of EWA Splatting. Its essential idea is not restricted to Gaussian point primitives or planar optical images. It is the use of a reconstruction kernel matched to a geometric measure and a rendering domain. Gaussian EWA is appropriate when an elliptical Gaussian is a suitable approximation to the measurement or reconstruction footprint. Surface EWA uses tangent-plane kernels and coverage-aware visibility. Spherical systems use angular covariances and equal-solid-angle sampling. Curvilinear systems convolve segments and polylines. Coherent sensing systems require complex, oscillatory kernels such as the Dirichlet response rather than Gaussian filtering.
7. Scope, limitations, and conceptual distinctions
EWA Splatting should be distinguished from several related but non-identical procedures.
EWA versus ordinary 3DGS. Both use projected Gaussian-shaped footprints and tiled rasterization, but EWA is associated with filtered reconstruction and, in volumetric formulations, integrated extinction. Original 3DGS uses learned per-primitive opacity and alpha compositing. The latter can optimize effectively with large primitive counts, although it is not generally equivalent to a view-independent extinction field (Celarek et al., 26 Feb 2025).
EWA versus exact volumetric rendering. EWA replaces ray-wise Gaussian-mixture integration with projected footprints and compositing. Ray marching handles overlap and transmittance more directly but incurs substantially greater computational cost. The practical preference is regime-dependent: accurate volumetric integration is more useful for few large primitives and genuinely participating media, while opacity-based 3DGS is often preferable for dense representations of solid objects (Celarek et al., 26 Feb 2025).
EWA versus UT projection. EWA linearizes the camera function and propagates covariance through a Jacobian. The Unscented Transform propagates sigma points through the nonlinear camera function and reconstructs moments afterward. UT supports distorted and time-dependent cameras more directly, but requires seven point projections for a three-dimensional Gaussian (Wu et al., 2024).
EWA versus surface splatting. Surface EWA uses kernels to reconstruct attributes of a surface layer and may derive opacity from aggregate coverage. Volumetric Gaussian splatting generally composites primitive opacities directly. In 3DSS, surface separation, coverage, and pre-shaded radiance are explicit (Younes et al., 7 May 2026).
EWA versus spherical Gaussian splatting. A spherical renderer such as UniTriSplat retains an EWA-like local quadratic kernel but changes the rasterization domain to the sphere and uses HEALPix equal-area cells. It does not, in the supplied formulation, perform exact integration over each spherical cell (Zhu et al., 29 Jun 2026).
EWA versus physics-matched splatting. Dirichlet Splatting demonstrates that a Gaussian footprint can be structurally inappropriate for coherent finite-window sensing. In such systems, the kernel is determined by the acquisition window, and complex phase and sidelobes are part of the measurement operator (Chen et al., 30 Sep 2026).
EWA versus analytic clipping. XClipGS preserves the EWA footprint and compositing pipeline while multiplying the footprint by a conditional Gaussian CDF. Its exactness is therefore an analytic extension within the affine EWA model, not a replacement for the complete renderer (Gao et al., 7 Aug 2026).
The principal limitations of EWA Splatting arise from local projection linearization, approximate visibility ordering, finite rasterization support, kernel-shape assumptions, and the distinction between discrete pixel evaluation and exact footprint integration. Nevertheless, its combination of covariance-based projected footprints, differentiable rasterization, filtering, and efficient compositing provides a common foundation for a broad family of rendering systems. Its most general formulation is:
The choice of measure, kernel, projection domain, visibility model, and compositing rule determines whether a particular implementation is best understood as classical EWA surface splatting, opacity-based 3DGS, extinction-based Gaussian rendering, spherical splatting, analytic clipped splatting, curvilinear splatting, or a physics-matched alternative.