---
title: 'Learn Screen-Space Subsurface Scattering (SSSS) '
url: https://www.emergentmind.com/topics/screen-space-subsurface-scattering-ssss
type: topic
---

# Learn Screen-Space Subsurface Scattering (SSSS) 

Screen-space subsurface scattering (SSSS) is a family of image-space approximations for rendering the spatially nonlocal appearance produced when light enters a translucent material, undergoes scattering, and exits at another surface location. Rather than solving volumetric light transport directly, SSSS generally operates on deferred-rendering buffers such as lighting, camera-space position, depth, normals, material parameters, and thickness-related information. Its characteristic operation is a spatially varying, color-dependent blur or learned filter over visible image pixels. SSSS is computationally attractive for interactive rendering, but its screen-space domain limits transport to information available in the current view and auxiliary buffers.

## 1. Physical basis and screen-space formulation

Subsurface scattering differs from local surface reflection because incident and outgoing light are associated with different surface locations. A BSSRDF expresses this nonlocal relationship schematically as

$$
S(x_i,\omega_i;x_o,\omega_o),
$$

where $x_i$ and $x_o$ are incident and exit locations, while $\omega_i$ and $\omega_o$ are incident and outgoing directions. Conventional analytic BSSRDFs often assume a planar half-space, homogeneous or surface-varying material properties, and diffusion-like transport. Direction-dependent diffusion models show that oblique incidence can produce anisotropic exit patterns rather than radially symmetric profiles [1609.02072].

SSSS replaces this surface-to-surface transport with an image-space approximation. A generic form is

$$
I_{\mathrm{sss}}(p)
\approx
\sum_{q\in\Omega(p)} K(p,q)\,I(q),
$$

where $p$ is the target pixel, $q$ indexes contributing pixels, $I(q)$ is an illumination or image contribution, and $K$ is a diffusion or scattering kernel. In conventional implementations, $K$ may depend on screen-space distance, depth, normals, thickness, material color, or scattering distance.

A principal reference formulation uses camera-space positions rather than merely two-dimensional pixel coordinates. For target pixel $p$ and source pixel $q$, the geometric distance is

$$
d_{pq}=\left\|P_{\mathrm{s}}(p)-P_{\mathrm{s}}(q)\right\|_2.
$$

A color-channel-dependent scattering result can then be written schematically as

$$
S_c(p)=
\sum_{q\in\Omega}
K_c\!\left(d_{pq};R_{\mathrm{scatt}}\right)L_c(q),
$$

where $K_c$ is a channel-specific scattering profile, $R_{\mathrm{scatt}}$ contains RGB scattering parameters, and $L_c(q)$ denotes the relevant source illumination. The profile is parameterized by the mean free path length for each color channel. Consequently, SSSS is spatially nonlocal and chromatic: different wavelengths propagate over different effective distances.

The approximation is usually applied to visible surfaces. It therefore differs from full volumetric rendering, which can account for hidden, back-facing, or off-screen geometry, refractive interfaces, multiple internal scattering events, and transport through arbitrary volumes. Screen-space methods can approximate these effects only insofar as they are encoded in the available buffers or learned from image-space examples.

### Distinction from related screen-space effects

SSSS is not synonymous with generic blur or with all screen-space shading effects. Ambient occlusion estimates local visibility or blocked directions; indirect illumination estimates bounced light; directional occlusion combines visibility with environment-map radiance; depth of field depends on focal-plane distance; motion blur follows object motion; and anti-aliasing smooths geometric discontinuities. SSSS specifically approximates light spreading through a translucent object.

A screen-space SSS filter is also distinct from a BRDF. A BRDF relates incident and outgoing directions at one surface point, whereas SSS requires a spatially nonlocal transport relationship. Likewise, a depth-aware blur can prevent some bleeding across discontinuities but does not by itself provide true object-space distance, thickness, or hidden-surface transport.

## 2. Classical kernels, profiles, and directional transport

### Diffusion profiles and separable filters

Real-time SSSS commonly approximates diffusion using a sum of separable Gaussian filters. Conceptually,

$$
K(r)\approx\sum_{m=1}^{M}w_mG_{\sigma_m}(r),
$$

where $r$ is a spatial distance, $G_{\sigma_m}$ is a Gaussian profile, and $w_m$ is its weight. Two-dimensional filtering is implemented with horizontal and vertical passes. This construction gives predictable execution and a cost approximately linear in image resolution, but it imposes a relatively low-dimensional diffusion model that is often isotropic or locally isotropic.

Depth-aware variants modulate the kernel using depth and normal differences. A generic weighting structure is

$$
w_{ij}
=
w_{\mathrm{spatial}}(i,j)
w_{\mathrm{depth}}(i,j)
w_{\mathrm{normal}}(i,j).
$$

Such factors reduce bleeding across silhouettes and depth discontinuities, but they do not remove the underlying screen-space visibility limitation.

### Effective shader parameters

Real-time skin shaders often use an effective scattering distance rather than intrinsic volumetric coefficients. A dual-exponential diffusion profile can be written as

$$
R(r,d)
=
A\,
\frac{e^{-r/d}+e^{-r/(3d)}}{8\pi d r},
$$

where $A$ is diffuse volume albedo and $d$ is the effective scattering distance. The parameter $d$ controls the radius and falloff of the scattering halo; it is a real-time shader parameter rather than a direct estimate of $\sigma_s^{-1}$ or a measured mean free path [2611.14451].

A differentiable SSSS implementation can sample screen-space locations while evaluating the profile using reconstructed three-dimensional distances. For target position $\mathbf{x}$ and reconstructed sample position $\mathbf{x}_i$,

$$
s_i=\left\|\mathbf{x}_i-\mathbf{x}\right\|_2.
$$

The screen-space offset is therefore used to locate candidate samples, while the profile is evaluated using view-space geometry. This separates image-space sampling from the transport-distance approximation.

### Non-perpendicular incidence

Perpendicular-incidence diffusion is often represented with a radially symmetric profile. Under oblique illumination, however, the refracted internal beam establishes a preferred direction. The multiple-scattering BSSRDF depends on albedo, incident angle, radial distance, and azimuth:

$$
S_p^{\mathrm{MS}}
(\rho,\theta,r,\phi).
$$

The angular dependence arises from the lateral displacement of the internal beam. A compact representation approximates the azimuthal profile with a General Wrapped Cauchy function,

$$
f_{\mathrm{GWC}}(\phi;\alpha,\beta,c)
=
\alpha+\beta\,\operatorname{pdf}_{\mathrm{WC}}(\phi;c),
$$

where $\alpha$ is an approximately uniform component, $\beta$ controls the directional component, and $c$ controls angular concentration. The model stores $(E,\beta,c)$ over $(\rho,\theta,r)$ and reconstructs $\alpha$ analytically. A table with 100 albedo samples, 10 incident-angle samples, and 64 radial samples occupies approximately 1 MiB and reproduces Photon Beam Diffusion with approximately $0.5\%$ mean relative error, including grazing angles [1609.02072].

This representation suggests a directional SSSS kernel of the form

$$
K(r,\phi\mid\rho,\theta)
\approx
f_{\mathrm{GWC}}
\left(
\phi;
\alpha(\rho,\theta,r),
\beta(\rho,\theta,r),
c(\rho,\theta,r)
\right).
$$

Such a kernel can represent directional bleeding at grazing angles, but it is more complex than a separable isotropic blur. It requires estimating the local incident angle and projecting the incidence-plane direction into screen space.

### Single and multiple scattering

Diffusion models principally describe multiple scattering. Near boundaries and in thin geometry, the first scattering event can remain directional, making diffusion inaccurate. A volumetric Gaussian-splatting method addresses this by combining a diffusion solve with an explicit single-scattering term along the camera chord:

$$
L_1(g_i)
=
\frac{1}{4\pi}
\int_0^{U_i}
Q(\mathbf{x}(u))e^{-\sigma_tu}\,du.
$$

The multiple-scattering component is obtained from a fluence field $\phi$ and an interior sample. This separation is relevant to SSSS because thin ears, noses, leaves, and other translucent structures can contain a significant single-scattering or transmission component [2610.04007].

## 3. Learned screen-space shading

“Deep Shading” treats SSSS as a learned approximation to a conventional screen-space SSS operator rather than as a new physically derived diffusion model. Reference images are generated with a conventional screen-space implementation based on Pixar’s scattering profile. A convolutional neural network then learns the mapping from deferred-shading attributes to RGB appearance [1603.06078].

The dedicated SSS network receives

$$
P_{\mathrm{s}},\qquad R_{\mathrm{scatt}},\qquad L,
$$

where $P_{\mathrm{s}}$ is camera-space position, $R_{\mathrm{scatt}}$ is the RGB mean-free-path parameter, and $L$ is direct illumination. Camera-space positions are used because absolute world positions do not provide additional shading information and could encourage memorization of geometry. The network does not require a separately supplied thickness field, curvature, surface tangent, irradiance gradient, or explicit diffusion radius.

The reference method evaluates contributions from other pixels according to the Euclidean distance between their camera-space positions. The learned method instead computes

$$
\widehat{S}_c(p)
=
\tilde f_{\mathbf w,c}(X)(p),
$$

where $X$ is the screen-space attribute tensor and $\mathbf w$ denotes the learned CNN parameters. The CNN does not explicitly evaluate the reference sum during inference.

### Architecture

The network is a fully convolutional U-shaped encoder-decoder. Its downsampling branch reduces image resolution, a bottleneck processes the coarsest representation, and an upsampling branch restores the original resolution. Skip connections connect corresponding levels. The architecture uses mean pooling for downsampling, bilinear interpolation for upsampling, grouped convolutions, and leaky ReLU activations.

The SSS configuration uses five resolution levels, $3\times3$ kernels, and $u_0=8$ channels at the first level. The channel progression on the down branch is

$$
8,\;16,\;32,\;64,\;128.
$$

The network contains approximately $133\,\mathrm{K}$ parameters and is applied independently to the red, green, and blue channels. Training is performed in Caffe, while inference is reimplemented in OpenGL shaders using array textures.

### Data and optimization

The complete dataset contains 61,000 pairs of deferred-shading buffers and reference images: 54,000 training images, 6,000 validation images, and 1,000 test images. Training and validation use ten scenes, while testing uses four different scenes. Images are $256\times256$ for SSS, and the network is evaluated at $768\times512$.

The training data vary camera views, object categories, random textures, and scattering parameters. The reference rendering uses a screen-space procedure based on Pixar’s scattering profile. Training data are stored as linear HDR images without gamma correction or tone mapping before training.

The actual loss is DSSIM rather than the standard squared loss. Images are tiled into $8\times8$ patches, SSIM is evaluated between corresponding patches and channels, and the loss is

$$
\operatorname{DSSIM}
=
\frac{1-\operatorname{SSIM}}{2}.
$$

Optimization uses stochastic mini-batch training, backpropagation, ADADELTA adaptive learning rates, and momentum $0.9$. The authors report that $L_1$ and $L_2$ losses tend to produce halos, whereas SSIM produces smoother fading behavior.

### Performance and limitations

The dedicated SSS model obtains a test SSIM of $0.905$ and a runtime of $21\,\mathrm{ms}$ at $768\times512$. The result is described as almost indistinguishable from the reference implementation. Its advantages include fixed GPU execution, a large effective receptive field through multiscale processing, and conditioning on geometry, direct lighting, and scattering parameters.

The model remains an approximation to a screen-space approximation. It requires an offline reference dataset, deferred buffers, memory for intermediate feature maps, and training coverage over the intended geometry, illumination, material, and resolution distributions. It can over-blur, ring, propagate influence in the wrong direction, or blur across silhouettes. It cannot represent contributions from surfaces that are occluded, clipped, culled, outside the image, or behind the camera.

## 4. Image-based and object-based neural representations

Several neural approaches extend beyond a conventional current-frame blur while remaining relevant to SSSS.

### Learned pixel-footprint transport

A neural acquisition method predicts an image-space scattering footprint for every visible surface pixel from six high-frequency phase-shifted structured-light images. The footprint is anisotropic, spatially varying, and conditioned on geometry and material appearance, but the method does not infer BSSRDF parameters or intrinsic optical coefficients [2606.02292].

A stereo projector-camera system captures horizontal and vertical sinusoidal patterns. Phase-shifted profilometry provides camera-projector correspondence and depth, while the network predicts the camera-space response caused by localized projector illumination. Measured dot patterns provide footprint targets during training. A typical grid spacing is $(55,55)$ projector pixels, and approximately 29,000 pairs of PSP tiles and measured footprints are used for training.

At relighting time, the predicted footprint is multiplied by the projector color associated with the corresponding projector coordinate and splatted into the output framebuffer. The representation is therefore a collection of view-conditioned, image-space impulse responses rather than a translation-invariant convolution. It can represent anisotropic and irregular footprints, but it includes baked appearance effects such as highlights; highlights do not move when the object is virtually rotated.

The method is SSSS-like because it performs image-space accumulation of spatially extended transport, but it is not a conventional runtime SSSS shader. It requires calibrated projector-camera acquisition, phase processing, neural inference, and footprint splatting. It does not recover absorption, scattering, reduced scattering, index of refraction, diffusion length, or phase-function parameters.

### Neural volumetric relighting

“Neural Relighting with Subsurface Scattering by Learning the Radiance Transfer Gradient” uses a volumetric neural representation rather than screen-space filtering. A density field and a neural radiance-transfer field are queried at three-dimensional sample points and integrated along camera rays [2306.09322].

For ray samples $\mathbf{x}_i$, the method predicts density $\hat{\sigma}(\mathbf{x}_i)$ and a nonnegative RGB transfer-gradient-like quantity $\hat{\mathbf h}(\mathbf{x}_i;\mathbf r;\omega)$. The rendered image is

$$
\hat{\mathbf I}(u,v;\omega)
=
\sum_{i=1}^{N}
\hat w(\mathbf{x}_i)
\hat{\mathbf h}(\mathbf{x}_i;\mathbf r;\omega),
$$

with volume-rendering weights

$$
\hat w(\mathbf{x}_i)
=
\hat T_i
\left(1-e^{-\hat\sigma(\mathbf{x}_i)\delta_i}\right).
$$

The method can represent multiple density layers, internal structure, global illumination, view dependence, and transport through translucent regions. It does not blur a framebuffer, use a screen-space diffusion kernel, or explicitly fit a BSSRDF. Its advantage over SSSS is access to object-centered volumetric information; its disadvantages are extensive multi-view, multi-light HDR capture, neural volume rendering, and lack of real-time performance.

### Neural full-object BSSRDFs

A neural BSSRDF represents the full-object response to a distant directional light. Its learned function takes incident direction, observation position, and outgoing direction:

$$
f_{\psi^*}(\omega_i,x_o,\omega_o).
$$

The outgoing radiance under an environment map is approximated by

$$
L_r(x_o,\omega_o)
\approx
\int_{4\pi}
L_i(\omega_i)
f_{\psi^*}(\omega_i,x_o,\omega_o)
\,d\omega_i.
$$

Unlike SSSS, this representation can incorporate off-screen and back-facing transport, self-shadowing, refractive interfaces, heterogeneous volumes, and embedded structures. It is object- and material-specific: geometry and volume properties are baked into the network weights rather than supplied as editable runtime parameters [2312.15711].

A direct MLP mode supports more than 60 samples per pixel per second in the reported setup, while a degree-2 spherical-harmonic variant achieves less than $0.01$ seconds per frame at reduced accuracy. The method is less suitable for dynamic deforming geometry, arbitrary local lights, or continuously changing material parameters.

### Gaussian-based representations

GTSR separates surface Gaussians from interior Gaussians. Surface Gaussians represent geometry, normals, depth, base color, roughness, and surface reflection; interior Gaussians represent view-dependent scattering color using spherical harmonics. Fresnel-weighted opacity connects the two sets during compositing [2603.22036].

The deferred branch renders surface normal, depth, base-color, and roughness maps and then applies a Disney-style BSDF. The final result is conceptually

$$
\mathbf C_{\mathrm{PBR}}
=
\mathbf C_s
+
\gamma\mathbf C_d
+
(1-F_{\mathrm{Schlick}})
\mathbf C_{\mathrm{scatter}},
$$

with $\gamma=0.3$ in the reported experiments. GTSR is not a physically based SSSS solver: it does not explicitly evaluate a BSSRDF, diffusion profile, refraction, or general multiple scattering. Its principal contribution is separating surface reconstruction from learned interior appearance.

VolS-GS instead converts a Gaussian scene into a $64^3$ occupancy grid and solves a diffusion equation over the reconstructed volume. The continuous transport equation is

$$
-\nabla\cdot(\kappa\nabla\phi)
+
\sigma_a\phi
=
b,
$$

with a Robin boundary condition. Finite-volume discretization yields a sparse system

$$
F\phi=b.
$$

The medium is parameterized with learned scattering and absorption coefficients, but these coefficients are fitted appearance parameters rather than validated physical measurements. A single-scattering term is added to the diffusion result, and the resulting radiance is composited using Gaussian splatting [2610.04007].

VolS-GS is geometry- and volume-aware rather than screen-space. It can propagate light through occupied interior regions and avoid spreading a Euclidean kernel across empty gaps in concave or disconnected geometry. Its reported rendering rate is approximately $64.5$ FPS on $256^2$ synthetic SSS-GS scenes on an A6000, with the transport solve accounting for roughly half of representative frame time.

## 5. Acquisition and inverse rendering for SSSS parameters

SSSS is a forward-rendering approximation, but several methods estimate the parameters and buffers required by a production shader.

### Polarization-based estimation

A polarization-based method jointly estimates shape, illumination, and homogeneous SSS parameters from a single polarization-camera capture. Its explicit SSS parameters are extinction coefficient $\sigma_t$, volumetric albedo $\alpha$, and Henyey–Greenstein anisotropy $g$ [2407.08149].

The reduced Stokes vector is

$$
\mathbf{s}=[s_0,s_1,s_2],
$$

with degree of linear polarization

$$
\rho
=
\frac{\sqrt{s_1^2+s_2^2}}{s_0}.
$$

The method uses four analyzer angles, $0$, $\pi/4$, $\pi/2$, and $3\pi/4$, and derives $I_{\max}$ and $I_{\min}$ as a MaxMin polarization representation. The network predicts normals, depth, illumination, and SSS parameters. On synthetic data, the full method reports MAE values of $0.1340$ for extinction, $0.0918$ for albedo, and $0.1554$ for anisotropy.

These outputs can inform an SSSS pipeline. Given

$$
\sigma_s=\alpha\sigma_t,
\qquad
\sigma_a=(1-\alpha)\sigma_t,
$$

a renderer can derive a reduced transport coefficient,

$$
\sigma_t'
=
\sigma_a+(1-g)\sigma_s.
$$

However, the method does not directly output a diffusion profile, thickness map, curvature map, per-pixel scattering radius, or screen-space convolution kernel. A separate diffusion calculation is required to convert the inferred optical parameters into a profile usable by SSSS.

### Differentiable production-shader fitting

A lightweight skin-rendering system directly optimizes the scattering distance used by a real-time Burley-style SSSS shader rather than estimating intrinsic optical coefficients. It uses a single-light, single-view capture setup and estimates spatially varying albedo, roughness, a temporary specular residual, and effective scattering distance [2611.14451].

The differentiable renderer samples a screen-space disk but reconstructs three-dimensional positions from depth. The diffusion profile is evaluated using the Euclidean distance between the target and sampled view-space positions. The optimization uses 128 samples and a masked loss focused on shadowed transition regions, where subsurface halos are most observable.

Facial regions are fitted separately and combined into a spatially varying scattering-distance map. Thickness is represented with a texture or approximated using inverted-normal screen-space ambient occlusion. The output is directly compatible with real-time material pipelines: base color, roughness, IOR/specular settings, thickness, scattering-distance, and optional region maps.

The real dataset contains 45 participants and 225 skin patches spanning the Monk Skin Tone Scale. The method reports relatively stable perceptual errors across tone groups and a masked-loss improvement in FLIP from $0.11$ to $0.06$ in the shadow-transition region. The estimated scattering distance generally decreases for darker, more melanin-rich skin, consistent with stronger absorption.

The method deliberately accepts the limitations of SSSS in exchange for direct deployment in interactive engines. It assumes locally homogeneous patches, uses monocular depth, does not optimize explicit single scattering, and remains subject to screen-space visibility, disocclusion, and depth-buffer limitations.

## 6. Applications, artifacts, and design trade-offs

### Real-time rendering

SSSS is suitable when the scene is represented by a deferred buffer, visible positions or depth are available, direct illumination can be generated quickly, and approximate screen-space transport is acceptable. Its principal advantages are low implementation complexity, predictable GPU execution, compatibility with rasterization, and the ability to integrate with standard material systems.

A screen-space filter is particularly effective for local soft transport on visible surfaces. It is less reliable for light entering at an off-screen or back-facing location, crossing a deep concavity, passing through thin geometry, or emerging after transport through hidden volume. Thickness maps, back-face depth, bent normals, object-space distances, and ray-traced transmission can improve these cases but increase complexity.

### Failure modes

Common SSSS artifacts include:

- **Over-blurring**: insufficient capacity or an excessive profile radius softens texture boundaries and fine translucent structure.
- **Halos and ringing**: inaccurate filters or pixelwise losses can create bright or dark bands around illuminated regions.
- **Depth-discontinuity leakage**: samples from unrelated surfaces or background regions can contribute across silhouettes.
- **Wrong-direction propagation**: isotropic filters cannot represent directional transport under oblique incidence.
- **Thin-geometry ambiguity**: front and back surfaces can be confused with nearby surfaces separated by empty space.
- **Background darkening**: filtering across the object boundary can alter background pixels.
- **Off-screen truncation**: missing surfaces outside the viewport cannot contribute.
- **Temporal instability**: camera motion, changing visibility, disocclusions, and changing screen-space neighborhoods can alter the result.
- **Material extrapolation failure**: learned filters may behave plausibly but incorrectly for scattering distances, spectral behavior, anisotropy, or layered materials outside the training distribution.

Independent RGB-channel networks or filters can also generate inconsistent chromatic transitions. Because scattering profiles are channel-dependent, errors may appear as colored halos or channel-specific ringing.

### Learned versus analytic approaches

Analytic SSSS profiles expose interpretable controls such as scattering distance, albedo, and channel-specific radii. They are easy to modify but depend on assumptions about geometry, thickness, and transport. Learned filters can condition on positions, illumination, material parameters, and broader image context, but they require training data and generalize only within the relevant distribution.

A learned representation can also absorb geometry, illumination, and material errors into appearance. This is particularly evident in object-specific neural BSSRDFs, neural volume fields, and Gaussian representations: they reproduce captured or path-traced appearance but do not necessarily expose editable absorption, scattering, anisotropy, or thickness parameters.

### Hybrid architectures

The research suggests several hybrid configurations:

1. **Classical local diffusion with learned residuals**: use a separable screen-space profile for the dominant local component and a learned residual for transmission, silhouettes, or complex material regions.
2. **Directional screen-space kernels**: condition a compact angular profile on incident angle and projected incidence direction.
3. **Thickness-aware SSSS**: combine front and back depth, object-space positions, normals, and extinction estimates.
4. **Surface-plus-interior representations**: retain explicit surface geometry while representing interior scattering separately.
5. **Object-specific quality modes**: use SSSS for ordinary dynamic assets and neural BSSRDF or volumetric transport for selected translucent objects.
6. **Single- and multiple-scattering decomposition**: use a diffusion-like screen-space component for multiple scattering and a transmission or ray-based component for thin regions.
7. **Polarization-assisted acquisition**: estimate normals, depth, and optical parameters before converting them into production-oriented SSSS controls.

These approaches do not eliminate the distinction between screen-space approximation and physical transport. They instead distribute the computation between inexpensive image-space operations and more informative object-space or learned components.

## 7. Position within subsurface-rendering research

SSSS occupies an intermediate position among several representations:

| Representation | Principal domain | Characteristic capability |
|---|---|---|
| Analytic SSSS | Image space | Fast visible-surface diffusion |
| Screen-space neural shading | Image space and deferred buffers | Learned approximation of screen-space operators |
| Texture-space diffusion | Surface or UV space | More stable object-centered filtering |
| Neural BSSRDF | Object and direction space | Full-object directional transport |
| Neural volumetric rendering | Three-dimensional volume | Density-dependent ray integration and internal structure |
| Gaussian transport | Sparse object-space volume | Learned or solved interior propagation |
| Path-traced BSSRDF | Surface and volume transport | Explicit or approximated physical light transport |

The defining property of SSSS is not merely that it renders a soft translucent appearance. It is that the approximation is performed in screen or image space, usually using visible-surface buffers and a spatial filter. Neural image-space footprints remain SSSS-like because they predict and splat image-space responses. Neural volumes, neural BSSRDFs, and finite-volume Gaussian solvers are related subsurface-rendering methods but are not conventional SSSS.

The central trade-off is therefore between computational locality and transport completeness. Screen-space methods offer interactive performance and straightforward integration but lack hidden geometry and full object-space transport. Object-centered neural and volumetric methods represent richer transport and heterogeneous structure but require acquisition, training, memory, and more expensive inference. Current research increasingly combines these regimes: analytic or learned screen-space approximations for ordinary rendering, and object-specific or volumetric representations where off-screen transport, internal structure, anisotropy, or material heterogeneity dominates.

Source: https://www.emergentmind.com/topics/screen-space-subsurface-scattering-ssss