Volume 32 (2013), Number 4

Eurographics Symposium on Rendering 2013

Nicolas Holzschuch and Szymon Rusinkiewicz
(Guest Editors)

Photon Beam Diffusion:

A Hybrid Monte Carlo Method for Subsurface Scattering
Ralf Habel1

Per H. Christensen2
1 Disney

(a)

Research Zrich

(b)

2 Pixar

(c)

Wojciech Jarosz1
Animation Studios

(d)

(e)

Figure 1: We solve the searchlight problem (a) using a Monte Carlo integration of dipoles distributed along either a normally
incident (b) or oblique (c) beam. We can apply our hybrid approach to standard subsurface scattering (d) and also to more
complex configurations using photon beams (e).
Abstract
We present photon beam diffusion, an efficient numerical method for accurately rendering translucent materials.
Our approach interprets incident light as a continuous beam of photons inside the material. Numerically integrating
diffusion from such extended sources has long been assumed computationally prohibitive, leading to the ubiquitous
single-depth dipole approximation and the recent analytic sum-of-Gaussians approach employed by Quantized
Diffusion. In this paper, we show that numerical integration of the extended beam is not only feasible, but provides
increased speed, flexibility, numerical stability, and ease of implementation, while retaining the benefits of previous
approaches. We leverage the improved diffusion model, but propose an efficient and numerically stable Monte
Carlo integration scheme that gives equivalent results using only 35 samples instead of 2060 Gaussians as in
previous work. Our method can account for finite and multi-layer materials, and additionally supports directional
incident effects at surfaces. We also propose a novel diffuse exact single-scattering term which can be integrated in
tandem with the multi-scattering approximation. Our numerical approach furthermore allows us to easily correct
inaccuracies of the diffusion model and even combine it with more general Monte Carlo rendering algorithms. We
provide practical details necessary for efficient implementation, and demonstrate the versatility of our technique by
incorporating it on top of several rendering algorithms in both research and production rendering systems.
Categories and Subject Descriptors (according to ACM CCS): I.3.7 [Computer Graphics]: Three-Dimensional
Graphics and RealismRay Tracing; I.6.8 [Simulation and Modeling]: SimulationMonte Carlo

1. Introduction
Many important materials exhibit noticeable translucency,
including wax, marble, juices, the foods we eat, and our own
skin. Accurate and efficient lighting simulation in such materials is a critical challenge in fields ranging from medical
physics to realistic image synthesis in computer graphics.
Light passes volumetrically through translucent materials
before being scattered or absorbed, so they are poorly apc 2013 The Author(s)

c 2013 The Eurographics Association and Blackwell PublishComputer Graphics Forum
ing Ltd. Published by Blackwell Publishing, 9600 Garsington Road, Oxford OX4 2DQ,
UK and 350 Main Street, Malden, MA 02148, USA.

proximated by traditional surface shading models. In fact,

subsurface scattering is just a simplified view of the more
general problem of light transport in participating media like
clouds, fog, smoke, and interstellar nebula.
Many rendering algorithms [PH10] numerically approximate light transport in participating media. Monte Carlo
methods, such as variants of path tracing [Kaj86] or photon
mapping [Jen96], are general since they can theoretically

R. Habel, P. Christensen & W. Jarosz / Photon Beam Diffusion

handle arbitrary geometries and material properties. Unfortunately, these methods are typically slow to converge, especially when the medium is highly scattering and many
bounces are needed to obtain converged results. Methods
based on the diffusion approximation [Sta95], on the other
hand, can be much more efficient in these situations, though
they place strict assumptions on the medium being simulated.

The improvements to diffusion theory, and the use of the

extended source term, made QD accurate for a much wider
range of translucent materials. Unfortunately, the quantization in QD has many practical drawbacks, including complicated and error-prone implementation, numerical instability,
and high computational complexity due to the large number
of Gaussians needed for accurate profiles.

In this paper, we present a new method for accurately rendering translucent materials called photon beam diffusion
(PBD). Our approach interprets incident light as a continuous beam of photons, and expresses its propagation through
the medium using enhancements to the state-of-the-art improved diffusion model [dI11]. Unlike previous work, we
formulate our solution as a Monte Carlo integration problem.
This provides increased flexibility, numerical stability, and
speed. Furthermore, our formulation allows us to easily extend the accuracy and capabilities of the diffusion model and
even combine it with more general Monte Carlo rendering
algorithms.

In our work, we show that we can compute the same extended source integral numerically in a straightforward and
very efficient wayeliminating the need for quantization and
the sum of Gaussians altogether. We show that with just 35
numerical samples we can obtain a profile of equal quality to
QD using 2066 Gaussians, providing more flexibility and
allowing us to easily approximate more general participating
media transport than possible with QD by combining with
other Monte Carlo rendering methods.

1.1. Previous Work

Diffusion-based Methods. The diffusion equation can
be solved using finite element methods [Sta95, AWB11,
WZT 08] which can handle heterogeneous materials and
complex boundaries, but which require building the finite elements and solving the diffusion equation numerically, which
is usually impractical for complex scenes.

Monte Carlo Methods. Concurrently to the work on

diffusion-based subsurface scattering methods, there have
been significant advances in Monte Carlo rendering techniques for general participating media. We focus only on the
most relevant previous work; a more comprehensive overview
is provided by Pharr and Humphreys [PH10].

Drawing on work from classical diffusion theory [FPW92],

Jensen et al. [JMLH01] introduced the diffusion dipole to
computer graphics and applied it to the rendering of translucent materials. Later, Donner and Jensen [DJ05] extended
this to a multipole to accurately handle thin and multi-layer
materials.

Variants of path tracing can handle arbitrary numbers of

bounces, but at a significant cost and with high variance. Donner et al. [DLR 09] used brute-force Monte Carlo particle
tracing to tabulate an empirical BSSRDF model which handles oblique illumination and a broader range of albedo values than possible with diffusion. Kulla and Fajardo [KF12]
recently proposed equiangular and decoupled importance
sampling strategies to reduce noise when path tracing singlescattering. We employ equiangular sampling, but for computing multiple scattering with diffusion.

All of these methods use the classical diffusion theory

approximation, which suffers from significant errors in nearsource and high-absorption regimes. Recently, dEon and
Irving [dI11] introduced, from other fields, a wealth of improvements to diffusion theory to address these problems,
notably: 1) a modified diffusion-type equation from neutron
transport [Gro56, Gro58] which explicitly decouples singlyand multiply-scattered light, and 2) a more accurate exitance
calculation from optics [KP97]. DEon and Irving also promoted the use of an extended source term instead of approximating it as an impulse at a single depth like the dipole.
Since no closed-form solution exists to the extended source
integral, and available numerical approaches were expensive,
they approximated the resulting diffusion profile as a sum
of Gaussians. To avoid fitting the Gaussians [dLE07], dEon
and Irving further exploited the fact that time-resolved or
quantized diffusion (QD) results in a Gaussian distribution
and used this as a mathematical basis for finding the Gaussian weights. Yan et al. [YZXW12] later extended the sum of
Gaussians to handle oblique illumination from environment
maps.

Monte Carlo algorithms that simulate multiple scattering

more efficiently typically rely on some form of light path
caching. Jensen and Christensen [JC98] generalized photon
mapping to participating media by caching photons in volumes. Later, Jarosz et al. [JZJ08, JNSJ11] showed significant
quality and performance improvements by considering the
entire length of camera rays (the beam radiance estimate) or
light paths (photon beams) when performing density estimation. This generalization relates to track-length estimators
used in the neutron transport field for some time [Spa66].
Many-light methods [KHA 12] turn photons into virtual
light sources instead of using density estimation. The rendering pass then simulates an extra bounce of illumination
by connecting these lights to shading points using visibility
queries. Recently, Novak et al. [NNDJ12] developed an improved many-light method for media which, instead of turning photons into virtual point sources, turns photon beams
into extended virtual ray light (VRL) sourcesproviding
faster convergence and reduced singularities for multiple scattering. All of these Monte Carlo methods can simulate more
general media than diffusion-based methods (accounting for
c 2013 The Author(s)

c 2013 The Eurographics Association and Blackwell Publishing Ltd.

R. Habel, P. Christensen & W. Jarosz / Photon Beam Diffusion

Table 1: Notation used throughout this paper

Symbol

Description

L(~x,~)
fs (~,~0 )
g
s
0s
a
t
t0

0
(~x)
~E(~x)
Q(~x)
D
tr

Radiance at position ~x from direction ~ [W m2 sr1 ]

Phase function (normalized to 1/4)
[sr1 ]
Average cosine of scattering

Scattering coefficient
[m1 ]
Reduced scattering coefficient: (1 g) s [m1 ]
Absorption coefficient
[m1 ]
Extinction coefficient: s + a
[m1 ]
Reduced extinction coefficient: 0s + a
[m1 ]
Relative index of refraction

Scattering albedo: s /t

Reduced scattering albedo: 0s /t0

Fluence
[W m2 ]
Vector flux (vector irradiance)
[W m2 ]
Source function
[W m3 ]
Diffusion coefficient p
[m]
Transport coefficient: a /D
[m1 ]

Units

importance sampling, eliminating the need for complex,

numerically unstable quantization.
We propose a new diffuse single-scattering term which

exactly matches expensive Monte Carlo simulations.

Our formulation of single and multiple scattering trivially

allows for oblique illumination.

We introduce an empirical correction factor which im-

proves the accuracy of the modified diffusion model for

near-source and oblique illumination regions.
We demonstrate our method for both traditional subsur-

face scattering rendering and applied on top of Monte

Carlo photon beam tracing, combining the complementary strengths of photon beams, virtual ray lights, photon
diffusion, and the improved diffusion model.
2. Light Transport in Scattering Media
Light transport in a participating medium can be described
by the radiance transport equation (RTE) [Cha60]:

heterogeneous media, anisotropic scattering and lighting distribution, and occlusions), but at generally higher cost, especially for highly scattering materials. As the name photon
beam diffusion suggests, we take inspiration from this line
of work by similarly considering a continuous beam of
photons, but we simulate multiple scattering from this beam
using diffusion instead of using density estimation or recursive Monte Carlo integration.
Hybrid Methods. A few hybrid methods have combined
diffusion and Monte Carlo techniques. Li et al. [LPT05]
proposed a hybrid approach which performs a few bounces
of path tracing from the eye, and approximates subsequent bounces with the classical dipole. Donner and
Jensens [DJ07] photon diffusion (PD) performs a photon
tracing pass and then interprets the photons as diffusion
sources. Like QD, this accounts for the extended incident
beam of light, and further accounts for oblique incident illumination and internal blockers (approximately), but it forfeits
accurate multilayered material support and uses the inferior
classical diffusion model. We can likewise compute diffusion from photon sources, but operate on continuous photon
beams [JNSJ11] instead of discrete photon points. This is
similar in principle to many-light methods, but instead of
computing exact one-bounce illumination from the photon
sources, these methods compute approximate multi-bounce
illumination using the diffusion approximation from photon
points (for PD) or from photon beams (for our method, PBD).
1.2. Contributions
Our photon beam diffusion approach provides improvements
across this wide variety of previous methods. Specifically:
We provide a more efficient and numerically stable way to

integrate the extended source function using Monte Carlo

c 2013 The Author(s)

c 2013 The Eurographics Association and Blackwell Publishing Ltd.

~
(~ )L(~
x,~) = t L(~x,~) + Q(~x,~)
Z

+s

(1)

L(~x,~) fs (~,~0 ) d~0 ,

which states that the change in radiance L in direction ~ is

the sum of three terms: a decrease in radiance dictated by the
extinction coefficient (t = s + a ), and an increase due to
the source function Q and the in-scattering integral on the
second line. We summarize our notation in Table 1. In the
most general form, simulating light transport in translucent
materials requires solving the RTE (accounting for the scattering within the medium) with suitable boundary conditions
imposed by the enclosing surfaces.
2.1. The BSSRDF and the Searchlight Problem
When rendering translucent materials, it is often convenient
to re-formulate this problem in analogy to the local surface
reflection integral. This results in an integral equation which
computes the outgoing radiance, Lo , at position and direction
(~xo ,~o ) as a convolution of the incident illumination, Li , and
the BSSRDF, S, over all incident positions and directions
(~xi ,~i ):
Z Z

Lo (~xo ,~o ) =

S(~xi ,~i ;~xo ,~o )Li (~xi ,~i )(~n ~i ) d~i dA(~xi ). (2)

A 2

For efficiency, S is often decomposed into reduced radiance,

single scattering, and multi-scattering terms, S = S(0) + S(1) +
Sd , so that each can be handled by specialized algorithms.
Most previous techniques in graphics have approximated S(1)
with the refractive single-scattering approximation proposed
by Jensen et al. [JMLH01]. Simulating the multi-scattering
term Sd accurately is expensive, so techniques in fields such as
medical physics, astrophysics, and recently computer graphics have relied on simplifications to this general problem
based on approximate solutions to the so-called searchlight
problem.

R. Habel, P. Christensen & W. Jarosz / Photon Beam Diffusion

"
0 zr (t) (1 + tr dr (t)) etr dr (t)
The searchlight problem, Figure 1(a), considers a simpli+
R~dE (~x,t) = C~E
4
dr3 (t)
fied setting where a focused beam of light is incident at the

origin on a semi-infinite planar homogeneous medium. Photons refract through a Fresnel boundary and travel in the
downward direction until they are scattered by the medium,
and ultimately get absorbed or escape the material. The distribution of photons exiting the upper boundary forms a spatial
reflectance profile Rd (~x), which is radially symmetric (1D) for
normally-incident light, Rd (~x) = Rd (k~xk). General solutions to
the searchlight problem typically rely on brute-force Monte
Carlo particle tracing for validation, though exact solutions
do exist for special cases.

#
(zr (t) + 2zb ) (1 + tr dv (t)) etr dv (t)
,
dv3 (t)

where we use the shorthand dr (t) = d(~x,~xr (t)) and analogously

for dv , and where zr (t) =~xr (t) ~n is the depth of the real source.
2.3. Quantized Diffusion
DEon and Irving noted that Equation (4) has no closed-form
solution, and proposed to approximate it using a sum of
Gaussians:

Most methods in graphics simplify Sd as a product of a

typically 1D, diffuse reflectance profile Rd and a directional,
energy-preserving Fresnel reshaping term [JMLH01, dI11]:
1
Ft (~xo ,~o )
Ft (~xi ,~i )Rd (~xo ~xi )
,

4C (1/)

Rd (~x) 0

where x = k~xk, (wR (i) wi ) are the weights, and vi are the variances of normalized 2D Gaussians G2D . The equations necessary to obtain the weights and variances of the Gaussians are
themselves complex summations of integrals depending on
further weights wR (v, i), w~ER (v, i), which we omit here.

where the reflectance profile is now centered at the incident

light position, ~xi , and C is a constant needed for normalization. To evaluate this expression efficiently, most methods
have relied on diffusion theory to obtain an analytic approximation for Rd .

The reflectance profile Rd can be obtained by considering an

exponentially-decreasing source beam within the material and
expressing its scattering and propagation using the diffusion
approximation. This results in the following integral:
Z

r(~x,~xr (t)) Q(t) dt,

(4)

where Q(t) = 0 t0 et t is the extended source function, and

r(~x,~xr (t)) = Rd (~x,t) + R~d (~x,t) is the radiant intensity on the
E
surface due to a dipole along the beam at ~xr (t) = t~ (see
Figure 1(b)). The radiant intensity depends on fluence and
vector irradiance terms, defined as:
Rd (~x,t) = C

0
4D

etr dr (t) etr dv (t)

dr (t)
dv (t)

!
,

We found that there are several practical downsides of

solving Equation (4) in this manner. Evaluating Equation (7)
is non-trivial, relatively expensive, and prone to error since
the formulas provided for the weights are highly susceptible
to numerical instabilitiesspecifically, when incorporating
the extinction of the source beam into the Gaussians weights.
Figure 2 shows a plot of the weights where the instability
occurs.
Fluence Weights( s = 1, a = 0.001, = 1.4)

Flux Weights( s = 1, a = 0.001, = 1.4)

0.08

0.04
Expected Behaviour
Quantized Diffusion

Expected Behaviour
Quantized Diffusion

wR

0.06

0.03

0.04

0.02

0
46

0.02

0.01

47

48

49
Variance

50

51

52

0
46

47

48

49
Variance

50

51

52

Figure 2: Gaussian weights computed as described by dEon

and Irving [dI11] with double precision (blue) show strong
instabilities compared to the expected behavior (green).

2.2. Improved Diffusion Model and Extended Source

Rd (~x,~) =

(7)

(wR (i) wi ) G2D (vi , x),

i=0

(3)

For a full review of diffusion theory and the improvements

which were previously described in tandem with the proposed
solution using quantized diffusion and Gaussians [dI11],
please refer to our supplemental technical report [HCJ13]. In
the following, we only present a short description resulting
in the integral from the improved diffusion model that needs
to be solved. We then show our solution using a Monte Carlo
approach, which is independent of the underlying model used.
Furthermore, we show how the same integration scheme can
be used to efficiently compute the exact single-scattering
from the searchlight problem and use it to define a new diffuse single-scattering term S(1) .

k1

wER

Sd (~xi ,~i ;~xo ,~o ) =

(6)

(5)

Though this problem can be circumvented by deriving

functional approximations (kindly provided to us by the authors [dI11]), this adds to the already high complexity of
the method. Moreover, Gaussian functions are actually a
poor representation for diffusion profiles, and therefore many
Gaussians are needed for an accurate reconstruction. Due
to the complexity of the calculations and the need for many
Gaussians, a quantized diffusion profile is roughly an order
of magnitude more expensive to evaluate than the standard
dipole.
On the other hand, the profiles resulting from quantized diffusion are significantly more accurate than the standard dipolesee Figure 3. The downsides are due to the
c 2013 The Author(s)

c 2013 The Eurographics Association and Blackwell Publishing Ltd.

R. Habel, P. Christensen & W. Jarosz / Photon Beam Diffusion

Multiple Scattering
s = 1.0, a = 10, ( = 0.090909), = 1.33

s = 1.0, a = 0.1, ( = 0.90909), = 1.33

s = 1.0, a = 0.01, ( = 0.9901), = 1.33

10

Monte Carlo
Jensen Dipole
Quantized Diffusion
Ours (5 samples)

10

0.1

0.2

0.3

0.4

0.5

10

Monte Carlo
Jensen Dipole
Quantized Diffusion
Ours (5 samples)

log (xRd (x))

log (xRd (x))

Monte Carlo
Jensen Dipole
Quantized Diffusion
Ours (5 samples)

10
log (xRd (x))

Monte Carlo
Jensen Dipole
Quantized Diffusion
Ours (5 samples)

10

10

10

10

log (xRd (x))

s = 1.0, a = 1, ( = 0.5), = 1.33

0.5

1.5
x

2.5

0.5

1.5
x

2.5

10

0.5

1.5
x

2.5

Single + Multiple Scattering

10

s = 1.0, a = 0.1, ( = 0.90909), = 1.33

10
Monte Carlo
Jensen Dipole
Quantized Diffusion
Ours (5 samples)

log (xRd (x))

Monte Carlo
Jensen Dipole
Quantized Diffusion
Ours (5 samples)

10
log (xRd (x))

s = 1.0, a = 1, ( = 0.5), = 1.33

10

10

10

s = 1.0, a = 0.01, ( = 0.9901), = 1.33

10
Monte Carlo
Jensen Dipole
Quantized Diffusion
Ours (5 samples)

10

Monte Carlo
Jensen Dipole
Quantized Diffusion
Ours (5 samples)

log (xRd (x))

s = 1.0, a = 10, ( = 0.090909), = 1.33

log (xRd (x))

10

10

10

0.1

0.2

0.3

0.4

0.5

10

0.5

1.5
x

2.5

0.5

1.5
x

2.5

10

0.5

1.5
x

2.5

Figure 3: Comparison of the standard dipole, QD, and our method (with correction) validated against Monte Carlo simulation
for multiple scattering only (top row) and combined with single scattering (bottom row). Our approach matches the quantized
diffusion profile, but uses only 5 samples compared to 66 Gaussians. By incorporating our correction factor, we are able to
surpass the fidelity of QD, avoiding divergence near the source and matching Monte Carlo much more faithfully.

quantization procedure itself (which is orthogonal to the

other improvements to diffusion imported to graphics by
dEon and Irvingsee also the recent work by dEon and
Hery [dE12, Her12]). This motivated us to leverage the improved model but solve Equation (4) in a different way.

lows us to efficiently integrate a new diffuse single-scattering

term motivated by the searchlight problem.
3.1. Importance Sampling
To evaluate Equation (8), we can use any number of importance sampling strategies to distribute the numerical samples.

3. Method
The core of our method is to numerically compute the extended source integral (4) using Monte Carlo integration with
importance sampling:
Rd (~x,~)

1
N

f (~x,~,ti )

pdf(ti |~x,~) ,

(8)

i=1

where f (~x,~,ti ) = r(~x,~xr (ti )) Q(ti ) is the integrand from (4) and
pdf(ti |~x,~) is the PDF of choosing ti given the shading location ~x and the beam direction ~.
Computing the integral in this way provides many practical
advantages. Firstly, it allows us to define our set of effective
basis functions (the summands of the above equation) by
carefully choosing the PDF. This allows us to obtain accurate
results with as little as N = 3..5 samples while supporting
asymmetric profiles from oblique incident illumination (see
Figure 1(b,c)), finite slabs, and multilayer materials. We also
introduce a correction term which improves the accuracy of
the underlying diffusion model near the incident beam, especially for highly scattering materials. This improvement is not
possible to incorporate into QD, but fits trivially within our
more flexible integration scheme. Finally, our approach alc 2013 The Author(s)

c 2013 The Eurographics Association and Blackwell Publishing Ltd.

Exponential Sampling. The traditional approachused by

path tracing, photon mapping, and photon diffusion for
choosing propagation distances in homogeneous mediais
exponentially-decreasing sampling with a PDF proportional
to the source term Q(t):
ti =

log(1 i )
t0

with

pdfexp (ti ) = t0 et ti ,

(9)

where i [0, 1) is a uniform random number. This PDF only

depends on t , and is independent of ~x and ~. This effectively
places dipole sources at different depths and weights them
according to the inverse exponential PDF, canceling out the
exponential variation in Q(t).
Equiangular Sampling. A different strategy is, at each ~x, to
distribute the samples uniformly in the angular domain subtended by the incident beam (Figure 1(b)). This corresponds
to the equiangular sampling method proposed by Kulla and
Fajardo [KF12] and also used by Novk et al. [NNDJ12].
This gives:
ti = h tan (i ) with pdfequ (ti |~x,~) =

h
(10)
(b a )(h2 + ti2 )

R. Habel, P. Christensen & W. Jarosz / Photon Beam Diffusion

Exponential Sampling ( s = 1.0, a = 0.1, = 1.33)
s=1
s=3
s=5
exact

the strategies. Using the standard balance heuristic would

draw samples from both strategies across the entire ~x domain.
This is wasteful and would not guarantee proper asymptotic
tail behavior. Luckily, we always know which strategy works
better in which region (tail vs. peak) and use this knowledge
to develop a modified MIS weighting heuristic, tailored to
our specific problem. Our estimator becomes:

s=1
s=3
s=5
exact

10
log Rd (x)

log Rd (x)

10

Exponential Sampling ( s = 1.0, a = 0.1, = 1.33)

10

10

10
0

10

15

0.2

0.4

0.6

0.8

Equiangular Sampling ( s = 1.0, a = 0.1, = 1.33)

Equiangular Sampling ( s = 1.0, a = 0.1, = 1.33)

s=1
s=3
s=5
exact

s=1
s=3
s=5
exact

10
log Rd (x)

log Rd (x)

10

Rd
+

10
5

10

15

0.2

0.4

0.6

0.8

MIS Sampling ( s = 1.0, a = 0.1, = 1.33)

MIS Sampling ( s = 1.0, a = 0.1, = 1.33)

s=1
s=3
s=5
exact

s=1
s=3
s=5
exact

10
log Rd (x)

log Rd (x)

10

f (~x,~,ti )wexp (ti , x)

)
exp (ti ) + weq (ti , x) pdfeq (ti |~x,~

i=1

1
N

wexp (t j , x) pdf

wexp (ti , x) pdf

N

(11)

f (~x,~,t j )weq (t j , x)
,
)
exp (t j ) + weq (t j , x) pdfeq (t j |~x,~

j=1

10

10

1
N

10

where wexp (ti , x) = clamp((x a)/(b a), 0, 1) and weq (ti , x) =

1 wexp (ti , x). Our weighting strategy dedicates samples entirely to the superior PDF except in a small transition zone
(x [a, b]) where we smoothly weight the two strategies. We
center the transition zone around one diffuse mean free path,
with a = 0.9/t0 and b = 1.1/t0 . This approach can be interpreted as deriving domain-specific representativity for the
two strategies used in the MIS estimator [PBPP11].

10

10
0

10
x

15

0.2

0.4

0.6

0.8

Figure 4: The profiles (left) and close-up of the peak (right)

of the sampling methods for different numbers of samples.
Exponential sampling always introduces some smoothing
near x = 0 while equiangular sampling cannot match the
tail behavior. Our combined MIS strategy (bottom) matches
both. Profiles for other absorption values are available in the
supplemental document.
where h is the distance between ~x and the beam ~, a and b
are the start and end angles of the integration in the angular
domain, and i = lerp(a , b , i ).
Discussion & Comparison. Standard Monte Carlo integration uses uniform random numbers for the i s. To avoid
per-pixel noise in our profiles, we instead use a deterministic
regular sequence i = i0.5
. This effectively creates a custom
N
quadrature rule which is defined by the positions and weights
of the importance sampling method used, directly using information from the integral through the scattering parameters.
We found that this produces superior results and achieves
high accuracy with a very low number (35) of samples.
Figure 4 (top and middle) compares the convergence behavior of exponential and equiangular sampling. In the exponential case, the tail of the profile is guaranteed to converge
when taking just 1 sample. On the other hand, equiangular sampling represents the peak well even with 1 sample,
whereas exponential sampling underestimates the peak.
Combining Strategies. Given two sampling strategies that
perform well in different situations the standard solution is to
use multiple importance sampling [VG95] (MIS) to combine

In Figure 4 (bottom) we show how our tailored MIS strategy, with just 3 samples, matches the entire profile, including
the peak and the asymptotic tail behavior.
3.2. Correction Factor
Donner and Jensen [DJ07] noticed an overestimation of exitance near the incident light source when using the classic
dipole and proposed an empirical correction factor (t) =
1 et t that diminishes near-surface regions of Q(t). DEon
and Irving [dI11] instead used the improved model with
Kienle-Patterson exitance calculations and showed that the
resulting profiles were consistently more accurate. However,
we found that this improved model also suffers from overestimation for near-surface sources. Figure 3 (top) shows that
QDs multiple scattering in fact diverges as ~x approaches 0,
though the finite number of Gaussians avoids an actual singularity at ~x = 0. To account for this, we introduce an empirical
correction factor inspired by Donner and Jensens, but tailored for the improved diffusion model:
(~x,t) = 1 e2t (d(~x,~xr (t))+t) .

(12)

In Figure 3 we compare our approach to both the classic

dipole and to QD (using 66 Gaussians) for a semi-infinite
medium with fixed s = 1 and three orders of magnitude of
absorption levels. We also include a ground truth computed
using our own brute-force Monte Carlo particle tracing code,
which we first validated against the well-tested MCML program [WJZ95].
While QD strongly overestimates at ~x 0, our profile
with the correction factor matches the shape of the Monte
Carlo reference profile well over the entire ~x domain and
absorption parameter range. It is important to point out that
since our correction factor depends not only on t , but also
c 2013 The Author(s)

c 2013 The Eurographics Association and Blackwell Publishing Ltd.

R. Habel, P. Christensen & W. Jarosz / Photon Beam Diffusion

Multilayer materials can be handled in the same way as

Donner and Jensen. Since Gaussians naturally support convolution required for multi-layer materials, QD admittedly
holds an advantage here over our method or Donner and
Jensens original multipole formulation.

Monte Carlo Multi-Scattering

PBD Multi-Scattering
QD Multi-Scattering (no )
Better Dipole [dEon 2012]
Classic Dipole [Jensen 2001]
Monte Carlo Single Scattering

3.4. Oblique Incident Light Angles

Diffuse Single Scattering

Similar to Yan et al. [YZXW12], who formulated QD for

oblique angles by offsetting the centers of the Gaussians,
we can apply our approach to oblique light angles. Due to
the explicit integration formulation, Equation (8) trivially allows us to choose any ~ that is not orthogonal to the surface
normal without any modifications. We simply compute locations along the refracted incident beam during numerical
integration as shown in Figure 1(c).

on the shading location ~x (indirectly through its dependence

on dr = d(~x,~xr (t))), it is impossible to incorporate it in the
sum-of-Gaussians approach used by QD. Our Monte Carlo
integration strategy, however, makes incorporating this correction factor trivial.

Reflectance Rd (x)

Transmittance Td (x)

s = 1.0, a = 0.1, = 1.33, d = 0.1

log (xTd (x))

log (xRd (x))

60

80

60

80

60

80

QD PD
PD

Single Scattering
0

60

80

60

80

s = 1.0, a = 0.1, = 1.33, d = 0.1

10
Monte Carlo
Jensen Dipole
Quantized Diffusion
Ours

10

Monte Carlo
Jensen Dipole
Quantized Diffusion
Ours
2

10

10

80

Ours

Handling finite slabs with our numerical integration scheme

is straightforward. We simply treat each sample (summand
of Equation (8)) as a dipole and mirror it at the extrapolated
boundary to perform a multipole extension for a finite slab
as in Donner and Jensen [DJ05]. This has the same effect as
the mirroring of Gaussians in QD. The equiangular sampling
in our MIS strategy avoids placing dipoles directly on the
reflective as well as transmissive side of the slab when close to
~x = 0 which allows for very thin slabs without any drawbacks.
Furthermore, our correction factor significantly improves
reflectance as well as transmittance compared to the improved
diffusion model for finite slabs as shown in Figure 6 below.

60

MC

3.3. Finite Slabs and Multilayer Materials

10

Multiple Scattering
0

Ours

Figure 5 shows a comparison of the different methods.

The scene consists of a semi-infinite medium with the surface illuminated by three thin, bright stripes of light. We use
0s = 1, a = (0.01, 0.1, 1), = 1, g = 0 for all validation results
to show low, mid, and high absorptive media in a single color
image. We include all results as HDR images in the supplemental material and encourage the reader to view them at
different exposures.

In Figure 7 we evaluate the accuracy of our method for

oblique angles on the same medium as in Figure 5. We compare to Monte Carlo for validation and we also compare to
PD with photons deposited using an exponential distribution [DJ07] for both the original PD formulation as well as
PD with the improved diffusion model. Steep incident angles
pose problems due to both the integration/sampling method

MC

Figure 5: Our method matches ground truth multi-scattering

(top) better than previous approaches, including QD (no
correction factor), better dipole, and classic dipole. Our diffuse single scattering (bottom) computed with just 5 samples,
matches Monte Carlo reference faithfully.

10
0.5

1.5

0.5

1.5

Figure 6: Reflectance (left) and transmittance (right) of a

finite slab with thickness d = 0.1/s .

c 2013 The Author(s)

c 2013 The Eurographics Association and Blackwell Publishing Ltd.

Figure 7: An incident light beam (with radius 0.2/s ) illuminates a semi-infinite medium at angles 0 , 60 and 80
degrees from normal incidence. We compare Monte Carlo
reference (top), our multi-scattering (middle) using 50 beams,
and photon diffusion [DJ07] with the improved model (QD
PD) and classic model (PD) using an equivalent number of
samples (50 5 = 250 photons). The bottom compares Monte
Carlo single-scattering reference to our solution with 5 samples. (Parameters: 0s = 1, a = (0.01, 0.1, 1), = 1, g = 0.)

R. Habel, P. Christensen & W. Jarosz / Photon Beam Diffusion

and approximation in the diffusion model. Our results match

Monte Carlo references accurately at a wide variety of incident angles and do not break down even at extreme grazing
angles (see supplemental material). The improvement in overall intensity and color compared to photon diffusion is due
to the improved diffusion model and our correction factor,
which becomes increasingly important for avoiding overestimation at steep incident angles. Also, even with the same
number of total dipole evaluations, our method benefits from
better sample placement which adapts the photon locations
to each shading location.
3.5. Diffuse Single Scattering
Single scattering is an important part of translucent material
appearance, especially near the incident light and at high
absorption. Jensen et al. [JMLH01] derived an approximate
single-scattering term assuming smooth Fresnel boundary
conditions. Unfortunately, this usually results in a very dim
contribution, and is therefore often omitted from production
implementations. However, since the improved model separates out single scattering we need to account for it explicitly.
Our goal is to include the same diffuse single-scattering
term computed by brute-force Monte Carlo particle tracing
simulations such as MCML [WJZ95]. This type of singlescattering term has typically been used only in validation
graphs, but not incorporated in any practical rendering algorithm since using brute-force particle tracing during rendering would be prohibitively expensive. However, since we are
explicitly integrating along the incident beam using Monte
Carlo integration, we can efficiently and trivially include
single scattering in the same way as the multi-scattering.
In fact, the only difference is that the Greens function
(the r(~x,~xr (t)) term in the integrand of Equation (4)) needs to
change to a single-scattering response:
r(1) (~x,~xr (t)) =

fs (~ ~~xr~x )et (dr (t)) Ft (o , )Ft (i , 1/) cos o

dr2 (t)

(13)
where ~~xr~x and dr (t) are the direction and distance from ~xr (t)
to ~x respectively, o = arccos(~~xr~x ~n), i = arccos(~ ~n), and
Ft and fs are the transmissive Fresnel and phase function
terms. Figure 1(b,c) illustrates the configuration.
We can then compute the single-scattering profile,
R(1) (~x,~), using the same Monte Carlo estimator (8), but with
f (~x,~,ti ) = r(1) (~x,~xr (ti )) Q(ti ). The single-scattering portion of
the BSSRDF, S(1), can then be defined in direct analogy to
Sd (3), but using R(1) instead of Rd . We can compute the
two profiles simultaneously (using the same points or rays),
adding minimal overhead to the calculations. However, if the
surface separates volumes with different indices of refraction,
we can use a separate sample sequence for single-scattering,
starting the integration not from the incident light angle, but
from the critical angle since contributions below the criti-

cal angle are zero. This optimization guarantees a smooth

single-scattering profile regardless of the number of samples.
In contrast to multi-scattering using diffusion, the singlescattering integral is exact under the assumed planar configuration and boundary conditions (Figure 1). In particular,
the phase function and Fresnel terms are accounted for explicitly instead of approximately using similarity theory and
the reflection parameter. Also, since we calculate a single
bounce, no negative sources (dipoles) appear in the integrand.
It is worth pointing out that this equation is mathematically
similar to computing the diffuse illumination on a surface
due to a VRL [NNDJ12] but with the addition of Fresnel
terms. Unlike the refractive single-scattering term of Jensen
et al. [JMLH01], which is highly directional and dim, our
diffuse single scattering is a significant part of the overall
radiant exitance profile (see Figures 3, 5, and 7).
Compared to single scattering at a Fresnel boundary as
described by Jensen et al. [JMLH01], we assume the exitant
boundary is diffusethough not Lambertian, since we reshape the exitant radiance by a Fresnel term as done with
multi-scattering. Jensen et al. on the other hand approximate
the length of the incident beam inside the medium with a
shadow ray that ignores refraction. Both of these are approximations. Jensen et al.s single scattering approach is likely
more amendable to media such as marble or milk with a
smooth Fresnel boundary while our model caters more to
skin or other materials with rough boundaries. A full investigation of the relative strengths of these two approaches would
be an interesting avenue for future work.
4. Implementation and More Results
In addition to the scientific results for validation in the previous sections, we additionally evaluated photon beam diffusion
on practical scenes. To demonstrate the versatility of our approach, we incorporated it into several rendering systems and
light transport algorithms, including a point-based subsurface
scattering implementation in Mitsuba [Jak10], point-based
and ray-traced subsurface scattering in PRMan, and on top
of a progressive photon beam tracer [JNT 11] to simulate
oblique illumination.
4.1. Subsurface Scattering with PBD Profiles
Our Mitsuba point-based implementation first distributes surface points using point repulsion and writes them in a point
cloud file. It then constructs an octree of the points, and hierarchically traverses the octree [JB02]. Figure 8 shows a direct
comparison between volumetric path tracing, photon beam
diffusion, and dipole diffusion rendered in Mitsuba.
PRMans implementation of subsurface scattering encompasses both point-based and ray-traced approaches. For the
point-based approach, the object surfaces are first tessellated
into micropolygons and shaded, and the shaded micropolygons are written as points in a point cloud file. PRMan then
c 2013 The Author(s)

c 2013 The Eurographics Association and Blackwell Publishing Ltd.

R. Habel, P. Christensen & W. Jarosz / Photon Beam Diffusion

ray tracing or octree traversal, we use the ray length or the

distance to the octree point or cluster and interpolate between
the two nearest cached profile values.
Though caching alleviates evaluation time impact on the
overall rendering time, efficient generation of profiles is important if the scattering parameters change over the surface
and a new profile needs to be generated for each shading
point. With the QD weight calculation and the evaluation of
the Gaussians, we achieve a speedup factor of 3.1 for a full
profile, comparing 5 photon beam diffusion samples with 44
Gaussians. In the less common case where only one sample
of a profile is needed, we achieve a speedup of 21.6.
Figure 8: Comparison between volumetric path-tracing (top
left), our method (top right), our method with multiple scattering only (bottom left), and dipole diffusion (bottom right). The
dipole produces color shifts and over-smoothing due to an
inaccurate profile. (Parameters: 0s = 1, a = (0.01, 0.1, 1), =
1, g = 0.)
constructs an octree over the points, and traverses the it hierarchically during rendering [JB02]. Ray-traced subsurface
scattering at a shading point is computed by selecting a point
inside the object (below the shading point) and shooting
rays in all directions. We compute the area, incident illumination, and subsurface contribution for each ray hit point.
The incident illumination can be cached in a multiresolution
irradiance cache for efficiency [CHS 12].
The interface and parameters of our method are the same
as for the classic dipole, though the different methods cause
a few practical issues that we address in the following.
4.1.1. Efficient Profile Evaluation
Computing high-quality subsurface scattering at a point requires evaluating the BSSRDF at a few hundred distances
either for subsurface ray hit points or for point and cluster
positions. Evaluating the BSSRDF hundreds of times for each
shading point is relatively slow.
For efficiency, dEon and Irving [dI11] used binning to
reduce the number of BSSRDF evaluations: they deposit
cache point energy into discrete radial bins during hierarchical traversal, and then multiply each bin by the integrated profiles to accumulate the final energy. This is relatively efficient
and accurate since their Gaussian profile can be analytically
integrated within each bin. We take a dual approach by tabulating the profile itself, to achieve a similar effect without
requiring bin integration.
Like dEon and Irving, for each shade point, before traversing the point hierarchy or shooting rays, we choose an
exponentially-varying set of distances x = 0, x1 , x1 , x1 2 ,
x1 3 , . . ., with a step size of 1.2 and x1 0.01/t . But instead of using these distances as discrete bins, we evaluate
and cache the reflectance profile at these distances. During
c 2013 The Author(s)

c 2013 The Eurographics Association and Blackwell Publishing Ltd.

4.1.2. Parameter Conversion

For practical applications, it is more user-friendly to parameterize the BSSRDF by diffuse albedo Rd (diffuse reflection
coefficient) and diffuse mean free path length (1/t0 ) than by
0s and a . For dipole diffusion, Jensen and Buhler [JB02]
derived an expression for Rd given 0 and noted that although
the expression isnt analytically invertible it is monotonically
increasing so the inversion can be done using secant root finding. We have developed a more efficient table-based method
for this inversion.
For a given index of refraction, we first generate a table of
Rd values for values of 0 [0, 1). Next, the table is inverted,
resulting in a table of 0 values for values of Rd . Interpolating
between two table entries, we can determine 0 while avoid-

ing expensive profile evaluations. Together with the diffuse

mean free path, we then get 0s and a .
If we had used secant root finding, the Rd evaluations required for photon beam diffusion would be more expensive
than for dipole diffusion: for dipole diffusion, computing Rd
given 0 only requires evaluating an analytic formula, while
for photon beam diffusion we need to perform a numerical
integration. But fortunately the same table inversion method
can be used for both diffusion models, making their parameter conversion times nearly identical. Thanks to these table
lookups and efficient profile evaluations, rendering with photon beam diffusion and single-scattering typically takes only
about 15% longer than using the classic dipole.
Figure 9 shows a human head with ray-traced dipole diffusion compared against photon beam diffusion with and
without single-scattering. Subsurface scattering with photon
beam diffusion is much sharper than dipole diffusion around
skin pores and bumps on the lips, and single scattering adds
further sharpness. A turntable animation of the head and a
complex scene rendered with photon beam diffusion can be
found in the supplemental material.
4.2. Applying PBD to Photon Beams
In addition to standard subsurface scattering rendering which
convolves the diffusion profile with the direction-less surface

R. Habel, P. Christensen & W. Jarosz / Photon Beam Diffusion

Our Method (Final)

Photon Beams (SS)

Photon Beam Diffusion (MS)

15 seconds
Photon Beam Diffusion (MS)

600 seconds
MC reference: Virtual Ray Lights (MS)

15 seconds
Photon Diffusion (MS)

Figure 10: Photon beam diffusion shows a faster multiple scattering convergence than PD due to the equiangular/exponential
MIS sampling, especially at oblique light angles (see insets). The left image shows our multiple scattering (MS) combined with
single scattering (SS) calculated using photon beams.
photon locations along the beams for each shading location
using equiangular sampling. Figure 10 shows a comparison
of the ground truth, photon diffusion, and our method, showing a faster convergence with less noise with our equiangular
diffusion sampling.
5. Conclusion and Future Work

Classic Dipole

Photon Beam Diffusion

PBD + Single Scatter

Figure 9: The classic dipole model (left) can lead to an overly

waxy appearance for human skin. Our PBD model (middle)
produces sharper profiles, and additionally incorporating our
diffuse single scattering term (right) preserves pore details
even better. Render times: 7 minutes on a multi-core PC.

irradiance, we have applied our improved profiles and numerical integration scheme on top of a photon beam tracer
to simulate diffusion from directional, oblique illumination.
Our implementation operates much like PD [DJ07], but we
store continuous, directional photon beams entering the surface instead of discrete photon points inside the medium (see
Figure 1(e) and compare to 1(d)). We use the stored photon
beams as extended beam sources and evaluate the multiple
scattering using our integration scheme and the improved diffusion model. Applying photon beam diffusion to oblique illumination on arbitrary geometries requires more careful handling of the boundary conditions. We blend between dipole,
multipole, and quadrupole boundary conditions as proposed
by Donner and Jensen [DJ07]. Our Monte Carlo integration
scheme (8) makes the link between these two explicit: each
Monte Carlo sample of the beam integral can be viewed as
a weighted photon at ~xr (ti ) inside the medium (a monopole),
which we can mirror exactly as in PD to compute the three
boundary conditions. By storing photon beams instead of discrete points during photon tracing, we can compute different

We have presented a hybrid methodphoton beam diffusion

that combines Monte Carlo integration with the diffusion
approximation. We can obtain the same improved subsurface
scattering results as quantized diffusion, but our method is
simpler, faster, more accurate, and numerically stable. On the
other hand, certain features such as multi-layered materials
are faster and more elegant in QD than with our explicit
integration. Our method can be used with point-based and raytraced subsurface scattering, and is only a few percent slower
than the classic dipole diffusion model. We integrated it in
both research and production renderers to validate the method.
We showed that photon beam diffusion can be combined
with more general rendering algorithms like photon beams,
minimizing artifacts associated with strongly oblique angles.
Our versatile numerical integration scheme opens up the
doors for several improvements. It is theoretically possible to
account for more general rough dielectric refraction distributions for both multi- and single-scattering, by incorporating
them directly into our numerical integration process. This
would provide a principled way to handle more complex
BSDFs in place of our current diffuse distribution at the
boundary. We also believe that a full combination of bidirectional path tracing and diffusion approximation is possible,
though the connection to the path vertices requires a more
fundamental reformulation of diffusion.
Acknowledgements. We would like to thank Eugene dEon
for explaining and clarifying QD and for providing the functional approximations. We thank Marios Papas for helpful
discussions. We thank Maurizio Nitti for the models shown in
the supplemental material and Chris Harvey for early testing
and for rendering the head images and animation. The head
data is courtesy of Infinite Realities and available via Creative
Commons.
c 2013 The Author(s)

c 2013 The Eurographics Association and Blackwell Publishing Ltd.

R. Habel, P. Christensen & W. Jarosz / Photon Beam Diffusion

References
[AWB11] A RBREE A., WALTER B., BALA K.: Heterogeneous
subsurface scattering using the finite element method. IEEE Transactions on Visualization and Computer Graphics 17, 7 (2011),
956969. 2
[Cha60]

C HANDRASEKAR S.: Radiative Transfer. Dover, 1960. 3

[CHS 12] C HRISTENSEN P. H., H ARKER G., S HADE J., S CHU BERT B., BATALI D.: Multiresolution radiosity caching for efficient preview and final quality global illumination in movies. Tech.
Rep. 12-06, Pixar, 2012. 9
[dE12] D E ON E.: A better dipole. Tech. rep., http://www.eugenedeon.com, 2012. 5
[dI11] D E ON E., I RVING G.: A quantized-diffusion model for
rendering translucent materials. ACM Transactions on Graphics
(Proc. SIGGRAPH) 30, 4 (2011), 56:156:14. 2, 4, 6, 9
[DJ05] D ONNER C., J ENSEN H. W.: Light diffusion in multilayered translucent materials. ACM Transactions on Graphics
(Proc. SIGGRAPH) 24, 3 (2005), 10321039. 2, 7
[DJ07] D ONNER C., J ENSEN H. W.: Rendering translucent materials using photon diffusion. Rendering Techniques (Proc. Eurographics Symposium on Rendering) (2007), 243252. 3, 6, 7,
10
[dLE07] D E ON E., L UEBKE D., E NDERTON E.: Efficient rendering of human skin. Rendering Techniques (Proc. Eurographics
Symposium on Rendering) (2007), 147158. 2

[JNSJ11] JAROSZ W., N OWROUZEZAHRAI D., S ADEGHI I.,

J ENSEN H. W.: A comprehensive theory of volumetric radiance
estimation using photon points and beams. ACM Transactions on
Graphics 30, 1 (2011), 5:15:19. 2, 3
[JNT 11] JAROSZ W., N OWROUZEZAHRAI D., T HOMAS R.,
S LOAN P.-P., Z WICKER M.: Progressive photon beams. ACM
Transactions on Graphics (Proc. SIGGRAPH Asia) 30, 6 (2011),
181:1181:12. 8
[JZJ08] JAROSZ W., Z WICKER M., J ENSEN H. W.: The beam
radiance estimate for volumetric photon mapping. Computer
Graphics Forum (Proc. Eurographics) 27, 2 (2008), 557566. 2
[Kaj86] K AJIYA J. T.: The rendering equation. Computer Graphics (Proc. SIGGRAPH) 20 (1986), 143150. 1
[KF12] K ULLA C., FAJARDO M.: Importance sampling techniques for path tracing in participating media. Computer Graphics Forum (Proc. Eurographics Symposium on Rendering) 31, 4
(2012), 15191528. 2, 5
[KHA 12]

K RIVNEK
J., H AAN M., A RBREE A., DACHS C., K ELLER A., WALTER B.: Optimizing realistic
rendering with many-light methods. In SIGGRAPH 2012 Courses
(2012). 2
BACHER

[KP97] K IENLE A., PATTERSON M. S.: Improved solutions of

the steady-state and the time-resolved diffusion equations for
reflectance from a semi-infinite turbid medium. Journal of the
Optical Society of America A 14, 1 (1997), 246254. 2

[DLR 09] D ONNER C., L AWRENCE J., R AMAMOORTHI R.,

H ACHISUKA T., J ENSEN H. W., NAYAR S.: An empirical bssrdf
model. ACM Transactions on Graphics (Proc. SIGGRAPH) 28, 3
(July 2009), 30:130:10. 2

[LPT05] L I H., P ELLACINI F., T ORRANCE K. E.: A hybrid

Monte Carlo method for accurate and efficient subsurface scattering. Rendering Techniques (Proc. Eurographics Symposium on
Rendering) (2005), 283290. 3

[FPW92] FARRELL T. J., PATTERSON M. S., W ILSON B.: A

diffusion theory model of spatially resolved, steady-state diffuse
reflectance for the noninvasive determination of tissue optical
properties in vivo. Medical Physics 19, 4 (1992), 879888. 2

[NNDJ12] N OVK J., N OWROUZEZAHRAI D., DACHSBACHER

C., JAROSZ W.: Virtual ray lights for rendering scenes with
participating media. ACM Transactions on Graphics (Proc. SIGGRAPH) 31, 4 (2012), 60:160:11. 2, 5, 8

[Gro56] G ROSJEAN C. C.: A high accuracy approximation for

solving multiple scattering problems in infinite homogeneous
media. Il Nuovo Cimento (1955-1965) 3 (1956), 12621275. 2

[PBPP11] PAJOT A., BARTHE L., PAULIN M., P OULIN P.: Representativity for robust and adaptive multiple importance sampling.
IEEE Transactions on Visualization and Computer Graphics 17,
8 (2011), 11081121. 6

[Gro58] G ROSJEAN C. C.: Multiple isotropic scattering in convex homogeneous media bounded by vacuum. In Proc. Second
International Conference on the Peaceful Uses of Atomic Energy
(1958), vol. 413. 2

[PH10] P HARR M., H UMPHREYS G.: Physically Based Rendering: From Theory To Implementation, 2nd ed. Morgan Kaufmann,
2010. 1, 2

[HCJ13] H ABEL R., C HRISTENSEN P. H., JAROSZ W.: Classic

and modified diffusion theory for subsurface scattering. Tech.
rep., Disney Research Zrich, 2013. 4

[Spa66] S PANIER J.: Two pairs of families of estimators for transport problems. SIAM Journal on Applied Mathematics 14, 4
(1966), 702713. 2

[Her12] H ERY C.: Texture mapping for the better dipole model.
Tech. Rep. 12-11, Pixar, 2012. 5

[Sta95] S TAM J.: Multiple scattering as a diffusion process. Rendering Techniques (Proc. Eurographics Workshop on Rendering)
(1995), 4150. 2

[Jak10] JAKOB W.: Mitsuba renderer, 2010. http://www.mitsubarenderer.org. 8

[JB02] J ENSEN H. W., B UHLER J.: A rapid hierarchical rendering technique for translucent materials. ACM Transactions on
Graphics (Proc. SIGGRAPH) 21, 3 (2002), 576581. 8, 9
[JC98] J ENSEN H. W., C HRISTENSEN P. H.: Efficient simulation
of light transport in scenes with participating media using photon
maps. Computer Graphics (Proc. SIGGRAPH) 32 (1998), 311
320. 2
[Jen96] J ENSEN H. W.: Global illumination using photon maps.
Rendering Techniques (Proc. Eurographics Workshop on Rendering) (1996), 2130. 1
[JMLH01]

J ENSEN H. W., M ARSCHNER S. R., L EVOY M., H AN P.: A practical model for subsurface light transport. Computer Graphics (Proc. SIGGRAPH) 35 (2001), 511518. 2, 3, 4,
8
RAHAN

c 2013 The Author(s)

c 2013 The Eurographics Association and Blackwell Publishing Ltd.

[VG95] V EACH E., G UIBAS L. J.: Optimally combining sampling

techniques for Monte Carlo rendering. Computer Graphics (Proc.
SIGGRAPH), 29 (1995), 419428. 6
[WJZ95] WANG L., JACQUES S., Z HENG L.: MCML: Monte
Carlo modeling of light transport in multi-layered tissues. Computer Methods in Programs and Biomedicine, 8 (1995), 313371.
6, 8
[WZT 08] WANG J., Z HAO S., T ONG X., L IN S., L IN Z., D ONG
Y., G UO B., S HUM H.-Y.: Modeling and rendering of heterogeneous translucent materials using the diffusion equation. ACM
Transactions on Graphics 27, 1 (2008), 118. 2
[YZXW12] YAN L.-Q., Z HOU Y., X U K., WANG R.: Accurate
translucent material rendering under spherical Gaussian lights.
Computer Graphics Forum (Proc. Pacific Graphics) 31, 7 (2012),
22672276. 2, 7