Global Illumination

Following Photons
• For each light, choose a number of outgoing directions (on the hemisphere or
sphere):
• Emit a photon in each direction

• Each photon travels in a straight line, until it intersects an object

• If Absorbed: Terminate the photon
• If Reflected/Transmitted/Scattered: Photon goes off in a new direction, until it
intersects an object

• When a photon goes through the camera aperture and hits a pixel (on the film), it
contributes to the color of that pixel
 Following Photons
• Most of the photons never hit the film (far too inefficient, impractical)
 Backward Tracing Photons
• For each pixel, send a ray through the aperture to backward trace a photon that
would have hit the pixel (as in ray tracing)
• If the ray hits an object, cast rays in all directions of the hemisphere in order to
backwards trace incoming photons
• Every new ray that hits another surface spawns an entire hemisphere of rays of its own
(exponential growth makes this impractical)
• Follow all rays until they hit a light source

• Once a ray hits a source, a path for photons (from the light to the pixel) has been
created
• Emit photons along this path, bounce them off all the objects along the path, check to see
if absorbed (if not absorbed, the photon continues on towards the pixel)
• The absorption of photons results in a specific color/brightness of light hitting the pixel
 Backward Tracing Photons
• Most paths take too long to find their way back to the light source (inefficient)
 Ray Tracing (efficient Backward Tracing)
• Ignore most incoming directions on the hemisphere, only keeping the most
important ones:

• Rays incoming directly from the light source have a lot of photons
• A Shadow Ray is used to account for this incoming light
• This is called direct illumination, since the light is coming directly from a light source
• Reflective objects bounce a lot of photons in the mirror reflection direction
• This incoming light is accounted for with a Reflected Ray
• Transparent objects transmit a lot of photons along the transmitted ray direction
• This incoming light is accounted for with a Transmitted Ray

• Downside: ray tracing ignores a lot of the incoming light, and thus cannot
reproduce many visual effects
 Bidirectional Ray Tracing
• Combine forward and backward tracing:
• Step 1: Emit photons from the light, bathe objects in those photons, and record
the result in a light map
• Photons bounce around illuminating shadowed regions, bleeding color, etc.
• Note: light maps don’t change when the camera moves (so they can be precomputed)
• Step 2: Ray trace the scene, using the light map to estimate indirect light (from
the ignored directions of the hemisphere)

• IMPORTANT: Still treat the most important directions on the hemisphere with ray
tracing, for increased accuracy
• Shadow Rays for direct illumination
• Reflected Rays
• Transmitted Rays
 Light Maps
• Light maps work great for soft shadows, color bleeding, etc.
• They can also generate many other interesting effects:
 Recall: Lighting Equation
• Multiplying the BRDF by an incoming irradiance gives the outgoing radiance
𝑑𝐿𝑜 𝑑𝑢𝑒 𝑡𝑜 𝑖 (𝜔𝑖 , 𝜔𝑜 ) = 𝐵𝑅𝐷𝐹 𝜔𝑖 , 𝜔𝑜 𝑑𝐸𝑖 (𝜔𝑖 )

• For even more realistic lighting, we’ll bounce light all around the scene
𝑑𝐸
• It’s tedious to convert between 𝐸 and 𝐿, so use 𝐿 = to obtain:
𝑑𝜔 cos 𝜃𝑙𝑖𝑔ℎ𝑡
𝑑𝐿𝑜 𝑑𝑢𝑒 𝑡𝑜 𝑖 (𝜔𝑖 , 𝜔𝑜 ) = 𝐵𝑅𝐷𝐹 𝜔𝑖 , 𝜔𝑜 𝐿𝑖 𝑑𝜔𝑖 cos 𝜃𝑖

• Then,

𝐿𝑜 𝜔𝑜 = න 𝐵𝑅𝐷𝐹 𝜔𝑖 , 𝜔𝑜 𝐿𝑖 cos 𝜃𝑖 𝑑𝜔𝑖

𝑖∈ℎ𝑒𝑚𝑖

Changed 𝑖 ∈ 𝑖𝑛 for “incoming directions” to 𝑖 ∈ ℎ𝑒𝑚𝑖 for

“hemisphere”
 Lighting Equation
• Explicitly add the dependencies on surface location 𝑥 and incoming angle 𝜔𝑖
• Add an emission term 𝐿𝑒 , so 𝑥 can be a location on the surface of an actual light too

𝐿𝑜 𝑥, 𝜔𝑜 = 𝐿𝑒 𝑥, 𝜔𝑜 + න 𝐵𝑅𝐷𝐹 𝑥, 𝜔𝑖 , 𝜔𝑜 𝐿𝑖 (𝑥, 𝜔𝑖 ) cos 𝜃𝑖 𝑑𝜔𝑖

𝑖∈ℎ𝑒𝑚𝑖

• Incoming light from direction 𝜔𝑖 left some other surface point 𝑥′ going in direction −𝜔𝑖
• So, replace 𝐿𝑖 𝑥, 𝜔𝑖 with 𝐿𝑜 (𝑥 ′ , −𝜔𝑖 )

𝐿𝑜 𝑥, 𝜔𝑜 = 𝐿𝑒 𝑥, 𝜔𝑜 + න 𝐵𝑅𝐷𝐹 𝑥, 𝜔𝑖 , 𝜔𝑜 𝐿𝑜 (𝑥 ′ , −𝜔𝑖 ) cos 𝜃𝑖 𝑑𝜔𝑖

𝑖∈ℎ𝑒𝑚𝑖
 An Implicit Equation
• Computing the outgoing radiance 𝐿𝑜 𝑥, 𝜔𝑜 on a particular surface requires knowing the
outgoing radiance 𝐿𝑜 (𝑥 ′ , −𝜔𝑖 ) from all the other (relevant) surfaces
• But the outgoing radiance from those other surfaces (typically) depends on the outgoing
radiance from the surface under consideration (circular dependencies)

𝐿𝑜 𝑥, 𝜔𝑜 = 𝐿𝑒 𝑥, 𝜔𝑜 + න 𝐿𝑜 (𝑥 ′ , −𝜔𝑖 ) 𝐵𝑅𝐷𝐹 𝑥, 𝜔𝑖 , 𝜔𝑜 cos 𝜃𝑖 𝑑𝜔𝑖

𝑖∈ℎ𝑒𝑚𝑖
Reflected Light Emission Reflected Light BRDF incident angle
UNKNOWN KNOWN UNKNOWN KNOWN KNOWN

• Fredholm Integral Equation of the second kind (extensively studied) given in canonical form
with kernel 𝑘 𝑢, 𝑣 by:
𝑙 𝑢 = 𝑒 𝑢 + ‫𝑢 𝑘 𝑣 𝑙 ׬‬, 𝑣 𝑑𝑣
 Aside: Participating Media
• “Air” typically contains participating
media (e.g. dust, droplets, smoke, etc.)
• 𝐿 should actually be defined over all of 3D
space (not just on 2D surfaces)
• The incoming light should be considered
in a sphere centered around each point in
3D space

• Neglecting this assumes that “air” is a

vacuum
• That restricts 𝐿 to surfaces
 Discretization of the Lighting Equation
• Choose 𝑝 points, each representing a chunk of surface area, which is a 2D
discretization (for participating media, volume chunks are 3D)
• For each of the 𝑝 points: Choose 𝑞 outgoing directions, each representing a
chunk of solid angles of the hemisphere (or sphere), which is a 2D discretization

• 𝐿𝑜 and 𝐿𝑒 then each have 𝑝 ∗ 𝑞 unknowns, a 4D (or 5D) discretization

• Represent them by vectors: 𝐿 and 𝐸, each with length 𝑝 ∗ 𝑞
• The light transport “kernel” matrix 𝐾 has size 𝑝 ∗ 𝑞 by 𝑝 ∗ 𝑞

• The linear system of equations is: 𝐿 = 𝐸 + 𝐾𝐿 or 𝐼 − 𝐾 𝐿 = 𝐸

• Solution: 𝐿 = 𝐼 − 𝐾 −1 𝐸 = 𝐼 + 𝐾 + 𝐾 2 + ⋯ 𝐸
• Since 𝐾 bounces only a fraction of the light (the rest is absorbed), higher powers are smaller
(and the infinite series can be truncated)
 Power Series
𝐿 = 𝐸 + 𝐾𝐸 + 2
𝐾 𝐸 + 3
𝐾 𝐸 +⋯

Emission directly
from Light Sources

Direct Illumination
(light bounces
only once) Global Illumination
(indirect lighting,
two bounces) Global
Illumination Etc.
(indirect lighting,
three bounces)
Power Series
 Tractability
• A typical scene might warrant thousands or tens of thousands of area chunks
• So, 𝑝 could be 1e3, 1e4, 1e5, 1e6, etc.
• Incoming light could vary significantly across the hemisphere
• So, 𝑞 might need to be 1e2, 1e3, 1e4, etc.
• 𝐿 and 𝐸 would then range in length from 1e5 to 1e10
• The matrix 𝐾 would then range in size from 1e5 by 1e5 up to 1e10 by 1e10

• 𝐾 would have between 1e10 and 1e20 entries!

• This tractability analysis is for the 4D problem (5D is even worse)

• The curse of dimensionality makes problems in 4D and 5D (and higher) hard to
discretize (with numerical quadrature)
 Addressing Tractability
• Separate the diffuse and specular contributions (to be treated separately)

Diffuse:
• Assume all materials are purely diffuse (i.e. no specular contributions)
• Compute the view-independent global illumination for the entire scene
• This can be done in a pre-processing step

Specular:
• Compute (view-dependent) specular illumination on-the-fly as the camera moves
• Use the Phong Shading model (or any other model)
 Radiosity & Albedo
• Radiosity: power per unit surface area leaving a surface (similar to irradiance, but outgoing
instead of incoming):
𝑑Φ
𝐵 𝑥 = = න 𝐿𝑜 (𝑥, 𝜔𝑜 ) cos 𝜃𝑜 𝑑𝜔𝑜
𝑑𝐴 ℎ𝑒𝑚𝑖
• When 𝐿𝑜 is independent of 𝜔𝑜 (i.e. for a purely diffuse surface):
𝑑Φ
𝐵 𝑥 = = 𝐿(𝑥) න cos 𝜃𝑜 𝑑𝜔𝑜 = 𝜋𝐿 𝑥
𝑑𝐴 ℎ𝑒𝑚𝑖
• Albedo: a “reflection coefficient” relating incoming light hitting a surface patch (irradiance 𝐸𝑖 )
to outgoing light emitted in all possible directions

𝜌 𝑥 =න 𝐵𝑅𝐷𝐹(𝑥, 𝜔𝑜 , 𝜔𝑖 ) cos 𝜃𝑜 𝑑𝜔𝑜

ℎ𝑒𝑚𝑖
• When the BRDF is independent of 𝜔𝑜 and 𝜔𝑖 (i.e. for a purely diffuse surface):

𝜌 𝑥 = 𝐵𝑅𝐷𝐹(𝑥) න cos 𝜃𝑜 𝑑𝜔𝑜 = 𝜋 𝐵𝑅𝐷𝐹(𝑥)

ℎ𝑒𝑚𝑖
 Purely Diffuse Surface Lighting Equation
• Recall: 𝐿𝑜 𝑥, 𝜔𝑜 = 𝐿𝑒 𝑥, 𝜔𝑜 + ‫∈𝑖׬‬ℎ𝑒𝑚𝑖 𝐿𝑜 (𝑥 ′ , −𝜔𝑖 )𝐵𝑅𝐷𝐹 𝑥, 𝜔𝑖 , 𝜔𝑜 cos 𝜃𝑖 𝑑𝜔𝑖
• Multiply through by cos𝜃𝑜 𝑑𝜔𝑜 and integrate over the hemisphere (i.e. over 𝑑𝜔𝑜 ) to obtain:

𝐵 𝑥 =𝐸 𝑥 +න 𝐵(𝑥′)𝐵𝑅𝐷𝐹 𝑥, 𝜔𝑖 , 𝜔𝑜 cos 𝜃𝑖 𝑑𝜔𝑖

𝑖∈ℎ𝑒𝑚𝑖

• 𝐵 is a 2D function (of 𝑥), whereas 𝐿 was a 4D function (of 𝑥 and 𝜔𝑜 )

• Assume that all surfaces have a diffuse BRDF independent of angle:

𝜌(𝑥)
𝐵 𝑥 =𝐸 𝑥 + න 𝐵 𝑥 ′ cos 𝜃𝑖 𝑑𝜔𝑖
𝜋 𝑖∈ℎ𝑒𝑚𝑖
 Recall: Solid Angle vs. Cross-Sectional Area
• The (orthogonal) cross-sectional area is 𝑑𝐴 𝑐𝑜𝑠𝜃
𝑑𝐴𝑠𝑝ℎ𝑒𝑟𝑒 𝑑𝐴 𝑐𝑜𝑠𝜃
• So, 𝑑𝜔 = = (so, the solid angle varies with the tilt 𝜃 and the distance 𝑟)
𝑟2 𝑟2

S0
unit
sphere
d

r N̂


surface
patch
dA
 Replace Solid Angle with Surface Area
𝑑𝐴 𝑐𝑜𝑠𝜃 𝑑𝐴′ 𝑐𝑜𝑠𝜃𝑜
• Note: 𝑑𝜔 = means that 𝑑𝜔𝑖 =
𝑟2 𝑥−𝑥 ′ 22

𝜌(𝑥) ′ x'
• So, 𝐵 𝑥 = 𝐸 𝑥 + ‫∈𝑖׬‬ℎ𝑒𝑚𝑖 𝐵 𝑥 cos 𝜃𝑖 𝑑𝜔𝑖 is: o
𝜋

𝐵 𝑥 = 𝐸 𝑥 + 𝜌(𝑥) න 𝐵 𝑥
cos 𝜃𝑖 cos 𝜃𝑜 ′
′ − i dA'
′ 2 𝑑𝐴
𝑖∈ℎ𝑒𝑚𝑖 𝜋 𝑥−𝑥 2 i
i
• Let 𝑉 𝑥, 𝑥 ′ = 1 when 𝑥 and 𝑥′ are mutually visible and
x
𝑉 𝑥, 𝑥 ′ = 0 otherwise, then:
′ ′
cos 𝜃𝑖 cos 𝜃𝑜 ′
𝐵 𝑥 = 𝐸 𝑥 + 𝜌(𝑥) න 𝐵 𝑥 𝑉 𝑥, 𝑥 ′ 2 𝑑𝐴
𝑎𝑙𝑙 𝑥 ′ 𝜋 𝑥−𝑥 2
 A Tractable Discretization
• Choose 𝑝 points, each representing a chunk of surface area (a 2D discretization)
cos 𝜃𝑖 cos 𝜃𝑗
• Then 𝐵𝑖 = 𝐸𝑖 + 𝜌𝑖 σ𝑗≠𝑖 𝐵𝑗 𝐹𝑖𝑗 with a purely geometric 𝐹𝑖𝑗 = 𝑉 𝑥𝑖 , 𝑥𝑗 2 𝐴𝑗
𝜋 𝑥𝑖 −𝑥𝑗
2
• Rearrange to 𝐵𝑖 − 𝜌𝑖 σ𝑗≠𝑖 𝐵𝑗 𝐹𝑖𝑗 = 𝐸𝑖 and put into matrix form:

1 −𝜌1 𝐹12 ⋯ −𝜌1 𝐹1𝑝 𝐵1 𝐸1

−𝜌2 𝐹21 1 ⋯ −𝜌2 𝐹2𝑝 𝐵2 𝐸2
⋮ = ⋮
⋮ ⋮ ⋱ ⋮
−𝜌𝑝 𝐹𝑝1 −𝜌𝑝 𝐹𝑝2 ⋯ 1 𝐵𝑝 𝐸𝑝
• For 𝑝 ranging from 1e3 to 1e6: 𝐵 and 𝐸 have the same size, and the matrix has 1e6 to 1e12
entries (still large, but 1e4 to 1e8 times smaller than previously)
 Form Factor
𝐹෠𝑖𝑗 𝐹෠𝑖𝑗
• Write 𝐹𝑖𝑗 = 𝑉 𝑥𝑖 , 𝑥𝑗 and 𝐹𝑗𝑖 = 𝑉 𝑥𝑖 , 𝑥𝑗 with
𝐴𝑖 𝐴𝑗
(symmetric) form factor:
cos 𝜃𝑖 cos 𝜃𝑗
෠
𝐹𝑖𝑗 = 2 𝐴𝑖 𝐴𝑗
𝜋 𝑥𝑖 − 𝑥𝑗
2
• 𝐹෠𝑖𝑗 represents how the light energy leaving one surface
impacts the other surface, and vice versa (and only
depends on the geometry, not on the light)

• The visibility between between 𝑥𝑖 and 𝑥𝑗 , i.e. 𝑉 𝑥𝑖 , 𝑥𝑗 ,

also only depends on the geometry (and can be included
into 𝐹෠𝑖𝑗 if desired)
 Understanding the Form Factor
• Place a unit hemisphere at a surface point 𝑥𝑖

• Project a surface onto the unit hemisphere, noting

𝑑𝐴 𝑐𝑜𝑠𝜃 𝐴𝑗 cos𝜃𝑗
that 𝑑𝜔 = gives 2 as the result
𝑟2 𝑥𝑖 −𝑥𝑗
2
• Project the result downwards onto the circular base
of the hemisphere, which multiples by cos 𝜃𝑖
• Recall ‫∈𝑖׬‬ℎ𝑒𝑚𝑖 cos 𝜃𝑖 𝑑𝜔𝑖 = 𝜋, the area of the unit circle
• Finally, divide the result by the total area 𝜋 to get the
fraction of the circle occupied

cos 𝜃𝑖 cos 𝜃𝑗
• Overall, this gives: 𝐹𝑖𝑗 = 2 𝐴𝑗
𝜋 𝑥𝑖 −𝑥𝑗
2
 Implementation
Precomputation:
• Create a hemicube, and divide each face into sub-squares
(as small as desired)
• For each sub-square, use hemisphere projection (from
the last slide) to pre-compute its contribution to 𝐹𝑖𝑗

For each surface chunk:

• Place the hemicube at a surface point 𝑥𝑖
• Surface patches (from other objects) are projected onto
the hemicube in order to approximate 𝐹𝑖𝑗 (using the pre-
computed values for the sub-squares)
• The five hemicube faces can each be treated as a film
plane where sub-squares are pixels (standard scanline
rasterization)
• The depth buffer is used for the visibility term
Hemicube Scanline Rasterization
 Iterative Solvers
• For large matrices, iterative solvers are typically far more accurate than direct methods (like
Cramer’s rule) for computing inverses
• Iterative methods start with an initial guess, and subsequently iteratively improve it

2 1 𝑥 8 𝑥 2
• Consider = with solution =
1 2 𝑦 10 𝑦 4
𝑥 0
• Start with an initial guess of 𝑦 =
0

• Jacobi iteration (solve both equations, using the old values to get the new values):
𝑛𝑒𝑤 8−𝑦 𝑜𝑙𝑑 𝑛𝑒𝑤 10−𝑥 𝑜𝑙𝑑
•𝑥 = and 𝑦 =
2 2

• Gauss Seidal iteration (always use the most up to date values):

𝑐𝑢𝑟𝑟𝑒𝑛𝑡 8−𝑦 𝑐𝑢𝑟𝑟𝑒𝑛𝑡 𝑐𝑢𝑟𝑟𝑒𝑛𝑡 10−𝑥 𝑐𝑢𝑟𝑟𝑒𝑛𝑡
•𝑥 = and 𝑦 =
2 2
 Jacobi vs. Gauss-Seidal
Iteration Jacobi Gauss Seidel
x y x y
1 0 0 0 0
2 4 5 4 3
3 1.5 3 2.5 3.75
4 2.5 4.25 2.125 3.9375
5 1.875 3.75 2.03125 3.984375
6 2.125 4.0625 2.007813 3.996094
7 1.96875 3.9375 2.001953 3.999023
8 2.03125 4.015625 2.000488 3.999756
9 1.9921875 3.984375 2.000122 3.999939
10 2.0078125 4.00390625 2.000031 3.999985
11 1.998046875 3.99609375 2.000008 3.999996
12 2.001953125 4.000976563 2.000002 3.999999
13 1.999511719 3.999023438 2 4
14 2.000488281 4.000244141 2 4
15 1.99987793 3.999755859 2 4
16 2.00012207 4.000061035 2 4
17 1.999969482 3.999938965 2 4
18 2.000030518 4.000015259 2 4
19 1.999992371 3.999984741 2 4
20 2.000007629 4.000003815 2 4
 Better Initial Guess
Iteration Jacobi Gauss Seidal
x y x y
1 2 3 2 3
2 2.5 4 2.5 3.75
3 2 3.75 2.125 3.9375
4 2.125 4 2.03125 3.984375
5 2 3.9375 2.007813 3.996094
6 2.03125 4 2.001953 3.999023
7 2 3.984375 2.000488 3.999756
8 2.0078125 4 2.000122 3.999939
9 2 3.99609375 2.000031 3.999985
10 2.001953125 4 2.000008 3.999996
11 2 3.999023438 2.000002 3.999999
12 2.000488281 4 2 4
13 2 3.999755859 2 4
14 2.00012207 4 2 4
15 2 3.999938965 2 4
16 2.000030518 4 2 4
17 2 3.999984741 2 4
18 2.000007629 4 2 4
19 2 3.999996185 2 4
20 2.000001907 4 2 4
 Iterative Radiosity
• Gathering - update one surface by collecting light energy from all surfaces
• Shooting - update all surfaces by distributing light energy from one surface
• Sorting and Shooting - choose the surface with the greatest un-shot light energy and use
shooting to distribute it to other surfaces
• start by shooting light energy out of the lights onto objects (the brightest light goes first)
• then the object that would reflect the most light goes next, etc.
• Sorting and Shooting with Ambient - start with an initial guess for ambient lighting and do
sorting and shooting afterwards
Iterative Radiosity