The Alchemy Screen-Space Ambient Obscurance Algorithm

Morgan McGuire∗ Brian Osman Michael Bukowski Padraic Hennessy

NVIDIA & Williams College Vicarious Visions Vicarious Visions Vicarious Visions

Figure 1: left: Environment lighting. right) Modulated by Alchemy ambient obscurance, computed from 12 samples per pixel at 1280×720
in 3 ms on GeForce 580. The algorithm is easy to tune, robust, and captures darkening at many scales and orientations.

Abstract meter scale. Ambient obscurance (AO) is an illumination term that

corrects the indirect light by scaling it proportional each point’s vis-
Ambient obscurance (AO) produces perceptually important illumi- ible obscurance on that scale. A point that is locally obscured from
nation effects such as darkened corners, cracks, and wrinkles; prox- most directions should receive little indirect illumination from dis-
imity darkening; and contact shadows. We present the AO algo- tant objects, while highly accessible points receive most indirect il-
rithm from the Alchemy engine used at Vicarious Visions in com- lumination. Obscurance is visually important for object definition,
mercial games. It is based on a new derivation of screen-space to provide a sense of scale, and as a spatial cue through contact
obscurance for robustness, and the insight that a falloff function shadows and darkening on concave surfaces. It is also computa-
can cancel terms in a visibility integral to favor efficient operations. tionally intense to estimate directly from scene geometry–any point
Alchemy creates contact shadows that conform to surfaces, cap- may be obscured from any direction. This is why screen space ap-
tures obscurance from geometry of varying scale, and provides four proximations, which are independent of the number of polygons,
intuitive appearance parameters: world-space radius and bias, and have proven very attractive in practice.
aesthetic intensity and contrast.
This paper presents the screen space AO algorithm we developed
The algorithm estimates obscurance at a pixel from sample points for a specific Guitar Hero title and subsequently generalized and
read from depth and normal buffers. It processes dynamic scenes integrated into the cross-platform Alchemy game engine. Figure 1
at HD 720p resolution in about 4.5 ms on Xbox 360 and 3 ms on demonstrates its visual impact. The left image shows a scene with
NVIDIA GeForce580. environment lighting only. The image on the right modulates that
lighting by Alchemy AO, which resolves the fine details and spa-
CR Categories: I.3.3 [Picture/Image Generation]: Display Al- tial relationships between objects. The algorithm follows from
gorithms; I.3.7 [Three-Dimensional Graphics and Realism]: Color, three insights: Derive a robust estimator from the rendering equa-
shading, shadowing, and texture tion; provide temporal coherence by making the estimator efficient
enough to evaluate many times per pixel; and achieve that effi-
Keywords: ambient occlusion, ambient obscurance, screen space ciency by shaping the falloff function to cancel expensive opera-
tions. Alchemy addresses the drawbacks of previous screen-space
AO methods, none of which satisfy all of the following require-
1 Introduction ments:

Indirect illumination is a significant factor in realistic lighting. Ev- 1. Robust: Conform obscurance to affected surfaces (e.g., no
ery game approximates indirect light sparsely over large distances, shadows “floating in air” near silhouettes), limit viewer de-
either via precomputation (e.g., [Larsson 2010]; environment maps pendence of intensity, and maximize temporal coherence.
and constant ambient are the classic examples) or dynamic gener- 2. Multiscale: Capture phenomena at multiple scales: shadowed
ation (e.g., [Kaplanyan and Dachsbacher 2010; Martin and Einars- deep pits, corner darkening, contact shadows, wrinkles.
son 2010]). Those sparse methods miss occlusion on the small, sub- 3. Artist-control: Provide intuitive parameters with large sweet-
∗ e-mail: morgan@cs.williams.edu,{bosman,mbukowski,phennessy}@vvisions.com spots and predictable quality.
4. Scalable: Compute in 3-5 ms, from Xbox 360 to Windows
Direct3D 11 hardware by varying quality.
Like all screen-space methods, its limitations are sample variance
(addressed by edge-aware filtering) and under-obscurance due to
unseen occluders behind the depth buffer surface and outside the
field of view. Rendering a guard band about the viewport can re-
duce the latter. We attribute the visual fidelity and robustness of
Alchemy AO compared to phenomenological methods to its radio- 1. Render an early−Z pass
metric derivation, however it is not physically correct. The value of 2. Render the camera−space normal and position buffers
the algorithm is that it makes practical concessions to artistic and 3. Render the ambient obscurance buffer
performance goals. 4. [Render other effects]
5. Cross−bilateral filter obscurance
1.1 Related Work 6. Composite obscurance, texture, and lighting in a forward pass

Listing 1: Rendering system context for obscurance.

Screen space Luft et al. [2006] observed that placing blurry
black lines at depth discontinuities in screen space by unsharp
masking increases visual contrast similarly to ambient occlusion. Listing 1 describes the AO algorithm in the context of a larger
A line of efficient screen-space AO methods rapidly followed, multi-pass rendering system. Our algorithm comprises obscurance
some published and some orally presented based on production use. pass 3 and filtering pass 5. Each of these may be implemented ei-
These developed efficient GPU sampling [Shanmugam and Arikan ther by rendering a full-screen rectangle with a shader that reads
2007; Mittring 2007; Kajalin 2009] and reconstruction [Filion and from the position and normal buffers, or as a forward pass over the
McNaughton 2008] methods, increasingly physically-based deriva- scene geometry.
tions [Bavoil and Sainz 2009a; Szirmay-Kalos et al. 2010; Loos
Environment light, shadows, and direct illumination are computed
and Sloan 2010], and multipass methods for addressing both per-
in pass 6, for forward lighting, or in step 4 for deferred shading.
formance and quality issues [Ritschel et al. 2009; Bavoil and Sainz
The final compositing pass uses multisample antialiasing (MSAA)
2009b]. Our technique builds on all of this work but revisits the
to apply screen-resolution obscurance with subpixel edge precision.
core derivation, which we focus on in section 2. We report experi-
It gathers obscurance from five pixels in a cross pattern and cross-
mental results comparing Alchemy to the two most recent methods
bilateral filters them based on camera-space proximity to the sur-
in section 3. Briefly, Volumetric Obscurance [Loos and Sloan 2010;
face being shaded. A deferred shading pass may be substituted for
Ownby et al. 2010] estimates occlusion as proportional to the vol-
pass 6 at the loss of antialiasing.
umetric occlusion of a point, which tends to produce black halos.
Alchemy uses projected solid-angle occlusion, like Horizon-Based
AO [Bavoil and Sainz 2009a], but uses a more efficient sampling 2.1 Background
scheme and avoids over-darkening by conservatively assuming that
unseen space is empty, like VO does. The low-level sampling tech- Let the distance-limited radiance function Lx (C, ω̂) denote radi-
niques of all methods are based on earlier work done at Crytek [Mit- ance incident at point C from a unit direction ω̂, due to points within
tring 2007; Kajalin 2009] and we incorporate filtering and aesthetic the ball of radius x about C. Let the visibility function V (C, P ) be
falloff methods from StarCraft II [Filion and McNaughton 2008]. 1 if there is no intersection between the open-ended line segment
CP and the scene, and 0 otherwise. The net incident radiance func-
tion is the sum of near field radiance and far field radiance mod-
Geometric Computing AO from geometry avoids the viewer- ulated by visibility [Arikan et al. 2005; McGuire 2010] for some
dependent errors from screen-space approximations, at the cost of given radius r,
performance for complex scenes.. We briefly describe geometric
methods of interest for real-time rendering, i.e., within a factor of L∞ (C, ω̂) = Lr (C, ω̂) + V (C, C + rω̂)L∞ (C + rω̂, ω̂). (1)
20× the performance of screen-space methods. See Real Time Ren-
dering [Akenine-Möller et al. 2008] for a broad survey. This decomposition is useful for combining estimates of near and
far field illumination from different algorithms. It is common in
Oat and Sander [2007] precomputed an accessible cone about each real-time rendering to choose radius r less than one meter and sep-
texel on a static mesh, which they then applied to dynamic illumi- arate the far-field V · L∞ term into shadowed direct illumination
nation. Reinbothe et al. [2009] dynamically voxelized polygonal and ambient-occluded environment lighting terms. It is also com-
scenes and sampled occlusion in voxel space. Bunnell [2005] com- mon to ignore small-scale global illumination by assuming Lr ≈ 0,
puted occlusion on a GPU in multiple passes with per-vertex proxy although recent screen-space gathering methods offer more accu-
geometry. rate alternatives [Ritschel et al. 2009; Soler et al. 2010].

Ambient Occlusion Fields [Kontkanen and Laine 2005] are pre- The ambient occlusion (see [Landis 2002]) of surface point C with
computed voxel occlusion values relative to a static mesh, which normal n̂ is
can be composed between moving rigid objects to estimate AO. AOr∗ (C, n̂) = 1 − A∗r (C, n̂), (2)
McGuire [2010] and Laine and Karras [2010] extended this idea to where 0 ≤ A∗r (C, n̂) ≤ 1 is the fractional visibility of the far field
dynamic meshes. at C [Cook and Torrance 1982]. Function A is visibility weighted
by projected area over the positive hemisphere,
2 Algorithm Z
1
A∗r (C, n̂) = V (C, C + rω̂) ω̂ · n̂ dω̂. (3)
π Ω
We begin by summarizing the key terms and equations governing
indirect illumination. Previous screen space methods approximate Note that (ω̂ · n̂) in eqn 3 accounts for the solid angle of occluders
these equations (or their resultant phenomena) in ways that are ex- projected onto the tangent plane at C. For example, this decreases
tremely efficient but can introduce significant error. We instead re- the impact of occluders only block incident light at shallow angles,
tain the full model and observe that the falloff function customarily which is radiometrically correct. Specifically, this is not the pro-
chosen for aesthetics can also be selected to eliminate expensive jected area factor that will later be applied to indirect illumination,
operations. With this, we are able to retain the full radiometric so this does not “double count” that factor.
model until the screen-space sampling step. That process leads to
a relatively robust solution. We augment the screen-space solution Shadows account for the visibility factor in direct illumination. ob-
with parameters for managing buffer precision and artistic goals in scurance is an approximation of shadowing for indirect illumination
production use. that modulates environment light. To transition smoothly between
 1.00  
the near- and far-field illumination models, ambient occlusion is 0.90  
often extended to ambient obscurance (AO) [Zhukov et al. 1998] 0.80  
Alchemy, u = 0.10 r
using a falloff function that reduces the influence of occlusion with 0.70   Alchemy, u = 0.05 r
distance. The analogous obscured function is 0.60   StarCraft II
0.50  
Z 0.40  
Ar (C, n̂) = 1− [1−V (C + ω̂ min(t(C, ω̂) + , r))]·g(t)·(n̂·ω̂) dω̂ 0.30  
Ω 0.20  
(4) 0.10  
where t is the distance from C to the first scene point encountered 0.00  
by a ray cast in direction ω̂ and 0 ≤ g(x) ≤ 1 is the aesthetically- 0   0.2   0.4   0.6   0.8   1  
selected falloff function, which commonly obeys g(x) = 0|x≥r and Occluder distance t / r
g(0) = 1. For the remainder of this paper we omit the parameters
of the A function, which are always r, C, and n̂. Figure 2: Plot of falloff g(t)vs. t for Alchemy and StarCraft II AO.

2.2 Derivation of the obscurance Estimator r

We now derive a robust estimator of obscurance with intuitive pa-
rameters for aesthetic falloff. Our particular falloff function leads C
to an extremely efficient implementation.

Let Γ ⊆ Ω be the subset of the hemisphere for which there exists v
some distance 0 < t(C, ω̂) < r at which the ray from C in direction
ω̂ ∈ Γ intersects the scene. That is, Γ is the set of all near-field P
occluders projected onto the unit sphere. Because Γ covers all near n
occluders, the visibility term from eq. 4 is always zero on it:
image plane
€ Q’
V (C, ω̂ min (t(C, ω̂) + , r)) = 0 ω̂∈Γ
. (5)
Figure 3: Geometric construction for Alchemy AO.
Changing the integration domain in eq. 4 to Γ gives

The New Alchemy Algorithm

Z
A=1− g (t(C, ω̂)) ω̂ · n̂ dω̂. (6) at random on its projection, and then read a depth or position buffer
Γ to find the camera-space scene point P = (xP , yP , z(Q0 )) on that
ray. Figure 3 summarizes the process. This selection biases the
We choose the falloff function g(t) = u · t · max(u, t)−2 , which re- samples because a ball about C projects to an ellipse, not a disk.
sembles the shifted hyperbola that Filion and McNaughton [2008] Like Loos and Sloan, we ignore this source of error because it is
reported artistically desirable in StarCraft II (figure 2) and leads to relatively small for a typical field of view up to 70 degrees.
a remarkable simplification under our geometric construct. Substi-
tute into eq. 6: We augment the solution with expressive parameters for intensity
scale (σ) and contrast (k) to allow intentional over- and under-
Z
ut(C, ω̂) occlusion, and nominally scale intensity by 1/π so that σ = 1.0
A=1− ω̂ · n̂ dω̂. (7) and k = 1.0 produce a useful baseline.
Γ max(u, t(C, ω̂))2
In practice, the falloff constant u in eq. 9 serves only to avoid di-
Rewrite this in terms of the vector ~v (ω̂) = ω̂t(C, ω̂) from C to vision by zero and interferes with setting σ independently, so we
the occluder (P in figure 3). Because ω̂ = ~v (ω̂)/||~v (ω̂)|| and eliminate it from the numerator and replace max(u2 , ~vi · ~vi ) in the
t(C, ω̂) = ||~v ||, denominator with ~vi · ~vi + . Choose  large enough to prevent un-
Z derflow and small enough that corners do not become white from
~v (ω̂) · n̂ light leaks. We found that values within two orders of magnitude of
A=1−u 2 v (ω̂) · ~
dω̂. (8)
Γ max (u , ~ v (ω̂))  = 0.0001 produced indistinguishable results.
Numerically estimate this integral by sampling a set {ω̂i } of s di- Bias distance β is the ambient-obscurance analog of shadow map
rections uniformly at random, giving: bias: increase β to reduce self-shadowing (figure 4), decrease β if
light leaks into corners. We scale it by z (which is negative) to
s
2πu X max(0, v~i · n̂) · H(r − ||~
vi ||) efficiently compensate for the dot products becoming increasingly
A≈1− . (9) sensitive to error in n̂ at distant points. Our final equation is:
s i=1 max(u2 , v~i · v~i )

where H(x) is the Heaviside step function. s

!k
2σ X max(0, v~i · n̂ + zC β)
A ≈ max 0, 1 − · (10)
2.3 Screen-Space Approximation s i=1 v~i · v~i + 

For practical application, we evaluate our radiometrically-based es-

timator from eq. 9 in screen space by sampling a depth or position The parameters are straightforward to adjust because they have in-
buffer, introduce concessions to finite precision, and extend the re- tuitive interpretations and reasonable values work for a large range
sult with aesthetic controls for artists. of scenes. Obscurance has a maximum world-space radius of r and
does not occur between objects closer than β. From there, one tunes
We choose each sample in the same manner as Loos and for appearance. Images in this paper use r = 0.5m, σ = 1, k = 1,
Sloan [2010]: consider the disk of radius r and center C that is par- and β = 10−4 m unless specified (N.B., art directors for our games
allel to the image plane, select a screen-space point Q0 uniformly prefer to overdarken with higher intensity and contrast than this).
Figure 4: Top view with 8-bit normals and 16-bit positions. The
radius of the large gear is about 0.8 m. Left: Self-occlusion with
β = 0 m. Right: Corrected with β = 10−4 m.

Figure 7: left: Environment-lit character; right: Alchemy AO en-

hances folds and wrinkles. r = 0.2

within that region. One would normally render a slightly larger

image and crop during blur; this adds about 0.2 ms to the render
time on GeForce 580.
Figure 1 shows obscurance in Industrial at varying scales and orien-
tations. This scene’s sharp concave and convex corners, many lay-
ers, and curves cover many important cases; figure 7 also demon-
strates obscurance for the organic shapes of the Orc model.
Figure 5: Diminishing sample count with distance. (The horizon- The darkening between vertices on the large platform disk in the
tal banding arises from interpolated vertex normals under coarse lower right of Industrial is undesirable but consistent with the
tessellation–if undesired, it can be eliminated by finer tessellation, model. Those are locations where the interpolated vertex nor-
face normals, or per-surface σ values.) mal indicates a tangent plane that passes beneath the nearby ver-
tices. This can be reduced by computing AO from face- instead of
interpolated-normals, or by increasing the tessellation.
Figure 8 shows that increasing radius r while holding the number
of samples constant increases the range of obscurance phenomena
at the cost of sharpness. The farther-back gears gradually become
darker as the radius grows to capture the influence of higher lay-
ers. Large radii are visually desirable but incur two costs. The first
is that more samples are required to reduce variance. The second
is that performance falls as r increases because samples near C
correspond to position-buffer locations likely to be in cache, while
points distant in screen space are likely not cached. All previous
screen-space ambient occlusion algorithms have this property and
it is widely known in the industry, although rarely reported in the
Figure 6: left: Raw obscurance from 12 samples per pixel; right: literature.
after two 13-tap 1D cross-bilateral filter passes.
Figure 9 shows the corresponding render times on GeForce 580
with 12 samples per pixel. The probability of a cache hit is the-
oretically proportional to r−2, but we observe a more favorable,
3 Results nearly-linear slowdown. We attribute this to the interaction of a
small sample count compared to cache line size, and a very flat
We report results on six scenes with diverse geometry: parabola. For r > 1.5m in this scene all fetches are cache misses,
Industrial, a giant machine with occlusion at many scales. so time plateaus.
Knight character interacting with an environment. The banded LOD and increase in performance when the screen-
Rocks and Grass nature scene. [Baraka 2010] space projection of r is small mean that the performance of
Sponza architectural benchmark with long sight-lines. [Meinl 2009] Alchemy AO increases when viewing a deep scene, or with a wide
Orc character head with many organic curves. [3D Hannibal 2008] field of view. This is precisely the case where other parts of a ren-
Dragon model scanned by Stanford (10 cm tall). [Stanford 1996] dering system tend to lose performance, such as particle systems
and rasterization itself, because more objects and a larger depth
The first two are from our studio to demonstrate application to pro- range are visible. This property helps balance the total frame ren-
duction game assets. The others are publicly available. dering time.
Accompanying this paper is a set of video results rendered on a Figure 10 compares Volumetric Obscurance [Loos and Sloan 2010],
Windows PC and on Xbox 360 to demonstrate temporal coherence the most recent previous work, to Alchemy AO. The top two rows
under camera and character motion. For the PC version, we show show the Industrial scene and the bottom row shows the Sponza
the guard band region on-screen to reveal the loss of obscurance scene. We tuned both algorithms for comparable performance: on
 Figure 12: left: Environment-lit untextured Rocks and Grass scene; right: With Alchemy AO.

Figure 13: Frames of the dynamic 2 m tall Frog Knight character and moving camera in the Knight scene on Xbox 360, with all lighting
terms and no materials. The rightmost frame shows the obscurance buffer. r = 0.76 m

Xbox 360 GeForce 460 SE GeForce 580 GTX

~D3D9 D3D9 D3D11 / GL4
Samples/band 3 4 4 3 2
Maximum bands 1 1 3 3 3
Registers 6 vec4 15 scalar 15 scalar 15 scalar 15 scalar
32-bit words/pix IO 28 28 40 31 22
Accessibility time 2.3 ms 0.6 ms 2.0 ms 1.5 ms 1.0 ms
Filter pixel width 11 11 13 13 9
Filter time 2.0 ms 1.6 ms 1.0 ms 1.0 ms 0.7 ms

Table 1: Sample counts to hold worst-case performance constant by varying quality for low, mid, and high-end systems.

a GeForce 460 SE GPU, Volumetric Obscurance computes obscu- by view dependence. Note the intense dark shadows under HBAO.
rance in 0.7 ms and Alchemy runs in 0.6 ms. Volumetric Obscu- The more conservative Alchemy result arises from adopting Loos’s
rance runs at two resolutions, taking six samples per pixel at each and Sloan’s [2010] assumption that unseen space is empty. The
resolution. Alchemy uses four samples per pixel in this test; it re- Alchemy result exhibits higher temporal coherence under camera
ceives fewer samples because it reads normals and positions instead motion because it operates at screen resolution (see video); HBAO
of just depth values. Volumetric Obscurance can run at half perfor- upsamples substantially and takes 22 ms if run at full resolution to
mance with normals to better capture fine detail, but that is redun- avoid aliasing.
dant with multiple resolutions.
Figure 12 shows a natural scene with dense clumps of thin grass
The Alchemy AO on the right of figure 10, exhibits desirable prop-
blades and large, smooth rocks. The appearance of foliage strongly
erties. The top row shows that it conforms to surfaces (b) instead
depends on ambient occlusion. Alchemy’s high sampling resolu-
of shadows “floating in air” near silhouettes (a). This is important
tion compared to other screen space methods makes it well suited
for properly planted pillars and feet. The second row demonstrates
to capturing the interaction of many thin objects.
that it is fast and robust enough to operate over a large radius, so oc-
clusion appears from a continuous range of geometric frequencies.
The third row combines these challenging cases over a continuous Figure 13 shows frames from the sequence captured on Xbox 360 in
distance range. our result video. The animated character deforms under animation
and establishes and breaks contact with the ground during the se-
Figure 11 compares Bavoil’s and Sainz’s [2009a] implementation quence. Unlike methods that process AO at very low resolution and
of Horizon-based AO (HBAO) to Alchemy AO for the Dragon upsample significantly, Alchemy samples multiple times at each
scene on GeForce 580. HBAO gives smoother results (a). Alchemy pixel and distributes samples over space and angle, so it provides
captures more fine details (b) and more consistent intensity (c) as high spatial resolution and a good estimate that are robust to object
the view changes, although both screen-space effects are limited and camera motion.
 Volumetric Obscurance Alchemy AO
[Loos and Sloan 2010] [this paper]

Figure 8: Varying obscurance radius r in the Industrial scene.

3.0   Figure 10: Quality comparison to the most recent previous work
2.5   at comparable performance. One aspect we especially sought to
Time (ms)

2.0   improve is that where the pillars in the top row intersect the floor a)
1.5   Volumetric Obscurance produces occlusion “floating in air” about
1.0  
the silhouette; b) Alchemy AO conforms to the ground orientation.
0.5  
0.0   Horizon-based AO Alchemy AO
[Bavoil and Sainz 2009a] [this paper]
0.0   0.5   1.0   1.5   2.0  
r in meters
Figure 9: Increasing radius increases the number of texture cache
misses and thus the render time.

Table 1 compares the algorithm across platforms for the view of

Sponza in figure 10, which shows a range of distances but no sky
(since sky pixels are inexpensive). We target 2-4 ms for the to-
tal cost of computing and filtering obscurance. Our algorithm is
bandwidth-limited, so to maintain approximately constant perfor-
mance we reduce the number of samples per pixel on lower-end
hardware, at the cost of higher noise on those platforms. RGB8
costs one four-byte word and RGB16F position or DEPTH32 both
cost two words per fetch due to memory layout, so our total band-
width cost is higher than the number of usable bytes that we read.
In the context of a game, one can exploit this to amortize the cost Figure 11: Dragon top and side views cropped from 1280×720.
of filtering over four effects. For example, one could include soft Left: HBAO with all optimizations enabled (8.5 ms). Right:
shadows, indirect illumination, or noisy stochastic particles in the Alchemy AO (2.6 ms) with r = 0.1.
unused three channels of an obscurance buffer for simultaneous
blurring. Microsoft’s Pix tool reports that the Xbox 360 implemen-
tation maintains 90% ALU utilization. variance) and sample radius.

We anticipate a future in which GPU performance increases and ge-

4 Discussion ometric methods for light transport become standard for real-time
rendering. However, we also anticipate resolution and geometric
The Alchemy AO algorithm meets a production need for very fast complexity increasing. It is possible that general light-transport
and plausible ambient occlusion with intuitive aesthetic parameters. methods will never be practical for capturing the small-scale occlu-
It combines the most effective elements of techniques previously sion modeled by AO–indeed, even offline rendering systems seek
proven in an industry context by developers at Crytek, Avalanche, efficient explicit AO approximations (e.g., [Bunnell 2005]). So we
and Blizzard and extends them with a new derivation for increased expect that Alchemy AO and methods like it may be useful for some
performance and robustness. The algorithm is stable under a large time to come.
range of parameters and can naturally leverage the performance
ability of a specific target GPU by adjusting the sample count (i.e., The most direct future work is to estimate local indirect illumina-
tion by sampling a color buffer as well as occlusion at each of the L AINE , S., AND K ARRAS , T. 2010. Two methods for fast ray-
samples taken by the Alchemy algorithm. Local screen-space spec- cast ambient occlusion. Computer Graphics Forum (Proc. Euro-
ular reflections work surprisingly well in many games, and there is graphics Symposium on Rendering 2010) 29, 4.
active work in screen-space diffuse and glossy reflection. The in-
teraction of multiple surfaces along the light transport path makes L ANDIS , H. 2002. Production ready global illumination. In SIG-
these physically similar to the AO problem and simulating indirect GRAPH 2002 courses, 331–338.
light presents comparable performance and robustness challenges. L ARSSON , D. 2010. Pre-computing lighting in games. In SIG-
So an algorithm like Alchemy’s is a natural candidate for extension. GRAPH 2010 courses, ACM. in Global illumination across in-
However, it remains open whether AO algorithms are generally ex- dustries, http://cgg.mff.cuni.cz/ jaroslav/gicourse2010/.
tensible to viable estimators of indirect light, in part because the
human perception of aliasing and bias in indirect light may be dif- L OOS , B. J., AND S LOAN , P.-P. 2010. Volumetric obscurance. In
ferent than that in shadows. Proceedings of I3D 2010, ACM, 151–156.
L UFT, T., C OLDITZ , C., AND D EUSSEN , O. 2006. Image en-
Acknowledgements hancement by unsharp masking the depth buffer. ACM Transac-
tions on Graphics 25, 3 (jul), 1206–1213.
We thank Naty Hoffman (Activision Studio Central) for checking
the equations carefully and supporting the publication of algorithms M ARTIN , S., AND E INARSSON , P. 2010. A real-time radiosity ar-
used in games published by Activision, artists Nathan Hawkinson, chitecture for video games. In SIGGRAPH 2010 courses, ACM.
Justin Wharton, Kevin Dobler, and Brent Gibson who created the M C G UIRE , M. 2010. Ambient occlusion volumes. In Proceedings
game scenes used in this paper, and the entire team at Vicarious of High Performance Graphics 2010.
Visions.
M EINL , F., 2009. Crytek sponza. Refinement of Marco Dabrovic’s
model. http://crytek.com/cryengine/cryengine3/downloads.
References
M ILLER , G. 1994. Efficient algorithms for local and global acces-
3D H ANNIBAL, 2008. Orc head. http://turbosquid.com/3d- sibility shading. In SIGGRAPH 1994, ACM, 319–326.
models/orc-head-3d-3ds/420140. M ITTRING , M. 2007. Finding next gen: Cryengine 2. In SIG-
A KENINE -M ÖLLER , T., H AINES , E., AND H OFFMAN , N. 2008. GRAPH 2007 courses, ACM, 97–121.
Real-Time Rendering 3rd Edition. A. K. Peters, Ltd., Natick, NVIDIA, 2009. Enabling ambient occlusion in games.
MA, USA. Online GeForce Guide, viewed on April 3, 2011.
A RIKAN , O., F ORSYTH , D. A., AND O’B RIEN , J. F. 2005. Fast http://geforce.com/#/Optimize/Guides/ambient-occlusion.
and detailed approximate global illumination by irradiance de- OAT, C., AND S ANDER , P. V. 2007. Ambient aperture lighting. In
composition. ACM Trans. Graph. 24, 3, 1108–1114. Proceedings of I3D 2007, ACM, 61–64.
BARAKA, 2010. Rock 02, May. OWNBY, J.-P., H ALL , R., AND H ALL , C. 2010. Rendering tech-
http://turbosquid.com/FullPreview/Index.cfm/ID/528532. niques in Toy Story 3. In SIGGRAPH 2010 courses, ACM.
BAVOIL , L., AND S AINZ , M. 2009. Multi-layer dual-resolution R EINBOTHE , C., B OUBEKEUR , T., AND A LEXA , M. 2009. Hy-
screen-space ambient occlusion. In ShaderX7 , W. Engel, Ed. brid ambient occlusion. EUROGRAPHICS 2009Areaspapers.
BAVOIL , L., AND S AINZ , M. 2009. Multi-layer dual-resolution R ITSCHEL , T., G ROSCH , T., AND S EIDEL , H.-P. 2009. Approxi-
screen-space ambient occlusion. In SIGGRAPH 2009: Talks, mating dynamic global illumination in image space. In Proceed-
ACM, 1–1. ings of I3D 2009, ACM, 75–82.
B UNNELL , M. 2005. Dynamic Ambient Occlusion and Indirect S HANMUGAM , P., AND A RIKAN , O. 2007. Hardware accelerated
Lighting. NVIDIA Corporation. ambient occlusion techniques on GPUs. In Proceedings of I3D
2007, ACM, 73–80.
C OOK , R. L., AND T ORRANCE , K. E. 1982. A reflectance model
for computer graphics. ACM Trans. Graph. 1 (January), 7–24. S OLER , C., H OEL , O., AND ROCHET, F. 2010. A deferred shading
pipeline for real-time indirect illumination. In SIGGRAPH 2010
F ILION , D., AND M C NAUGHTON , R. 2008. Effects & techniques. talks, ACM.
InSIGGRAPH2008 courses, ACM,133–164.
S TANFORD, 1996. Dragon reconstructed from cyberware scan.
I ONES , A., K RUPKIN , A., S BERT, M., AND Z HUKOV, S. 2003. http://graphics.stanford.edu/data/3Dscanrep/.
Fast, realistic lighting for video games. IEEE Comput. Graph.
Appl. 23, 3, 54–64. S TEWART, A. J., AND L ANGER , M. S. 1997. Towards accurate
recovery of shape from shading under diffuse lighting. IEEE
K AJALIN , V. 2009. Screen space ambient occlusion. In ShaderX7 , Transactions on Pattern Analysis and Machine Intelligence 19,
W. Engel, Ed. Charles River Media, March. 1020–1025.
K APLANYAN , A., AND DACHSBACHER , C. 2010. Cascaded light S ZIRMAY-K ALOS , L., U MENHOFFER , T., T ÓTH , B., S Z ÉCSI , L.,
propagation volumes for real-time indirect illumination. In Pro- AND S BERT, M. 2010. Volumetric ambient occlusion for real-
ceedings of I3D 2010, ACM, I3D ’10, 99–107. time rendering and games. IEEE CG&A 30, 1, 70–79.
K APLANYAN , A. 2010. CryENGINE 3: Reaching the speed of Z HUKOV, S., I ONES , A., AND K RONIN , G. 1998. An ambient
light. In SIGGRAPH 2010 courses, ACM. light illumination model. In Rendering Techniques ’98, Springer-
Verlag Wien New York, G. Drettakis and N. Max, Eds., Euro-
KONTKANEN , J., AND L AINE , S. 2005. Ambient occlusion fields. graphics, 45–56.
In Proceedings I3D 2005, ACM Press, 41–48.