1
2
3
Very complex rendering: Dynamic sweat and damage, animated trails. Dynamic wrinkles, muscle maps. Subsurface scattering (via UV-space blur in NIS). SSAO. Extremely dense models (sub-pixel triangles...). Almost used all the texture samplers on the skin. 3d Crowd with facial animation. 3d Cloth / softbody for impacts on the face... See GDC 2010 Fight Night 4 presentation...
4
5
6
Tools that help tuning: In game image viewing/filtering tools Photoshop scripts Debug shaders Debug/Reviewing environments (lighting) Artists need to be able to isolate issues, not judge only the final result.
7
Aliasing ... Try to crank up resolution and AA on pc games, see how much more realistic they look!
8
Photorealistic scenes can be found across generations... From top left, clockwise: Half-Life 2, Call of duty 4 Modern Warfare, Grand Theft Auto 4, Left for dead, Call of Juarez – Bound in blood, Tourist Trophy (PS2), Gran Turismo 4 (PS2) Most screenshots are from “Dead End Thrills” This is a selection of “older” games but I’m cheating here: it’s all about environments. Humans are harder. Videogame photography links: http://enwandrews.tumblr.com/ http://shotbyrobert.com/wordpress/?page_id=1011 http://deadendthrills.com/
9
Some big changes are still perceptually believable. The black and white image also has more global and local contrast. Global blurring does not change our ability of recognizing the “human” quality of the face (but blurring only the skin detail while retaining the edges would for example!) Most of the visual quality are relative. The lips in the bottom-left picture appear to be redder than the one on the top, even if only the skin around them has a slight cool tint, the lips did not change. Also some “minor” changes in absolute values can make huge differences. The bottommid image has a bit of green added to the midtones, the bottom-right one only changed the reds making them more pink.
10
We need photorealistic visual targets – photos, collages, high-quality renderings. Concept art won’t cut it, as it’s good to evaluate some aspects (volume, composition) but not all of what’s needed to generate a photorealistic image (colour, texture, details)
11
12
13
Possible solutions: - Re-compute normals from skinned geometry It’s tricky as the geometry in game can be in a different format and use different rules than the original in your DCC 3d modeling package. I.E. Quads vs Triangles, subdivision surfaces, projection of the normals from a high-res model into low-res etc... -Compute an extended set of weights for normal skinning. Let's say we derive face normals by averaging the vertices of a face. And that we compute vertex normals by some form of averaging of the faces of a vertex (usually, weighted by the areas). Let's assume that the face areas do not change under skinning. In that case we can compute a set of bone weights and indices that is the average of the bone weights and indices that act upon the faces of a given vertex.
14
15
16
17
18
19
Shaders are some sort of black box. They model a given function (BRDF, or better, a given approximation of the whole rendering equation), and each parameter can change part of the function’s shape. The art-direction and visual targets define an “objective function”, that we don’t know mathematically in any form. Artists are “optimizers”, they try to find the parameters that best fit the shader function to the art-direction. More parameters, more degrees of freedom = harder work for the artists. We need to guide the artists, provide intuitive models, limit parameter ranges and dependencies. Math and physics guides our process, they make us understand what to add and why, if a parameter makes sense or not, what is the root cause of a given visual defect.
20
Adding more parameters/hacks -- Increases complexity. Over time shaders become unwieldly, it becomes hard to understand what causes what. -- Feedback loop / a problem that makes itself worse. The more complex the shader, the less it’s understood, the more we will need to add ad-hoc hacks when things go wrong. -- Performance problems -- Makes changes harder: it becomes actually HARDER to art-direct, it becomes harder to iterate and tune (NAIL THE LOOK). Everyone is shifting towards less parameters, more physically based models. Many talks at Siggraph2010 were about GI in movies and its advantages ...It would be a nice project to create an automatic shader-tuner, that fits parameters to a given reference image. The right comparison metrices are needed, and there is no hope to make it work over all the dimensions of a typical shader parameter space, but it could be possible for example to fit SH lighting having fixed the material parameters for example.
21
22
23
Not only less parameters, coarser granularity! Material parameters: Skin in FN4 was specified mostly per-skintone, sometimes per boxer Skin in FNC is mostly tuned globally, only 2-3 parameters per skintone, none per boxer Lighting parameters: FN4 used 3-4 lighting cubemaps per venue, FNC only one FN4 tuned lighting per scenario (i.e. Entrances vs In-Game), FNC practically doesn’t. Cubemaps are SLOW to iterate on! Artist’s hacks: bending the parameters (the shader “shape”) too much, trying to fix a given visual problem with the wrong controls. “Encourages” communication: if a given visual problem is not fixable within the current parameter ranges, talk to an engineer Example of “artist friendly” parameters: Normalize specular so that changing exponent does not change the intensity It can also make some computation faster, i.e. Translate brightness/contrast parameters in a single mAdd Parameter ranges: Avoid having artists “pushing” parameters to levels that do not make sense. If they feel that need either there is a misunderstanding on how the lighting works, or the lighting is not working correctly! Texture parameter “space”: Gamma, different encodings and compression settings. Author textures in 16bit, find the best representation in the pipeline. Author texture separately, merge and mix them in the pipeline (i.e. Multiple grayscale texture merged in different channels of a single in-game one, AO maps pre-multiplied into diffuse and specular etc...)
24
25
26
27
28
29
30
Bad math in the specular: We tinted specular using diffuse lighting color, this actually decreased hue variety. Standard diffuse does not work. In FN4, in-game we used an ambient map (redness) to simulate SSS. During the NIS this map was replaced by a dynamically generated one (UV-space blur) This solution can’t work with more varied lighting setups (FN4 is almost always strongly top-lit). Also the map was not entirely reasonable the way it was authored. Sidenote: Noise is good to break highlights, if we wanted to simulate peach-fuzz it would be good to add a bit of noise in the specular at grazing angles
31
Don’t judge too much:
Don’t let your experience hinder your immagination Not the time to be strictly practical Tinker, Try, Hack, Read, Try Look at other medias After this, all the walls around my desk were cluttered with papers, screenshots, ideas
32
33
Performance problems: We need to compute diffuse lighting twice, first for SSS, then for the direct component Problems if we need to render many characters. FN4 is the ideal case, in game we have only two boxers Screenshots: Jensen’s multipole SSS model, D’Eon-Luebke realtime implementation
34
35
36
It’s possible to use curvature to approximate the geometrical neighborhood. We didn’t do that in our test. Similar to: “Curvature-Dependent Local Illumination Approximation for Translucent Materials” Kubo-Hariu-Wemler-Morishima And “Curvature-Dependent Reflectance Function for Interactive Rendering of Subsurface Scattering” “Curvature-based shading of translucent materials, such as human skin” The most recent and complete effort is “Pre-Integrated Skin Shading” by Eric Penner: http://www.ericpenner.net/portfolio/
37
38
39
Avoid blurring: Weight occlusion based on sample distance. Line sampling. Bent normals: did not make a big enough difference for us, rencently Crysis2 shipped with it Consider normals: Sample in 3D -> Orient sampling hemisphere along surface normals Sample in 2D -> Consider distance between normal-plane and sampled depth. Specular occlusion: sample on a line towards the specular reflection. Useful especially on the rim (fresnel angles)!
40
41
42
Power after cube fetch Wrong: can’t really change the exponent, only changes contrast. Dependant on the HDR range of the cubemap. Such contrast adjustments shift colors (saturation)!
43
Encoding multiple phong exponents in the mips 360 and PS3 (and older PC cards) do not filter across cube faces -> ATI Cubemapgen library solves this by making sure that the borders of the cube faces match Your DXT compressor has to be aware of this too, use the same reference colors for blocks across the seams! Cubemapgen does not does that
44
We like phong because can be baked in cubemaps!
45
46
47
48
49
50
51
We use a few “ubershaders” with static defines for features, so it’s easy to produce specializations
52
53
54
55
Faster iteration is better than good technology. I don’t think this applies to only to artists! I think any programmer experienced how much more important is to have fast iteration when working with shaders. We don’t have great debugging tools for them (compared to other languages) but being able to hotswap them easily and to have an immediate visual feedback is usually better than having great tools but slow iteration. The same applies to scripting languages. I’ve even found myself more prone to explore an idea with a lesser potential if it was implementable within the boundaries of what I could iterate fast, than better ideas that required slower changes. For FN4 most of our lighters time was spent tuning lighting cubemaps (80%) in photoshop FNC: drastically reduced the number of cubemaps, adopted an in-house tool for their autoring Another big pain were expression-driven corrective bones in Maya, we have many of them. Expressions written in MEL were hand-translated in C++ code! We employed a new node-based scripting engine for these, with a runtime in game and in Maya. Ownership: The more “distance” there is between artists and the game (mainline), the less ownership is possible. Outsourcing works too. The artists responsible for a given outsourcer does also effectively own their external changes, test them and check them into mainline 56
57
58
We added a fairly complex motion blur to Champion. It’s does not only blur camera movements, but also skinned actor ones. It’s also capable of blurring outside the boxer silouhette, and it’s fairly robust. We did some tests very early in prepro using optical flow on a filmed version of the game to quickly add the effect and we preferred that. Later on we retested our hypotesis using an actual prototype in game and blind testing. Most of the pre-pro was still done thinking about techniques that could work at 60hz. This was not a huge bottleneck as our worst case performance in FN4 and in FNC turns out to be the in-game NIS, that even in FN4 ran at 30hz. In-game FNC could still run at 60hz even if it wasn’t optimized for it. SSDO could still be a good technique. It requires better tools, light bleeding could be solved in the future with more accurate ray-casting, importance sampling according to the cubemap etc... Some examples of effects driven by the analytic main light: Ear translucency can be faked with a cheap trick, we boosted specular cubemaps with it and so on.. Again, accurate occlusion is Very important for shapes and volumes -> likeness
59
Hybrid Point/Directional lights are nice. FN4 flashes were a product of both IIRC (dot(N,D)*dot(N,P)), works nicely.
60
61
62
Having an analytic light gets us INFORMATION, we can use it to drive many more effects. Ear fake translucency is just a light set opposite to the main and masked with an ears mask…
63
It’s easy to break the rendering realism by having visible artifacts, especially aliasing or upsampling defects...
64
65
66
67
68
69
Note: FNC SSAO occludes specular, main light diffuse (using min(shadow,ao)) and ambient. It’s used as a shadowing component, not only as ambient occlusion.
70
71
72
73
74
75
