Physically-Based Shading at Disney

1 introduction

following our success with physically-based hair shading on Tangled [27], we began considering physically based shading models for a broader range of materials. with the physically-based hair model, we were able to achieve a great degree of visual richness while maintaning artistic control. however, it proved challenging to integrate the lighting of the hair with the rest of the scene which had still used traditional “ad-hoc” shading models and punctual lights. for subsequent films we wanted to increase the richness of all our materials while making lighting responses more consistent between materials and environments and also wanted to improve artist produtivity through the use of simplified controls.

when we began our investigation it was not obvious which models to use or even how physically-based we wanted to be. should we be perfectly energy conserving? should we favor pyhsical parameters like index-of-refraction?

for diffuse, Lambert seemed to be the accepted norm, while specular seemed to get most of the attention in the literature. some modeks such as Ashikhmin-Shirley (2000) [3] aimed to be intuitive and practical while physically plausible, while others such as He et al. (1991) [12] provided a more comprehensive physical model. stil others aimed at improved data fitting [xxxx], but few of these are appropriate for direct manipulation. we could have implemented several models and let the artists choose and combine them, but then we would have been back to the parameter explosion we were trying to get away from.

one study of a large variety of measured materials was Ngan et al. (2005) [21] which compared five popular models. some models fared 表现 better than others overall, but interestingly, there was a strong corrleation between the models’ performances 模型的性能——some materials were well represented by all the models, and for others, no model proved suitable. adding an additional specular lobe helped in only a few of the cases. this begs the question, what is not being represented in the difficult materials?

to answer this question and to evaluate BRDF models more intuitively we developed a new BRDF viewer that could display and compare both measured and analytic BRDFs. we discovered new, intuitive ways to view measured BRDF data and we found interesting features in the measured materials that were not well-represented by known models.

in these course notes we will share observations from studying measured materials along with insights we have gleaned 收集 abouth which models fit the measured data and where they fall short. we will then present our new model which is now being used on all current productions. we will also describe our experience of adopting this new model in production and discuss how we were able to add the right level of artistic control while preserving 保存 simplicity and roubustness.

2 the microfacet model
we will define our BRDF and compare with measured materials in terms of the microfacet model[xxx], the microfacet model postulates 假设 that if a surface reflection can occur between a given light vector l and view vector v, then there must exist some portion of the surface, or microfacet, with a normal aligned halfway between the l and v vectors. this “half-vector”, sometimes, referred to as the microsurface normal, is thus defined as h=(l+v)/|l+v|. a general form of the microfacet model for isotropic materials is:

the diffuse term is a function of unknown form. Lambert diffuse is often assumed and is represented by a constant value. for the specular term, D is the microfacet distribution function and is responsible for the shape of hte specular peak, F is the fresnel refection coefficient, and G is the geometric attenuation or shadowing factor.

θl and θv are the angles of incidence of the l and v vectors with respect to the normal, θh is the angle between the normal and the half-vector, and θd is the “difference” angle between l and the half-vector (or, symmetrically, v and h).

most physcially plausible models not specifically described in microfacet form can still be interpreted as microfacet models in that they have a distribution function, a Fresnel factor, and some additional factor which could be considered a geometric shadowing factor. the only real difference between microfacet models and other models is whether they include the explicit 1/(4cos(θl)cos(θv)) factor that comes from the microfacet derivation. for models that do not include this factor, an implied shadowing factor can be determined by multiplying the model by 4cos(θl)cos(θv) after factoring out the D and F factors.

3 visualizing measured BRDFs

3.1 The “MERL 100”

a set of 100 isotropic BRDF material samples was captured by Matusik et al. in 2003 [18] covering a wide range of materials including paints, woods, metals, fabric, stone, rubber, plastic, and other synthetic materials. this data set is freely available from Mitsubishi Electric Reserach Laboratories at www. merl.com/brdf and is commonly used for evaluating new BRDF models. slices of these BRDFs are shown in figure 1.

ecah BRDF in the MERL 100 is densely sampled into a 90 by 90 by 180 cube along the θh, θd, and φd axes respectively. these correspond to 1 degree increments except for the φh axis which was warped to concentrate data samples near the specular peak. the measurements have been filtered and extrapolated as needed so that there are no holes in the data. this is good in that the data is easy to use, but it’s not clear how accurate the data is, particularly near the horizon. because of this, some researchers discard data near the horizon when performing fitting, but this data is still useful to consider as it can have a profound effect on the material appearance.

3.2 BRDF Explorer

to examine the MERL measured materials and compare with analytic models, we developed a new tool, the BRDF explorer, shown in figure 2. it is availabe as open source at github.com/wdas/brdf and has the following features:

1. ability to load multiple analytic BRDFs written in GLSL.
2. ability to load measured BRDFs, including the anisotropic material samples captured by Ngan et al.
3. multiple data plots (3d hemisphereical view, polar plot, and various cartesian plots).
4. computed albedo plot (i.e. directional-hemisphereical reflectance).
5. image slice view with expousre controls

this tool has been invaluable in comparing measured materials with existing analytic models as well as in developing our new model. surprisingly, it has also proven very useful for artists as an interactive BRDF editor, giving them a deeper understanding of the model parameters and BRDF space.

one of the simplest, most intuitive ways to visualize a measured material is to simply view it as a stack of images, and we have found this to be a very powerful tool to gain an intuition about the data. as it turns out, all of the interesting features in the MERL 100 mateirals are visible in the fid = 90 slice.

5.3 diffuse model details
some models include a diffuse fresnel factor such as:

F(theta) is the fresnel factor for reflection.
[note : from the fresnel law for refraction, and to preserve Helmholtz reciprocity, it is necessary to account for refraction twice, once on the way in and once on the way out of of the surface.]
as seen in the measured data observations, and based on our past studio experience, the Lambert diffsue model is often too dark on the edges, and adding a Fresnel factor to make it more physically plausible only makes it darker.兰伯特已经暗了，加上物理的折射，更加暗了。

basedo on our observations, we developed a novel empirical model for diffuse retroreflection that transitions between a diffuse fresnel shadow for smooth surfaces and an added highlight for rough surfaces.

a possible explanation for this effect may be that for rough surfaces light enters and exits the sides of micro-surface features causing an increase in refraction at grazing angles. in any event, our artists like it, and it is similar to features we used to have in our ad-hoc model except that it is now more plausible and has a physical basis.

in our model, we ignore the index of refraction for the diffuse fresnel factor and assume no incident diffuse loss. this allows us to directly specify the incident diffuse color. we use the Schlick Fresnel approximation and modify the grazing retroreflection response to go to a specific value determined from roughness rather than zero.

our base diffuse model is:

this produces a diffuse fresnel shadow that reduces the incident diffuse reflectance by 0.5 at grazing angles for smooth surfaces and increases the response by up to 2.5 for rough surfaces. this seems to provide a reasonable match to MERL data and was also found to be artistically pleasing. BRDF image slices of our model for various roughness values are shown in figure 17.

5.4 specular D details
of the popular models, GGX has the longest tail. this model is in fact equivalent to the Trowbridge-Reitz (1975) distribution favored by Blinn (1977) for its ability to match experimental data. however, this distribution still does not have a long enough tail for many materials.
GGX具有最长的拖尾

Trowbridge and Reitz compared their distribution function along with several other distributions to measurements of ground glass. one of the other distributions, from Berry (1923), has a very similar form but with an exponent of 1 instead of 2 resulting an even longer tail.
什么是拖尾，拖尾长有什么优点，参考：
https://www.zhihu.com/question/48050245/answer/108888213
直观来说GGX最大的特点就是高光周围有很漂亮的拖尾效果
请看现实世界中的高光反射

this suggests a more general distribution with a variable exponent, introduced here and dubbed 被称为 Generalized-Trowbridge-Reitz, or GTR:

in each of these distributions, c is a scaling constant, and alpha is a roughness parameter with values between zero and one;

alpha=0 produces a perfectly smooth distribution (i.e. a delta function at thetah=0 and alpha=1 produces a perfectly rough or uniform distribution).

5.6 speculara G details

This remapping was based on comparisons with measured data as well as artist feedback that the
specular was just \too hot" for small roughness values. This gives us a G function that varies with
roughness, is at least partially physically-based, and seems plausible. For our clearcoat specular we
don’t have a Smith G derivation and simply use the GGX G with a xed roughness of 0.25, found to
be plausible and artistically pleasing.