efylmzr
commented
2 years ago Disney Fresnel
Lately, I observed another problem related to Disney Fresnel Function while creating images for my report. See the following: (do not mind flatness, the images are wrong there.):
If you look at the spec_tint parameter, you can see that there is no change in the images as the parameter changes. It is not related to the lighting or selected parameters but implementation (I was using the same implementation with PBRT). In Disney 2012 course notes, it is stated that schlick approximation is used for all specular lobes. It is calculated as: F_schlick = F0 + (1-F0) (1- cos(i)^5) They calculated F0 as follows: F0 = metallic base_color + (1-metallic) F0_dielectric_eta ( (1-spec_tint) + spec_tint c_tint) where c_tint = color / luminosity and F0_dielectric_eta is calculated with eta. So for metallic part of F0, base_color is directly used. For dielectric part, spec_tint amount of F0_dielectric is shifted by c_tint parameter which is a normalized color.
Everything is fine so far. In 2015 course notes they decided to use the correct Fresnel implementation for dielectric part where as specular_tint and metallic is still approximated by Schlick. But how they did it is ambiguous. (It is not clear how to do that because they blended F0 parameter in 2012 where true implementation of F_dielectric gives the result for all angles so can not be blended with F0). For this problem, PBRT applies, F = F_true_dielectric (1-metallic) + metallic F_schlick where F_schlick is the same value given above. This is nonsense since F_schlick above also takes into account dielectric part. So when metallic=0, we can not see any tint in specular dielectric.
Since they said that specular tint is also evaluated with schlick approximation, I applied:
F =F_true_dielectric (1-metallic) (1-spec_tint) + F_schlick_tint (1-metallic) spec_tint + metallic F_schlick_metallic
where
F_schlick_metallic = color + (1-color) (1- cos(i)^5)
F_schlick_tint = F0_dielectric_eta c_tint + (1-F0_dielectric_eta c_tint ) * (1- cos(i)^5)
I am just applying schlick approximation twice (for specular tinted dielectric part and metallic part.) So if spec_tint is given as 1, true value of Fresnel function is not used at all. After applying this, I got the following:
So this result is more correct, bot I do not know the true implementation.
Clearcoat Masking Function
Clearcoat masking function does not match the correct formula.
PotatoPot6676
gracien-app
Hi, I am currently implementing the Disney BSDF in Mitsuba as a project, and I was using PBRT as my baseline implementation. Apparently PBRT has some issues, especially in transmission and thin sheet approximation. I will try to cover them all. Also there are some indetermined issues which are not explained well (or not explained at all) in Disney course notes. I saw in the comments of the code that you communicated with Brent Burley, maybe you know the correct answers.
Schlick Approximation
This is a small mistake, the ordering in
lerp()function is wrong inFr_Schlick()function. It should bereturn Lerp(R0,SchlickWeight(cosTheta), 1);. instead ofreturn Lerp(SchlickWeight(cosTheta), R0, 1);Also this Schlick approximation is wrong if eta<1. It should be adjusted in case th rendering is done under water :).
BSDF Lobe
The rates for some lobes are different from what is explained in the paper. For example, the rate for Microfacet Transmission should be (1-metallic)*spec_trans . here is a rendering with metallic=1, spec_trans=1 :
The reason for the bad image is that transmission lobe is evaluated which should not be when metallic parameter is 1.
I created a flow chart for my report which can be useful.
Also sheen is not present in BSDF lobe (in PBRT implementation) which is explicitly mentioned in the course notes.
Well, from what I have understood from the 2015 Disney course notes, when all of the other lobes are turned of except BSDF , we should get the same thing as rough dielectric implementation with GGX distribution. When I render that configuration in PBRT, I get the following:
PBRT uniformly selects one of the active lobes while sampling so the convergence is pretty bad . But the main problem here is not taking care of the total internal reflection cases. Every time a total internal reflection occurs, half of the samples (in this configuration since only Microfacet reflection and transmission is turned on) are wasted, and paths can not reach anywhere for some points. After taking care of that issue in Mitsuba, I got the following (again equal probability between lobes) :
I also blended the brdf and bsdf lobe (metallic=0, spec_trans = 0 to 1) and created some renderings. With PBRT, I got this :
I was also getting the same thing with my implementation. The problem here is sampling and evaluating all of the lobes when the light is inside. (wi.z() < 0). Evaluating diffuse or retroreflection for internal reflection case did not seem right to me.(Course notes does not explain anything about this.) The most intuitive thing to do was only turning on BSDF minor lobes (MF Reflection, MF Transmission, Clearcoat and Sheen (I did not sample sheen though, just evaluated) ) when the light is inside. So when I tried that, I got the following in Mitsuba: (with importance sampling based on material parameters.):
The I used Fresnel coefficients while sampling. I got this:
Complete glass configuration was even better (it actually becomes same as the rough dielectric implementation in glass configuration):
Everything was converging quiet faster, I was getting nice looking images and it was passing most of the chi2 tests. The life was good :) until I did a rendering of a glass with clearcoat and sheen :
I do not know what is hapening here, but when I use clearcoat and sheen components for inside, everywhere is full of fireflies. (Sheen is understandable, I do not sample it.) Maybe sampling 2 different specular lobe for inside is too complicated due to internal reflections etc. In the end, I discarded sheen and clearcoat for inside case ...
Thin Model
There are only two paragraphs related to the thin model. So I am not sure about this part also but it seems that the paper is talking about a different model than the main BSDF. One thing they say in the paper is that spec_trans parameter fully blends between a fully diffuse model and a fully specular model. PBRT implementation is different from this info. When I render a thin sheet with PBRT with spec_trans = 0, I get this:
which still has specular parts. It uses the main BSDF model lobe coefficients. I think, the coefficients should be like this:
There is no mention to the clearcoat and metallic parts so I just excluded them. When I render thin sheets with spec_trans = 0 .. 1, I get this with my implementation:
Another issue is that PBRT uses normal Microfacet Transmission implementation which is not the case. This creates bending when we are going through the thin sheet. For a thin glass, such effect is not present. Instead of that, MicrofacetReflection should be sampled and inverted to the back side of the sheet. And instead of asymmetric BRDF function, (1-F) D G / (4cos(i)cos(o)) must be used. To see the effect of bending, check this rendering (PBRT):
Also, instead of multiplying the square root of the color, we should multiply the true color value in bsdf evaluation when refracting. (Sqrt is for 3D case. After 2 refractions, true color value is multiplied. )
These are all the potential problems that I see. If you think that there is something wrong above, please share info.