5 replies to this topic

[lipsryme]

I'm trying to implement a microfacet brdf but my results are wrong somehow.
That's why I was hoping someone can point out what I'm doing wrong or what I'm missing.

-> see below

Edited by lipsryme, 03 November 2012 - 01:05 PM.

[Tombstone]

Hi

I see at least your half vector is wrong.

"float3 H = normalize(-L - V);" should be: float3 H = normalize(-L + V);

[lipsryme]

Actually doing it like that makes it not work at all.
My View direction is being calculated like this:

  // Calculate Position from linear view space depth
  float3 ViewRay = input.PositionVS.xyz;
  float depth = DepthTarget.Sample(PointSampler, input.UV).r;
  float3 ViewDir = ViewRay * depth; // subtraction by the cam position not needed since we're already in view space

Have to say I got it working now with a phong distribution but anything else (the commented out part) gives me wrong results.

  // Compute Fresnel Term (F) using Schlick's approximation
  float LdotH = max(0, dot(-L, H));
  float base = 1 - LdotH;
  float exponent = pow(base, 5.0f);
  float3 F = saturate(SpecularAlbedo + (1 - SpecularAlbedo) * exponent);

  // Lambert
  float3 kd = NdotL;
  Lr.rgb += (kd / PI) * Ei;


  // Normalized Blinn-Phong distribution (D)
  float D = 0.0f;
  float NdotH = max(0, dot(N, H));
  D = ((SpecularPower + 2) / 2 * PI) * pow(NdotH, SpecularPower);

  // Beckmann distribution (D)
  //float NdotH = max(0, dot(N, H));
  //float NdotH2 = NdotH * NdotH;
  //float NdotH4 = NdotH * NdotH * NdotH * NdotH;
  //float SpecularPower2 = SpecularPower * SpecularPower;
  //float Beckmann_exponent = -((1 - NdotH2) / (SpecularPower2 * NdotH2));
  //float D = 1 * exp(Beckmann_exponent) / PI * SpecularPower2 * NdotH4;

  // Trowbridge-Reitz (GGX) distribution
  //float NdotH = max(0, dot(N, H));
  //float SpecPow2 = SpecularPower * SpecularPower;
  //float D = SpecPow2 / PI * pow(((NdotH*NdotH) * (SpecPow2 - 1) + 1), 2);

  // Cook-Torrance approximation (G)
  float NdotV = max(0, dot(N, V));
  float G = 1.0f;

  float3 nominator = F * G * D;
  float denominator = 4 * ((LdotH+0.000001f) * (LdotH+0.000001f));
  Ls.rgb += nominator / denominator;

Beckmann or Trowbridge-Reitz doesn't...

Edited by lipsryme, 03 November 2012 - 01:06 PM.

[Bacterius]

Pretty sure the Beckmann distribution is actually this (and same comment for GGX, I would guess):

float D = 1 * exp(Beckmann_exponent) / (PI * SpecularPower2 * NdotH4);

This is my path tracer code for Beckmann, it uses a raw BRDF so you'll have to modify the inputs (the incident vector goes towards the surface, the exitant vector goes away from it). The vector * vector operation is the dot product. Also I'm not using the tangent formula trick yet whereas you seem to be.. sorry:

http://pastebin.com/pHqCH1dc

Might have gone overboard with the absolute values but I was dealing with refraction at the same time and didn't want to take risks

EDIT: dammit, why are the code tags munching my newlines?
EDIT: pastebin it is!

Edited by Bacterius, 03 November 2012 - 08:39 PM.

[CryZe]

Your GGX distribution is missing a parenthesis. Everything should be in the denominator except the alpha^2. So the correct implementation would be:

float NdotH = max(0, dot(N, H));
float SpecPow2 = SpecularPower * SpecularPower;
float D = SpecPow2 / (PI * pow(((NdotH*NdotH) * (SpecPow2 - 1) + 1), 2));

float G = min(1, min(2 * NdH * NdV / VdH, 2 * NdH * NdL / VdH));

Don't do it this way, you are wasting instructions this way. Do it like that:
float G = min(1, 2 * NdH / VdH * min(NdV, NdL));

[lipsryme]

I'm still getting crazy high values (basically the entire mesh white). Does the input alpha have to be ridiculously high ?
edit: hmm no higher values don't seem to fix this.
edit2: ok so when only outputting D or D*F it looks more like it should (still very bright). Is there something wrong with my Geometry term ?

Edited by lipsryme, 05 November 2012 - 09:02 AM.