Note：UE5源码解析-数字人渲染篇

1.皮肤

关键文件

Engine\Shaders\Private\SeparableSSS.ush 主要用来实现4S的Shader,主要是用在后处理是对屏幕空间做处理
Engine\Source\Runtime\Engine\Private\Rendering\BurleyNormalizedSSS.cpp 主要计算高斯核
Engine\Source\Runtime\Engine\Private\Rendering\SubsurfaceProfile.cpp 对纹理做预处理

论文：

Approximate Reflectance Profiles for Efficient Subsurface Scattering https://graphics.pixar.com/library/ApproxBSSRDF/paper.pdf
Separable Subsurface Scattering http://iryoku.com/separable-sss/
Real-Time Realistic Skin Translucency http://www.iryoku.com/translucency/downloads/Real-Time-Realistic-Skin-Translucency.pdf

Engine\Source\Runtime\Engine\Private\Rendering\BurleyNormalizedSSS.cpp

这里是有一个纯C++的开源项目https://github.com/iryoku/separable-sss

//用两个指数函数的和除以距离R，近似的逼近漫反射曲线，见Approximate Reflectance Profiles for Efficient Subsurface Scattering论文
inline float Burley_ScatteringProfile(float r, float A,float S, float L)
{   //2PIR(r)rfloat D = 1 / S;float R = r / L;const float Inv8Pi = 1.0 / (8 * PI);float NegRbyD = -R / D;float RrDotR = A*FMath::Max((exp(NegRbyD) + exp(NegRbyD / 3.0)) / (D*L)*Inv8Pi, 0.0);return RrDotR;
}

//拟合曲线
float GetSearchLightDiffuseScalingFactor(float SurfaceAlbedo)
{return 3.5 + 100 * FMath::Pow(SurfaceAlbedo - 0.33, 4);
}

void ComputeMirroredBSSSKernel(FLinearColor* TargetBuffer, uint32 TargetBufferSize,FLinearColor SurfaceAlbedo, FLinearColor DiffuseMeanFreePath, float ScatterRadius)
{check(TargetBuffer);check(TargetBufferSize > 0);uint32 nNonMirroredSamples = TargetBufferSize;int32 nTotalSamples = nNonMirroredSamples * 2 - 1;//这里的计算见Approximate Reflectance Profiles for Efficient Subsurface Scattering这篇论文,公式（8）FVector ScalingFactor = GetSearchLightDiffuseScalingFactor(SurfaceAlbedo);// we could generate Out directly but the original code form SeparableSSS wasn't done like that so we convert it later// .a is in mmcheck(nTotalSamples < 64);FLinearColor kernel[64];{const float Range = (nTotalSamples > 20 ? 3.0f : 2.0f);// tweak constantconst float Exponent = 2.0f;// Calculate the offsets:float step = 2.0f * Range / (nTotalSamples - 1);for (int i = 0; i < nTotalSamples; i++){float o = -Range + float(i) * step;float sign = o < 0.0f ? -1.0f : 1.0f;kernel[i].A = Range * sign * FMath::Abs(FMath::Pow(o, Exponent)) / FMath::Pow(Range, Exponent);}// Center sample should always be zero, but might not be due to potential roundoff error.kernel[nTotalSamples / 2].A = 0.0f;//Scale the profile sampling radius. This scale enables the sampling between [-3*SpaceScale,+3*SpaceScale] instead of //the default [-3,3] range when fetching kernel parameters.const float SpaceScale = ScatterRadius * 10.0f;// from cm to mm// Calculate the weights:for (int32 i = 0; i < nTotalSamples; i++){float w0 = i > 0 ? FMath::Abs(kernel[i].A - kernel[i - 1].A) : 0.0f;float w1 = i < nTotalSamples - 1 ? FMath::Abs(kernel[i].A - kernel[i + 1].A) : 0.0f;float area = (w0 + w1) / 2.0f;//这里的计算见Approximate Reflectance Profiles for Efficient Subsurface Scattering这篇论文,公式（2）FVector t = area * Burley_ScatteringProfile(FMath::Abs(kernel[i].A)*SpaceScale, SurfaceAlbedo, ScalingFactor,DiffuseMeanFreePath);kernel[i].R = t.X;kernel[i].G = t.Y;kernel[i].B = t.Z;}// We still need to do a small tweak to get the radius to visually match. Multiplying by 4.0 seems to fix it.const float StepScale = 4.0f;for (int32 i = 0; i < nTotalSamples; i++){kernel[i].A *= StepScale;}// We want the offset 0.0 to come first:FLinearColor t = kernel[nTotalSamples / 2];for (int i = nTotalSamples / 2; i > 0; i--){kernel[i] = kernel[i - 1];}kernel[0] = t;// Normalize the weights in RGB{FVector sum = FVector(0, 0, 0);for (int i = 0; i < nTotalSamples; i++){sum.X += kernel[i].R;sum.Y += kernel[i].G;sum.Z += kernel[i].B;}for (int i = 0; i < nTotalSamples; i++){kernel[i].R /= sum.X;kernel[i].G /= sum.Y;kernel[i].B /= sum.Z;}}/* we do that in the shader for better quality with half res// Tweak them using the desired strength. The first one is://     lerp(1.0, kernel[0].rgb, strength)kernel[0].R = FMath::Lerp(1.0f, kernel[0].R, SubsurfaceColor.R);kernel[0].G = FMath::Lerp(1.0f, kernel[0].G, SubsurfaceColor.G);kernel[0].B = FMath::Lerp(1.0f, kernel[0].B, SubsurfaceColor.B);for (int i = 1; i < nTotalSamples; i++){kernel[i].R *= SubsurfaceColor.R;kernel[i].G *= SubsurfaceColor.G;kernel[i].B *= SubsurfaceColor.B;}*/}// generate output (remove negative samples){// center sampleTargetBuffer[0] = kernel[0];// all positive samplesfor (uint32 i = 0; i < nNonMirroredSamples - 1; i++){TargetBuffer[i + 1] = kernel[nNonMirroredSamples + i];}}
}

Engine\Shaders\Private\SeparableSSS.ush 可分离的次表面散射（Separable Subsurface Scattering）

//-----------------------------------------------------------------------------
// Separable SSS Reflectance Pixel Shader// @param texcoord The usual quad texture coordinates.
// @param dir Direction of the blur: First pass:   float2(1.0, 0.0), Second pass:  float2(0.0, 1.0)  调节横向和纵向
// @param initStencil indicates whether the stencil buffer should be initialized. Should be set to 'true' for the first pass if not previously initialized, to enable optimization of the second pass.
float4 SSSSBlurPS(uint2 BufferPos, float2 BufferUV, float2 dir, bool initStencil, float2 Extent)
{// Fetch color of current pixel:float4 colorM = SSSSSampleSceneColorPoint(BufferUV);// we store the depth in alphafloat OutDepth = colorM.a;colorM.a = GetMaskFromDepthInAlpha(colorM.a);// we don't need to process pixels that are not part of the subsurafce scattering (optimization, also prevents divide by later on)BRANCH if(!colorM.a){// todo: need to check for proper clear
//      discard;return 0.0f;}// 0..1float SSSStrength = GetSubsurfaceStrength(BufferUV);#if !SSSS_COMPUTESHADER// Initialize the stencil buffer in case it was not already available:if (initStencil) // (Checked in compile time, it's optimized away)if (SSSStrength < 1 / 256.0f) discard;
#endiffloat SSSScaleX = SubsurfaceParams.x;float scale = SSSScaleX / OutDepth;// Calculate the final step to fetch the surrounding pixels:float2 finalStep = scale * dir;// ideally this comes from a half res buffer as well - there are some minor artifactsfinalStep *= SSSStrength; // Modulate it using the opacity (0..1 range)FSeparableFilterParameters SeparableFilterParameters;#if STRATA_ENABLEDconst FStrataSubsurfaceData SSSData = LoadStataSSSData(BufferUV);// For SSS_PROFILE_ID_PERPIXEL is managed throught the burley passesconst uint SubsurfaceProfileInt = SSSData.bIsProfile ? SSSData.ProfileId : SSS_PROFILE_ID_PERPIXEL;SeparableFilterParameters.SubsurfaceProfileInt = SubsurfaceProfileInt;FillBurleyParameters(SSSData, SeparableFilterParameters);
#elseconst FGBufferData GBufferData = GetGBufferData(BufferUV);// 0..255, which SubSurface profile to pick// ideally this comes from a half res buffer as well - there are some minor artifactsconst uint SubsurfaceProfileInt = ExtractSubsurfaceProfileInt(GBufferData);SeparableFilterParameters.SubsurfaceProfileInt = SubsurfaceProfileInt;
#endif// Accumulate the center sample:float3 colorAccum = 0;// >0 to avoid division by 0, not 100% correct to not visiblefloat3 colorInvDiv = 0.00001f;// center samplehalf3  CentralKernelWeight = GetSubsurfaceProfileKernel(0, SeparableFilterParameters).rgb;colorInvDiv += CentralKernelWeight;colorAccum = colorM.rgb * CentralKernelWeight;float3 BoundaryColorBleed = GetSubsurfaceProfileBoundaryColorBleed(SubsurfaceProfileInt);// Accumulate the other samples:SSSS_UNROLLfor (int i = 1; i < SSSS_N_KERNELWEIGHTCOUNT; i++) {// Kernel.a = 0..SUBSURFACE_KERNEL_SIZE (radius)half4 Kernel = GetSubsurfaceProfileKernel(i, SeparableFilterParameters);float4 LocalAccum = 0;float2 UVOffset = Kernel.a * finalStep;// The kernel is symtrical, we want to use that property.// Half the GetSubsurfaceProfileKernel() calls (more expensive if done by texture)// Half the weighting computations (saves 3mul per lookup sample)SSSS_UNROLLfor (int Side = -1; Side <= 1; Side += 2){// Fetch color and depth for current sample:float2 LocalUV = BufferUV + UVOffset * Side;float4 color = SSSSSampleSceneColor(LocalUV);uint LocalSubsurfaceProfileInt = SSSSSampleProfileId(LocalUV);float3 ColorTint = LocalSubsurfaceProfileInt == SubsurfaceProfileInt ? 1.0f : BoundaryColorBleed;float LocalDepth = color.a;color.a = GetMaskFromDepthInAlpha(color.a);#if SSSS_FOLLOW_SURFACE == 1// If the difference in depth is huge, we weight the sample less or not at allfloat s = saturate(12000.0f / 400000 * SubsurfaceParams.y *//        float s = saturate(300.0f/400000 * SubsurfaceParams.y *abs(OutDepth - LocalDepth));color.a *= 1 - s;
#endif// approximation, ideally we would reconstruct the mask with GetMaskFromDepthInAlpha() and do manual bilinear filter// needed?color.rgb *= color.a * ColorTint;// Accumulate left and right LocalAccum += color;}// Accumulate to final value (left and right sample with the same weight)colorAccum += Kernel.rgb * LocalAccum.rgb;colorInvDiv += Kernel.rgb * LocalAccum.a;}// normalize (some samples are rejected because of depth or the other material is no SSS, compensate for that)// done for each color channel to avoid color shiftfloat3 OutColor = colorAccum / colorInvDiv; // alpha stored the SceneDepth (0 if there is no subsurface scattering)return float4(OutColor, OutDepth);
}

渲染流程：BSDF （散射）= BRDF（反射） + BTDF（透射）

1.通过得到的高斯核，对纹理进行预处理

2.经过着色后，对屏幕的像素做处理

2.毛发

关键文件：

物理模拟部分代码：Engine\Plugins\Runtime\HairStrands
渲染部分代码：Engine\Shaders\Private\HairStrands 和 Engine\Source\Runtime\Renderer\Private\HairStrands

论文：

Light Scattering from Human Hair Fibers http://graphics.stanford.edu/papers/hair/hair-sg03final.pdf
A Data-Driven Light Scattering Model for Hair https://graphics.pixar.com/library/DataDrivenHairScattering/paper.pdf
Dual Scattering Approximation for Fast Multiple Scattering in Hair http://www.cemyuksel.com/research/dualscattering/dualscattering.pdf

Engine\Shaders\Private\HairBsdf.ush

float Hair_g(float B, float Theta)
{return exp(-0.5 * Pow2(Theta) / (B * B)) / (sqrt(2 * PI) * B);
}float Hair_F(float CosTheta)
{const float n = 1.55;const float F0 = Pow2((1 - n) / (1 + n));return F0 + (1 - F0) * Pow5(1 - CosTheta);
}

///
// Hair BSDF
// Approximation to HairShadingRef using concepts from the following papers:
// [Marschner et al. 2003, "Light Scattering from Human Hair Fibers"]
// [Pekelis et al. 2015, "A Data-Driven Light Scattering Model for Hair"]float3 HairShading( FGBufferData GBuffer, float3 L, float3 V, half3 N, float Shadow, FHairTransmittanceData HairTransmittance, float InBacklit, float Area, uint2 Random )
{// to prevent NaN with decals// OR-18489 HERO: IGGY: RMB on E ability causes blinding hair effect// OR-17578 HERO: HAMMER: E causes blinding light on heroes with hairfloat ClampedRoughness = clamp(GBuffer.Roughness, 1/255.0f, 1.0f);//const float3 DiffuseN   = OctahedronToUnitVector( GBuffer.CustomData.xy * 2 - 1 );const float Backlit  = min(InBacklit, HairTransmittance.bUseBacklit ? GBuffer.CustomData.z : 1);#if HAIR_REFERENCE// todo: ClampedRoughness is missing for this code pathfloat3 S = HairShadingRef( GBuffer, L, V, N, Random );//float3 S = HairShadingMarschner( GBuffer, L, V, N );
#else// N is the vector parallel to hair pointing toward root//论文Light Scattering from Human Hair Fibers给出的经验值定义，用于后面的R,TT,TRT计算const float VoL       = dot(V,L);                                                      const float SinThetaL = clamp(dot(N,L), -1.f, 1.f);const float SinThetaV = clamp(dot(N,V), -1.f, 1.f);float CosThetaD = cos( 0.5 * abs( asinFast( SinThetaV ) - asinFast( SinThetaL ) ) );//CosThetaD = abs( CosThetaD ) < 0.01 ? 0.01 : CosThetaD;const float3 Lp = L - SinThetaL * N;const float3 Vp = V - SinThetaV * N;const float CosPhi = dot(Lp,Vp) * rsqrt( dot(Lp,Lp) * dot(Vp,Vp) + 1e-4 );const float CosHalfPhi = sqrt( saturate( 0.5 + 0.5 * CosPhi ) );//const float Phi = acosFast( CosPhi );float n = 1.55;//float n_prime = sqrt( n*n - 1 + Pow2( CosThetaD ) ) / CosThetaD;float n_prime = 1.19 / CosThetaD + 0.36 * CosThetaD;float Shift = 0.035;float Alpha[] ={-Shift * 2,Shift,Shift * 4,};  float B[] ={Area + Pow2(ClampedRoughness),Area + Pow2(ClampedRoughness) / 2,Area + Pow2(ClampedRoughness) * 2,};float3 S = 0;//Mp是纵向散射函数，Np是方位角散射函数，Fp是菲涅尔函数，Tp是吸收函数//Rif (HairTransmittance.ScatteringComponent & HAIR_COMPONENT_R){const float sa = sin(Alpha[0]);const float ca = cos(Alpha[0]);float Shift = 2 * sa * (ca * CosHalfPhi * sqrt(1 - SinThetaV * SinThetaV) + sa * SinThetaV);float BScale = HairTransmittance.bUseSeparableR ? sqrt(2.0) * CosHalfPhi : 1;float Mp = Hair_g(B[0] * BScale, SinThetaL + SinThetaV - Shift);float Np = 0.25 * CosHalfPhi;float Fp = Hair_F(sqrt(saturate(0.5 + 0.5 * VoL)));S += Mp * Np * Fp * (GBuffer.Specular * 2) * lerp(1, Backlit, saturate(-VoL));}// TTif (HairTransmittance.ScatteringComponent & HAIR_COMPONENT_TT){float Mp = Hair_g( B[1], SinThetaL + SinThetaV - Alpha[1] );float a = 1 / n_prime;//float h = CosHalfPhi * rsqrt( 1 + a*a - 2*a * sqrt( 0.5 - 0.5 * CosPhi ) );//float h = CosHalfPhi * ( ( 1 - Pow2( CosHalfPhi ) ) * a + 1 );float h = CosHalfPhi * ( 1 + a * ( 0.6 - 0.8 * CosPhi ) );//float h = 0.4;//float yi = asinFast(h);//float yt = asinFast(h / n_prime);float f = Hair_F( CosThetaD * sqrt( saturate( 1 - h*h ) ) );float Fp = Pow2(1 - f);//float3 Tp = pow( GBuffer.BaseColor, 0.5 * ( 1 + cos(2*yt) ) / CosThetaD );//float3 Tp = pow( GBuffer.BaseColor, 0.5 * cos(yt) / CosThetaD );float3 Tp = 0;if (HairTransmittance.bUseLegacyAbsorption){Tp = pow(abs(GBuffer.BaseColor), 0.5 * sqrt(1 - Pow2(h * a)) / CosThetaD);}else{// Compute absorption color which would match user intent after multiple scatteringconst float3 AbsorptionColor = HairColorToAbsorption(GBuffer.BaseColor);Tp = exp(-AbsorptionColor * 2 * abs(1 - Pow2(h * a) / CosThetaD));}//float t = asin( 1 / n_prime );//float d = ( sqrt(2) - t ) / ( 1 - t );//float s = -0.5 * PI * (1 - 1 / n_prime) * log( 2*d - 1 - 2 * sqrt( d * (d - 1) ) );//float s = 0.35;//float Np = exp( (Phi - PI) / s ) / ( s * Pow2( 1 + exp( (Phi - PI) / s ) ) );//float Np = 0.71 * exp( -1.65 * Pow2(Phi - PI) );float Np = exp( -3.65 * CosPhi - 3.98 );S += Mp * Np * Fp * Tp * Backlit;}// TRTif (HairTransmittance.ScatteringComponent & HAIR_COMPONENT_TRT){float Mp = Hair_g( B[2], SinThetaL + SinThetaV - Alpha[2] );//float h = 0.75;float f = Hair_F( CosThetaD * 0.5 );float Fp = Pow2(1 - f) * f;//float3 Tp = pow( GBuffer.BaseColor, 1.6 / CosThetaD );float3 Tp = pow(abs(GBuffer.BaseColor), 0.8 / CosThetaD );//float s = 0.15;//float Np = 0.75 * exp( Phi / s ) / ( s * Pow2( 1 + exp( Phi / s ) ) );float Np = exp( 17 * CosPhi - 16.78 );S += Mp * Np * Fp * Tp;}
#endifif (HairTransmittance.ScatteringComponent & HAIR_COMPONENT_MULTISCATTER){S  = EvaluateHairMultipleScattering(HairTransmittance, ClampedRoughness, S);S += KajiyaKayDiffuseAttenuation(GBuffer, L, V, N, Shadow);}S = -min(-S, 0.0);return S;
}

HairStrands（Marschner模型）S=反射（R）+透射（TT）+次反射（TRT）

参考链接：

《GPU Gems 3》：真实感皮肤渲染技术总结 https://zhuanlan.zhihu.com/p/42433792
Approximate Reflectance Profiles for Efficient Subsurface Scattering https://graphics.pixar.com/library/ApproxBSSRDF/paper.pdf
Separable Subsurface Scattering http://iryoku.com/separable-sss/
Real-Time Realistic Skin Translucency http://www.iryoku.com/translucency/downloads/Real-Time-Realistic-Skin-Translucency.pdf
UE4 Hair Strands浅析 https://zhuanlan.zhihu.com/p/128669105
Light Scattering from Human Hair Fibers http://graphics.stanford.edu/papers/hair/hair-sg03final.pdf
A Data-Driven Light Scattering Model for Hair https://graphics.pixar.com/library/DataDrivenHairScattering/paper.pdf
Dual Scattering Approximation for Fast Multiple Scattering in Hair http://www.cemyuksel.com/research/dualscattering/dualscattering.pdf
tp://www.iryoku.com/translucency/downloads/Real-Time-Realistic-Skin-Translucency.pdf

Note：UE5源码解析-数字人渲染篇相关推荐

美团动态线程池实践思路开源项目（DynamicTp），线程池源码解析及通知告警篇
大家好,这篇文章我们来聊下动态线程池开源项目(DynamicTp)的通知告警模块.目前项目提供以下通知告警功能,每一个通知项都可以独立配置是否开启.告警阈值.告警间隔时间.平台等,具体代码请看core ...
MACE源码解析【ARM卷积篇(一) 】1*N和N*1卷积实现
MACE Mobile AI Compute Engine (MACE) 是一个专为移动端异构计算平台优化的神经网络计算框架,旨在深度神经网络部署在移动端,是一个SoC上的神经网络实现.主要涉及的硬件 ...
缩放手势 ScaleGestureDetector 源码解析，这一篇就够了
其实在我们日常的编程中,对于缩放手势的使用并不是很经常,这一手势主要是用在图片浏览方面,比如下方例子.但是(敲重点),作为 Android 入门的基础来说,学习 ScaleGestureDetecto ...
面试官系统精讲Java源码及大厂真题 - 33 CountDownLatch、Atomic 等其它源码解析
33 CountDownLatch.Atomic 等其它源码解析每个人的生命都是一只小船,理想是小船的风帆. 引导语本小节和大家一起来看看 CountDownLatch 和 Atomic 打头的原 ...
Zuul源码解析（一）
说在前面我们公司有一个线上服务报错通知群,经常报网关服务的一个 EOFException 异常.这个异常报出来好久了,如下图所示,艾特相关的人也不去处理,大概是不重要异常吧,反正看样子是不影响线上核 ...
mysql 网络io_分布式 | DBLE 网络模块源码解析（一）：网络 IO 基础知识
作者:路路热爱技术.乐于分享的技术人,目前主要从事数据库相关技术的研究. 本文来源:原创投稿 *爱可生开源社区出品,原创内容未经授权不得随意使用,转载请联系小编并注明来源. 前言对于计算机学科来说 ...
clickhouse原理解析与开发实战 pdf_重识SSM，“超高频面试点+源码解析+实战PDF”，一次性干掉全拿走...
重识SSM,"超高频面试点"+"源码解析"+"实战PDF",一次性干掉全拿走!! 01 超高频面试点知识篇 1.1 Spring超高频面试点 ...
Android开发知识（二十二）LayoutInflater装载xml布局过程的源码解析
文章目录前言 LayoutInflater实例 LayoutInflater的装载过程 include 标签解析 merge 标签解析 attachToRoot参数解析 View创建过程 (1)判断 ...
Spring5源码 - 13 Spring事件监听机制_@EventListener源码解析
文章目录 Pre 概览开天辟地的时候初始化的处理器 @EventListener EventListenerMethodProcessor afterSingletonsInstantiated 小 ...

Note：UE5源码解析-数字人渲染篇

Note：UE5源码解析-数字人渲染篇

1.皮肤

2.毛发

Note：UE5源码解析-数字人渲染篇相关推荐

最新文章

热门文章