Thought I'd stop by and post something that I've been working on for the last while. A skin shader that aims to make realistic faces while maintaining high performance and avoiding texture/image space diffusion; I'm running this in CryEngine2 so I'm pretty limited in what I can do. I can't do lightmap passes so texture space diffusion is out of the question (and it's slow) and I can't make custom post-process stages for image space diffusion, thus making both methods nigh impossible.
So for those of you who are interested in a cheap alternative to diffusion approximations and the likes, read on!
I've been working around these limits by approximating the core concepts of various rendering models (basically just creating layers and faking blurs heh). Here's a render from a few days ago (a few things have changed but the shader is mostly the same):
Head model available here.
I'm using the Kelemen Szirmay-Kalos for my BRDF as proposed by NVIDIA. I'm using 4 terms for my specular model (instead of the 1 used on that page), using 4 beckmann distributions (saved to a texture) and weighing them at the at the end with a simple dot product.
Specularity:
For the actual subsurface scattering I do a mix of things. I use a multi-layer system (like those found in popular modeling programs like Maya or 3ds Max, for example the MentalRay FastSkin shader). I also use blended normals for local normal-based diffusion (Naughty Dog "Character Lighting and Shading") to get some nice scattering effects as well as some of my own little tweaks (a closer rgb relationship in blended normals to keep red bleeding lower for a pleasing look).
The texture maps for each layer (subdermal and epidermal) can be done by hand, or they can be generated by the shader (merely a few simple color operations on the diffuse texture, and sampling a mip-map for the subdermal texture for a blurry effect). The overall tint can be changed through the shader's settings. The approximated textures generated by the shader are good enough for regular use, but pre-made textures have the benefit of having more control and artistic freedom. The model used in the example images uses a generated epidermal map and a pre-made subdermal texture (for back scattering, see below).
Texture Maps (yes, my translucency texture sucks :P):
For backscattering (yes, the shader supports back scattering, although it's primitive and will probably need changing eventually) I simply light the model using a wrapped negative NdotL value (opposite of the light) and mask it using a translucency texture (placed in the alpha channel of the subdermal texture). It's not physically correct but it does the job just fine. I'd love to get translucent shadow maps working but limitations are stopping that from happening and I haven't found any other decent alternatives yet (if someone knows of any or has any ideas, please let me know! :D).
Back scattering (mind the jittery shadows):
For each layer I calculate a different lighting result. For the diffuse texture, I simply calculate blended normals as usual, using a standard NdotL. For the epidermal layer, I do the same but with a slightly more wrapped NdotL, and for the subdermal I use an even more wrapped NdotL. This is to simulate the layers being "blurred" so that when weighed together at the end of the shader, they create a proper subsurface effect. The weights for each layer can be controlled in the shaders properties.
To properly scatter shadows I do something very similar to blended normals. For each layer, after blending the normals, I blend between a hard shadow and a soft shadow, allowing a red-bleed to occur on the shadows. The strength of sharpening and softening for each channel is changed for each layer (diffuse is very hard, epidermal is softer and subdermal is very soft). Rather than performing blurs on the shadow maps which can result in poorer performance, I cheat... big time. I simply run pow() for each channel (green & blue are merged in both blended normals and blended shadows), giving a harder shadow. This basically hardens and shrinks the radius of the green/blue channels while the red channel remains nice and smooth. The result is desaturated a bit so it doesn't come off as too red. Since I'm doing this in CryEngine2 I decided to pre-blur the shadow maps since jittering can get ugly after hardening the shadows (it's merely a screen texture, so blurring it only needs to be done once: I'm doing a mere 4 tap blur (4 texture samples surrounding the original). I've had some good results using this method but it still needs a bit of work.
Bleeding shadows (100% diffuse vs 100% epidermal/subdermal (no diffuse))
The final effects are detail bump mapping (basically a tiled detail normal map layered over the original to give the impression of more detail), rim lighting and the option to change the skin's melanin amount. These are pretty standard effects really (the melanin shader is just lerping between 1 and the diffuse texture multiplied by it's luminance using a melanin value).
Here's a comparison between standard lambertian lighting with a physically based blinn-phong specular (Naty Hoffman), the skin shader without SSS and then with SSS.
Performance is great, it's ~the same as a standard fully-featured material shader, you probably wouldn't notice much impact on performance if you swapped it for say, a metal shader.
Compare that to the performance it takes to render texture-space diffusion which is orders of magnitude slower than this, I'd say I've hit a good mark. It's not as physically accurate as NVIDIA's Head Demo, but looking at the results and the simplicity of the shader, I think I'll be able to close the gap eventually. I hope to push forward and obtain something more correct, but I think this is a good step forward.
To anyone who read: thanks for reading. ;) Here's a last minute screen-capture.
[Edited by - Styves on October 18, 2010 2:22:46 AM]
.
