Is there a reason NOT to use a metalness workflow in PBR?
polycounter lvl 11
The purpose of metallness vs. spec/gloss has been made clear. Only use it when working on metallic objects. However if your object is a mix of metal and dielectric then flood the dialectic portion of them black. The dielectric always comes out just fine in a metalness map. I figured that if I have an all dielectric material then I'll just make a roughness/all black metalness map and call it a day rather than spend the time/effort making a specular as well.
Sounds like metalness is objectively better, yet people use spec/gloss all the time. Am I missing something?
Sounds like metalness is objectively better, yet people use spec/gloss all the time. Am I missing something?
Replies
Two things that you should always tweak the specular of in a metalness workflow are human skin (which should be set to 0.4125 in UE4 spec) and water (which should be set to 0.25 in UE4 spec.) Otherwise, you shouldn't need to worry about it much. Also, if you are going to make a flooded black metalness it's usually considered better to set it with a value in the UE4 material editor rather than with a texture, but if you're already channel packing it shouldn't be a big deal.
Specular in UE4 is a limited range of refractive indices (represented as percentages) to work with which makes it impossible to use as if it were a typical specular workflow. It is there for if you want very specific control over its limited range of 0% to 8% specular (1-~1.79 IoR). Its default of 0.5 is a 1.5 IoR, working out to a refractive coefficient of 0.04 (look up schlick's approximation and the Ro equation for why). In other terms, 4% of light is specularly reflected. Good enough for everything but metals. So if you care about having that fine of control then use the specular input knowing how those values are getting interpreted and in a physical sense. But for 99% of the stuff out there the default value is very close. The metallic workflow takes advantage of the fact that metal materials have no real albedo, and blends between the specular value and using the albedo as a refractive coefficient for each RGB channel. The downside is the trouble of softly blending between dielectric and metallic on the same material.
Oh and I do R - AO G - Roughness B - Metallic for my packed textures. With appropriate AO and well authored albedo/roughness/metallic I have never needed anything else.
I am one who uses specular input all the time to erase excessive highlight from cracks , surface pores , small details shadows baked into surface etc. And then have to compensate decreasing overall reflectivity somehow fighting with PBR stuff. Especially when the normal map has not enough resolution to model surface bumps accurately enough. i.e. a crack is 1pix wide and have no room in the normal map actually.
Now this means that for normal map information we have to do a slight modification to maintain what is perceptually correct roughness when objects are at different distances. When we move further away from an object our sample size (area of each surface within one pixel) increases.
The textures begin to mip. And in this process you can imagine that information is lost due to that resolution decrease. This generally isn't an issue for albedo textures and the like, but it's a massive problem for normals. Because light is being bent in various directions for each pixel of the normal map texture, this has a similar effect to how microfacets cause reflections to appear rough. WIth very detail normal maps an artist can cause a surface to look nice and rough up close, but as the camera zooms out in engine and all those perturbed normals were lost, the object appears extremely glossy! To counteract this we are applying a precalculation to transform normals into a microfacet roughness representation and just add that increasing roughness to each of the object's roughness mip levels. Valve has a great presentation on this problem, also presenting a hacky solution for geometric normals being converted to roughness.
PDF Slideshow -
http://media.steampowered.com/apps/valve/2015/Alex_Vlachos_Advanced_VR_Rendering_GDC2015.pdf
Video Slideshow -
http://www.gdcvault.com/play/1021771/Advanced-VR
In UE4 it's accomplished for normal maps through Composite Texture:
https://docs.unrealengine.com/latest/INT/Engine/Content/Types/Textures/Composite/index.html
A fun anecdote, I learned about this as a kid due to visiting the Biltmore House (iirc one of the top 2 biggest houses in the US, it's a tourist attraction now) and they told us that the builders hand chipped all the bricks specifically to create a uniform rough look for the building at a distance to prevent any glare. Who would have figured that'd be a useful bit of information later on!
Edit: Oh yeah. As a reversion of this thought process, your problem of having a 1px crack and not being able to represent it with normals is easy to solve. This is where the highest resolution of your normal map has reached its limits, so you represent it via increased roughness
Unfortunately even being 100% rough those tiny cracks with no representation in the normal map still reflect essential amount of highlight , making kind of "laminate cover" feeling and dissolving right in the next mip. Plus having too strong contrast between polished and rough pixels creates same kind of artifact (halos) as in between metals and nonmetals . In fact in practice I really can't use contrast black and white representation in metallic channel and in roughness one too. It's always a kind of compromise to reach better overall crispness.
That makes the whole metallic approach a kind of tricky and more time consuming than the old speclevel/gloss one imo. So I totally understand why people still use it
So this is an interesting problem that I aimed to briefly talk about with the suggestion to use the ambient occlusion slot. The reality is that normal maps are by nature not actually energy conserving. Where a high poly model may shadow or mask light, the normal map has no information for capturing this. However by darkening the specular you are occluding direct light that could just as easily be reflected from within a crevice as well as darkening indirect light which kinda makes sense, but without any sense of the material occlusion. I try to model with these things in mind, using geometry to aid this type of energy conservation. One solution would be to capture a correlated position-normal dataset that you could enforce energy conservation across, but this is beyond what's currently feasible and in the future tessellation will solve this problem anyways.
Could you share the halo artifacts you're talking about between rough and polished pixels? I haven't had this problem. As far as using an IoR based specular it is a better method for controlling reflections, but a last gen spec/gloss isn't gonna be less time consuming to work with than spec/metallic PBR setups. I pump out textures like twice as fast now haha. And not to say that this is what you're doing, but the modern pseudo PBR systems are a huge improvement in every regard. If people avoid embracing modern rendering systems because they can't get good results then they are hurting themselves.
base color -black, metallic -black, roughness -is dark rectangles on white background. It's not that strong as metal halos but still makes noisy/grainy surfaces a bit less crisp comparing with old spec level shaders where highlight intensity was expressed directly by pixel values.
btw, picture also shows how much highlight intensity 100% black and 100% rough surface still reflects while in reality it would be most often filled with micro shadowing from the small bumps a normal map often can't represent correctly .
The shadowing-masking parameter of the BRDF is the main thing controlling energy conservation. Roughness is a principle component of it. And because the normal map feeds into roughness as the mipmaps blend, this method makes a better approximation to the energy conservation that would happen IRL. The other thing that should get modeled is occlusion, which yeah the normal map can't represent. But by prebaking that AO information and applying it to the indirect lighting function you are dealing with the shadowing that would happen at such small levels in a more physically appropriate way. You also allow direct lighting to work as it's supposed to by not arbitrarily adjusting a refractive index value.
Realtime has a lot of shortcomings at the moment, particularly multiple scattering which is the current main source of energy loss. But if composite textures and occlusion are leveraged appropriately I've found no need for that kind of tweaking.