Approximate indirect specular occlusion

[aevyrie]

Objective

• The current PBR renderer over-brightens indirect specular reflections, which tends to cause objects to appear to glow, because specular occlusion is not accounted for.

Solution

• Attenuate indirect specular term with an approximation for specular occlusion, using [Lagarde et al., 2014] (pg. 76).

Before	After	Animation

---

Changelog

• Ambient occlusion now applies to indirect specular reflections to approximate how objects occlude specular light.

Migration Guide

• Renamed PbrInput::occlusion to diffuse_occlusion, and added specular_occlusion.

[IceSentry]

I really like how it looks especially considering how small the change is but since it breaks the Physically Based part we should probably make this configurable.

The suggestion in discord to do indirect_light += environment_light.diffuse * occlusion + environment_light.specular * specular_occlusion would make that pretty easy.

The comment should maybe be shortened a bit, maybe don't mention the wheel well example? It seems a bit too specific and over explained to me.

[aevyrie]

it breaks the Physically Based part

I think this is debatable. Not accounting for specular occlusion is also not physical, but it requires raytracing to simulate accurately. This gets us closer to the ground truth by having something for the specular occlusion term. The next most accurate would be using SSR for the occlusion check, followed by full path tracing.

[nicopap]

I do agree that it looks better when occluding the specular as well. The Fresnel defeats pretty much the purpose of the occlusion.

I followed closely the discussion on discord I think this is worth merging.

I've yet to see a scene where it looks worse, so I think it should be merged, since it's not even a tradeoff, it's just straight up a positive change.

Of note is that occlusion is not strictly defined by SSAO, but also by the occlusion texture. But looking at the glTF2 spec I don't see any details about whether this should affect environment light specular reflection or not, so it's fine as-is.

Regarding the suggestion of having it parametrable. I think it should be part of a different PR. It's a different level of controversy, as it includes more tradeoffs.

It seems @DGriffin91 has a more accurate implementation of specular occlusion, but until its ready, I think it's fine to merge this.

[DGriffin91]

Specular ambient occlusion is included in both reference implementation that bevy uses for GTAO (XeGTAO), and is also specified in original paper: [Jimenez et al., 2016] (pg. 3)

Specular occlusion from [Lagarde et al., 2014] (pg. 76) is also specified in Filament, the primary reference PBR implementation that bevy uses: https://google.github.io/filament/Filament.md.html#lighting/occlusion/specularocclusion

(XeGTAO implements both the simpler Lagarde14 method as well as the improved Jimenez16 one)

Using the diffuse ambient occlusion term directly for specular occlusion will result in over occlusion in areas that would receive less ambient diffuse light but still have a path for specular light, while also significantly under occluding in most cases.

We should continue implementing GTAO to include GTSO. And then also provide a SSR oriented method for more accurate specular occlusion that also better captures more distant occlusion.

[aevyrie]

The filament docs and related paper you link are equivalent to what I've done here, for rough materials. For rough materials, AO and specular occlusion are 1:1. I would just need to add the wrapping function to match Filament/Frostbite for roughness values < 0.3.

It also looks like XeGTAO also uses this (Legarde) method in the simple case.

https://github.com/GameTechDev/XeGTAO/blob/0b276c0ce820475c2adf6e2f3b696b696c172f43/Source/Rendering/Shaders/Filament/ambient_occlusion.va.fs#L78-L79

[JMS55]

Needs a migration guide and changelog for the diffuse_occlusion naming change.

Approving this PR, assuming that ignoring baked occlusion makes sense.

[aevyrie]

Note, I implemented a min(baked_ao, ssao) for specular, but this had no effect because SSAO was always darker. We decided to just ignore the baked AO for specular occlusion.