2021.3.2 Is Multiple importance sampling broken?

[RevC2]

Hi,
I’m just testing 2021.3.2 to see how much the version is stable and I’ve noticed that doing a bake with “Multiple importance sampling” activated gives back odd results.

Image 01 Bake without MIS activated, that’s more or less what I expect from the bake

Image 02 Bake WITH MIS, informantion in lightmaps looks odd, with weird “brush stroke” look

Am I missing something or could someone else confirm this?

[AcidArrow]

The results make sense based on my understanding of MIS, yes.

You need more samples btw if you need cleaner results, not MIS.

Read this for an explanation of MIS (and other stuff).

Confusion about new Lightmap parameters

The environment used to be sampled as part of the indirect samples. Everytime an indirect ray was shot from a lightmap texel, or bounced off a surface, and didn’t hit anything, the environment was evaluated. The problem with this approach is that high frequency environment maps would lead to a lot of noise due to undersampling. With the latest changes the first difference is that environment sampling has been split into direct and indirect environment sampling. Direct environment sampling is co…

[RevC2]

Hey AcidArrow,
thanks for the reply. Yeah I know, the lack of samples and denoising it’s just to show a “raw” lightmap, and what looks like to be MIS wrong results. I’m not really confused to why it’s noisy beacuse of the low intentional settings

BTW the artifacts remains with more samples/denoise active

This is with higher samples + denoising, still the weird artefacts are present.

[AcidArrow]

I guess you didn’t read the post I linked, I’ll quote the relevant part here

When enabling MIS, we analyze the environment map and find spots of high influence in it. When generating the ray directions, we alternate between either picking a random direction like before, or picking a direction from the pre-calculated points.
MIS works best for high frequency environments, but will produce noisier results, the smoother the environment gets. The worst case is an environment of uniform color.

[RevC2]

@AcidArrow

I guess you didn’t read the post I linked, I’ll quote the relevant part here

Well, I did read it and it says nothing new as far as what MIS actually is. The point is that this weird result never occured before in dozens of other Unity versions/projects I’ve worked on so I’m trying to figure out why the usual “room + sun + simple procedural sky” gave me such bad lightmaps. If you take a look again at “Image02” you’ll see an unusable lightmap, not because of the NOISE, but because of the strange gradients, “strokes” and splotches which never ever occurred on a simple setup like this. I did a quick test in 2020.3.33 with same geometry and same lighting setup (Sun + simple procedural sky) and the result it’s correct and what to be expected:

I’ll try another couple of tests in the meanwhile, I’m curious to know I maybe my project has issues

[kristijonas_unity]

MIS works well with high contrast HDRIs. If the HDRI you are using has mostly homogeneous luminosity values, enabling MIS will result in artifacts and/or noise.

We have a feature in the backlog which will enable MIS contextually behind the scene. For now, just disable MIS if it results in more artifacts.

[RevC2]

@kristijonas_unity

MIS works well with high contrast HDRIs. If the HDRI you are using has mostly homogeneous luminosity values, enabling MIS will result in artifacts and/or noise.

We have a feature in the backlog which will enable MIS contextually behind the scene. For now, just disable MIS if it results in more artifacts.

Ok, i guess. I’d still like to understand why it has worked fine with pbr skies for two years and hasn’t in this example (again, same pbr sky I always used)

Thanks anyway

[kristijonas_unity]

If you have clear evidence that the output using MIS has changed recently, while using identical lighting settings/conditions, then please let us know by reporting a bug. It could possibly be a regression