Use amount and perceptual roughness constants from Filament.

[fire]

Reviewing the constants from Filament and changed to the roughness amount of 1.0 and the roughness limit to be 0.002025. https://github.com/google/filament/blob/main/shaders/src/shading_lit.fs

https://github.com/google/filament/blob/9914eb84a4d613d609ae77a1d031fadf973db02c/shaders/src/common_material.fs#L6

Need some help doing screenshot comparison @Calinou .

Updated:

1. Update sample to latest Godot master Calinou/test_specular_aliasing#1
2. PR Use amount and perceptual roughness constants from Filament. #56822

Edited:

Uses the 0.002025 constant now.

[Calinou]

Need some help doing screenshot comparison @Calinou .

I made a slider comparison with Calinou/test_specular_aliasing@61836cb: https://imgsli.com/OTEwNzM/1/2
(Note: I pushed a new commit since that uses lossless texture compression to work around the VRAM mipmap issue.)

This PR seems to get ever so slightly closer to ground truth, so I think the new values are a good idea.

[fire]

#define MIN_ROUGHNESS 0.002025 on rendering/anti_aliasing/screen_space_roughness_limiter/limit

Removes a lot of the aliasing artifacts but I can't tell if it's correct.

[fire]

@Calinou can you post an imgsli of the new 0.002025 constant?

[fire]

For another pr, it may be less confusing to remove https://github.com/godotengine/godot/blob/master/servers/rendering/renderer_rd/shaders/roughness_limiter.glsl.

[Calinou]

@Calinou can you post an imgsli of the new 0.002025 constant?

Done: https://imgsli.com/OTEwODc/1/2

It seems the new roughness limiter values are barely different from having the roughness limiter disabled in the first place.

[fire]

For someone who was wondering what barely different from having the roughness limiter disabled in the first place means here's a video.

2022-01-15.15-01-05.mp4

[Calinou]

With the new changes, I can still barely see a difference between the current roughness limiter and the new one: https://imgsli.com/OTEyOTE

The images are different, as demonstrated by dssim:

[fire]

I was told that:

The function mainly affects specular aliasing on geometry edges

So I was recommended to test something metallic but mostly geometry instead of shader detail, like not a geometric primitive plane.

This matches the intuition that the current system uses surface normals.

Edited:

Something like a chair or table, or if you just want something obvious a bunch of metal bars and spikes.

[fire]

The paper authors use bistro with all objects turned metallic.

From clayjohn.

[fire]

After a conversation with @clayjohn, he said that we should test to demonstrate improvement.

After the great work of wickedinsignal, we now have a great test map at #63374.

[fire]

I wasn't able to find time to test this again. Probably needs to be salvaged.

[fire]

Godot's settings now almost match the pdf's with a difference of like 0.03