PBR, IBL, (very?) Technical Question. Sphere vs Cube.

oXYnary

So Torque 3d MIT is aiming for full PBR implementation soon.  I'm on one of the test branches currently supporting it.  It will be using the metallic/roughness workflow.

Anyhow, with the reflection map system that is being used relies on Torque's old system of giving the engine 6 different maps that it will combine into 1 cube map vs giving it a complete cubemap or use a panorama/spherical system. From what I understand, for higher roughness values, the reflection is just shown at a lower bit mipmap setting. 

The issue is at very low roughness I can start to see the seams of the cubemap in the reflection.  I was using 512 for every side, but we had to crank it up to 2048 for each side so the mipmap was at a high enough resolution still to not show the seams.  I think what is happening is that since the engine put each side together separately versus as a whole as with a spherical or standard combined cubemap is that at the edges of each seam its acting like having no padding with UVs.  As in the lower resolutions are only blending that edge to itself versus across to the other sides.

Does this sound correct?  What is the most common Environmental Reflection Shape and format used for PBR?

EarthQuake

I think how we do this is that we process each mip to be properly blurred, we don't just use the mips as is. Also, for the lowest mips we don't use super low resolution images, they have a fixed minimum level and blur the blurriest ones more, rather than just using lower mips as this doesn't look great. From my understanding, we save the blurrier versions with less resolution, but we don't use the lower resolution mips for blurriness per say. I think max cubemap res we use is 512.

Also think the various blurriness levels are specially processed to avoid seams.

I don't know the exact file format we save the IBL content in, but I think it's some sort of fairly standard DDS cubemap format.

Don't quote me on any of this because I'm not sure how correct it is.

Eric Chadwick

This might help
http://wiki.polycount.com/wiki/Diffusely_convolved_cube_map

oXYnary

Auh, so your link Eric implies that you manually create each MipMap level from a combined version before separating for the engine?

Eric Chadwick

No, you use a tool that does the mip filtering properly, creating the blurry mips by sampling all six images together at the same time. Cubemapgen does this, Lys does this, probably a few other tools do this. Most tools do not, because they don't have any idea about what a cube map is.

JedTheKrampus

Have the engine use cmft to convolve the cubemaps:

https://github.com/dariomanesku/cmft

Also, check out ARB_seamless_cubemap for OpenGL to solve the seams problem. Depending on the engine min spec it might not be available in which case you'll need to slightly alter the texcoords of the cubemap lookup.

https://www.opengl.org/registry/specs/ARB/seamless_cube_map.txt