2021-12-7 Blender Rendering Meeting

I presume you are referring to MNEE. It is a specialised technique for rendering refractive caustics so I would assume not. However, in principal the technique could be extended to other surface types.

2 Likes

Exciting to see Metal backend is close to release state!

1 Like

Is there patch/task page for this?

3 Likes

Yes please, I would love to follow/test the hydra delegate, this is amazing news,

1 Like

How soon can we expect Metal backend in Alpha?

I am sorry I don’t have a definite date for you but it’s probably in the coming weeks.

1 Like

Sorry I only have the link above and no task or patch.

1 Like

Oh, that’s totally ok!

Regarding USD / Hydra, I’ve updated the notes to clarify this. There is no actual patch yet, there is interest from NVIDIA in contributing this and when it becomes concrete we’ll make a task.

7 Likes

Just to make sure I get it right: are the upcoming improved caustics a result of overall improved light transport in Cycles, like bidirectional path tracing / VCM, or is MNEE a trick to only improve caustics?

You can find the MNEE patch here: ⚙ D13533 Adding Manifold Next Event Estimation Sampling Technique

From the looks of it, it seems MNEE in that patch a trick to improve caustics, not an “overall improvement to Cycles light transport”.

There was a task created for path guiding (⚓ T92571 Cycles: path guiding) that should bring improvements to the convergence of reflective and refractive caustics. This is more likely to be a “overall improved light transport in Cycles”. However, we need to wait for a patch for this to be made.

3 Likes

Thanks for the info and links. So the MNEE implementation is not a replacement of the path guiding initiative, but a method to improve caustics, while improved overall light transport will still be implemented?

MNEE is an implementation that specifically helps with certain refractive caustics (at least in it’s current form). Reflective caustics are not covered.

From my understanding other improvements to Cycles light transport (like path guiding) is still planned for the future. However this is just my interpretation of things. You’d want comments form the Cycles developers for exact plans.

1 Like

Instead of bi-directional path tracing, Metropolis Light Transport directly, a god renderer is born

1 Like

These algorithms have their own issues. Metropolis Light Transport has incorrect brightness issues (shown in the Practical Path Guiding paper’s kitchen scene comparison), MNEE has issues with more complex shapes (In the MNEE paper there is a “failed case”), Path Guiding according to SNEE’s paper seems to also have its own issue that I don’t quite understand.

The more I read the more I realize that there is just no one perfect algorithm. But hopefully some algorithms are complementary like MNEE and Path Guiding, let’s wait and see how it goes.

2 Likes

So am I in the right thread to ask about if EEVEE will get baking capabilities in the future?

What is the status / progress of AMD GPU Linux support in 3.1 ??

1 Like

Metropolis Light Transport has incorrect brightness issues, that sounds like an easy problem to solve with AI, it could be trained on what correct brightness images look like, then simply fix the image post render just like AI de-noise, all it would need to do is adjust the exposure to correct the image before it gets converted down to a PNG, JPG etc. Or am I wrong?

No I didn’t mean exposure, they are different.
Here is the part of the path guiding paper:

Also, render engines like Cycles render in open domain, or scene linear state, which means the intensity info it generates is radiometric and the highest end of it’s range is infinity.It doesn’t make sense to have it adjusted to correct result only when you convert it to close domain PNG etc, what about EXRs then? It just doesn’t make sense.

EDIT: Actually now I think of it, I might be confused about MLT and MEMLT, I now remember I have read somewhere that MLT has temporal flickering issue, not sure though.
EDIT: confirmed my previous edit:

Anyway I think my previous conlcusion still stands

I’m not surprised that the problem is more complicated than I thought but I still hope AI could be trained to correct it. Out of interest when using Luxrender MLT in it has a slider for probability of large mutation which if i recall when it it will decide to give up and work on another area of the image, if set very low the MLT algorithm will strictly keep working on any bright areas of the image like caustics for many samples, or if set high it will work on the bright areas but likely diverge and look for other areas to work on after a few samples… And yes it stated that MLT works by identifying bright areas of the image to work on as those are the areas of interest for light transport. Sorry if this is confusing im just going from memory.