Rendering tips please

[Mike_RB]

I'm evaluating mantra for rendering an upcoming project. The subject matter is something like a silver car bashing through stuff. My background comes from lightwave, mental Ray, and modo.

So lots of blurry reflections, subdivision surfaces and displacement.

My initial tests seem to indicate the surface model plus pbr plus env light is the way to go. However I have some specific questions that seem a little vague in the docs

For an rtpbr render I was using 8x8 pixel samples with 1min and 64max shading samples using variance 0.015 noise and gamma color space. I left all other quality values at 1 and this seemed to give a pretty good result of 64 aa/mblur with a range of 64-4096 shading samples per pixel. I'm not scared of 4096 being used here and there to clean up the noise as blurry reflection and environment lights need to be sampled heavily in the cracks.

So all good there. However I'm getting errors on the farm with ifds where the render will just stop at 24% or 35% and quit. No real error. I get an exr back that has 1/4 filled in with the render from the top (not buckets). Checked ram usage. Not a problem.

So then I started to investigate the other rendering modes as our resident houdini genius (docent know rendering) informed me that mp is always more stable.

Worth investigating. So I put a low quality mppbr render on the farm and it was fine.

Then I tried to dial in the quality. As far as I can tell Ray variance dosent work at all with mp? So then I started nudging up the light quality, Min shading samples and shading quality dicing. None of them worked well on their own. But a little of all 3 mostly did the trick. It seems like there is a lot more fiddling with mppbr instead of just letting the variance take care of it. What is the best strategy here? 8x8 pixel sample. 1 shading sample 1 light quality with with shading quality dicing at 20 or 30 to get 400-900 micro polygons in each pixel? Wont I hit a memory problem like that?

Thanks for reading this and I look forward to hearing how you guys are dialing in the quality you need without the Ray variance.

Thanks

Mike

[LaidlawFX]

Saturday when I have the time I can probably detail some more info specifically.

Here is some basic stuff to help you out, it's not finished, especially in the area you need it, but it is a better organized version of the help documents, lol, least from my perspective

Rendering Documentation

[Mike_RB]

Thanks for the PDF

After doing some more tests I've found that variance Ray sampling does work with mppbr. So my latest efforts are 7x7 pixel samples. 4-16 shading samples. Shading Quality dicing at 8, light quality at 1

This seems to do a pretty good job.

Another question: using the direct Ray extra image plane I see that the blocks of shading with quality dicing at pretty big. Not "micro" by any means. Quality 1 is way bigger than a single pixel. Maybe I need to change dicing to raster space? Otherwise what does the 1 represent?

Thanks

[LaidlawFX]

Thanks for the PDF

After doing some more tests I've found that variance Ray sampling does work with mppbr. So my latest efforts are 7x7 pixel samples. 4-16 shading samples. Shading Quality dicing at 8, light quality at 1

This seems to do a pretty good job.

Another question: using the direct Ray extra image plane I see that the blocks of shading with quality dicing at pretty big. Not "micro" by any means. Quality 1 is way bigger than a single pixel. Maybe I need to change dicing to raster space? Otherwise what does the 1 represent?

Thanks

I tried looking at the direct ray image plane and I wasn't getting anything at all. I don't know what the direct correlation between dicing and that would be. Do you have a doc I could look at?

The dicing tabs on the mantra surface are multipliers for the dicing, so what ever 1 is multiplied or divided by in the code. So not a 100% 1 dicing to 1 pixel comparison, I remember somewhere it being described as that, or better yet, it is good to explain it like that in terms of discussion, in order to understand the concept better. I feel like it is though, but I can't source it at the moment. It is possible that your other render setting are effecting the size of what you see in the direct ray image plane. It is not always the most efficient way of doing it, but it certainly works as it runs more shading calculations by making the chunks smaller, you just end up doing more geometry and shader processes as opposed to one or the other. There are also dicing tabs on the geometry to where they originated prior to h11, that could be affecting your speed, and quality.

For the dicing here is some borrowed text, to explain it. It would be best to say I get the feel for it, and know what to play with, but I do not 100% understand it at all, since I didn't create it myself. So it easier to give you the raw info.

Micropolygon measuring is the three measuring devices to determine if primitives are “too big” and need to be diced into micropolygons. Uniform measuring generates uniform divisions. The size of the divisions are controlled by the shading quality on the mantra rop, this is a global multiplier. Locally there is the shading quality parameter at the object level on geometry nodes under the render, dicing tab. Raster space measuring measures geometry in screen space. This is roughly equivalent to the “nonraster -z 0” measurer, so it has been deprecated in favor of the following approach. Non-raster measuring measures geometry in 3D. The z-importance is a bias for the z-component of the surface. A z-importance of 0 means that the x and y components of the object will be the only metric in determining the size of the object. This is roughly equivalent to raster space measurement. By increasing the z-importance to 1, the z measurement becomes more meaningful. It is possible to increase the z-importance beyond 1.

If you think of a flat primitive in the XY plane, the z-importance has no effect. However, if a primitive is nearly in the XZ plane, z-importance has more influence on the dicing. With a z-importance of 0, only the projected measurements will be used, which will result in long, thin strips being created. With a z-importance of 1, the grid will be more uniformly sub-divided. With a value greater than 1, more divisions will be performed in Z. This is important when displacement mapping is being performed.

[Mike_RB]

As every pixel has at least 1 direct shading sample you need to click the button to frame the values so you can see the result. The float value is the number of direct samples. Pretty useful image plane.

[LaidlawFX]

As every pixel has at least 1 direct shading sample you need to click the button to frame the values so you can see the result. The float value is the number of direct samples. Pretty useful image plane.

Cool