Displacement Precision Issues 2018.6 / Mtoa 3.2.X and 3.3.0

[lennart.oberscheidt]

Hello guys,

I already posted a longer text about some arnold displacement precision issues that we noticed during a current project, however, for whatever reason the post seems to be hanging as it has to be moderated. Anyway, after more then a day of troubleshooting we've come to the conclusion that there are currently multiply seemingly serious precision issues in arnold that i never noticed before, or in any other render engine in the severity we are experiencing them currently. Now, we all know about the limitation of floats, however, especially displacement in arnold seems to be very instable even when only working a few hundred units off the origin. An asteroid 4000 units translated in y not only renders completely different at 0,0,0 and 0,4000,0, it actually displays very differently in the IPR at -3 and -2 samples before the first buckets start to render or hit the object (no view dependant tesselation). Example images included. Unfortunately, i cannot share the scene file currently (and the example shown is a reduced example) but the issue remains and can be reproduced in different scenes with different objects, even simple cubes. Anyone noticed similar issues before? Displacement actually starts to look differently when only moving the objects +- 50 units. I'm pretty sure we rendered displaced objects several hundred units off the origin without problems many times, so I'm currently pulling my hair out figuring out whats going wrong in (seemingly) all recent arnold versions and even simple examples. We also noticed a lot of jagged shadows / edges on objects close to the camera that are about 4000 units off the origin, but the issue becomes apparent as soon as we move them away a few hundred units. Now, we can manage that and frankly we are geeting very close to those objects (VR/360°) projects and currently work with large scenes for the project, however, I'm pretty confused: I used to move whole scenes backwards towards the origin while the camera flew through them during long takes before to circumvent precision loss before, but never have I noticed artifacts that obvious and never in completely normal unit ranges of a few hundred to thousands.

Any known problems with displacement currently? Any pointers what the problem might be?

Displacement Issues:

IPR before first bucket at 0,4000,0

IPR before first bucket, object at origin (correct)

Unusually apparent general precision issues:

[thiago.ize]

How big are your triangles? By how many units are you displacing?

4000 units away from the origin, the smallest increment a 32-bit float can record is about 0.0004 units. Are your displacements perhaps moving vertices in such a way that a 0.0004 error would be noticeable?

Is this a problem in the baked in geometry or in the bump? If you disable autobump on the polymesh and any other bump nodes you might have, do you get similar looking renders at the origin and far from the origin?

Is there an example of the -AA differences? I couldn't tell from the images you supplied.

[ashley.handscomb.retallack]

What happens if you disable autobump?

Check for it in the arnold section of the displacement shader and the displacement options in the mesh node.

https://docs.arnoldrenderer.com/x/xwtkAg#Displacement-Autobump

https://docs.arnoldrenderer.com/x/Zg1kAg

[thiago.ize]

As I just mentioned in your other question, I think the shadowing errors might be fixed in the just released 6.0.1. The smoother looking autobump artifact for the initial low-res blocky pass we haven't gotten to yet.

[lennart.oberscheidt]

Hi Thiago,

Thanks for taking a look into this and to hear back from you – thank you very much! I already noticed this in the Arnold 6 release notes and was planning on installing it tomorrow. I will check out the issue with an older scene file and test setup in the coming days and report back to you! In the end, we rendered everything closer to the origin, which worked:

Cheers,

Lennart