2013 - Unified Sampling - Great Results

[JJJenkins]

I've been testing the improvements made to Unified from the version of mental ray included in 2012 to 2013. I have gotten some serious gains with my initial tests and am extremely pleased with the results. I literally did absolutely nothing different between 2012 and 2013. I set the scene up in 2012, rendered, then opened it up in 2013 and hit render. That's it. Based on the info on the awesome elemental ray blog, I could probably lower my quality in 2013, get the same visual result, but lower the render times even more. I'm going to experiment with the error cutoff and see if that improves render time as well. But first I wanted to do a direct, cut and dry comparison of Unified between the 2 versions of mental ray.

The difference in the samples buffer is amazing. The reduction of eye rays from one version to the next is significant. Actually, I got a significant reduction in all of the rays shot, not just the eye rays.

For this scene I kept it simple and used one mia_material_x with a poly carbonate look. 1 glossy sample and 4 samples for AO. The Builtin IBL was my only light source. I used final gather with super low settings (a little too low actually).

Unified Settings:
Min - 1
Max - 500
Quality - 4
Error Cutoff - 0

Render time went from 29.22 to 17.51!

Check out the output window:

2012:

RC 0.18 info : rendering statistics
RC 0.18 info : type number per eye ray
RC 0.18 info : eye rays 43099593 1.00
RC 0.18 info : reflection rays 56174916 1.30
RC 0.18 info : shadow rays 366797700 8.51
RC 0.18 info : environment rays 24182792 0.56
RC 0.18 info : probe rays 177916955 4.13
RC 0.18 info : fg points interpolated 68795129 1.60
RC 0.18 info : on average 32.49 finalgather points used per interpolation

RC 0.18 progr: rendering finished
RC 0.18 info : wallclock 0:29:22.37 for rendering
RC 0.18 info : allocated 651 MB, max resident 953 MB
GAPM 0.18 info : triangle count (including retessellation) : 5713899
PHEN 0.18 info : Reflection rays skipped by threshold: 8744810

2013:

RC 0.18 922 MB info : rendering statistics
RC 0.18 922 MB info : type number per eye ray
RC 0.18 922 MB info : eye rays 29450633 1.00
RC 0.18 922 MB info : reflection rays 37633341 1.28
RC 0.18 922 MB info : shadow rays 253317933 8.60
RC 0.18 922 MB info : environment rays 19338079 0.66
RC 0.18 922 MB info : probe rays 121407338 4.12
RC 0.18 922 MB info : fg points interpolated 44114611 1.50
RC 0.18 922 MB info : on average 33.03 finalgather points used per interpolation

RC 0.18 679 MB progr: rendering finished
RC 0.18 679 MB info : wallclock 0:17:51.18 for rendering
RC 0.18 679 MB info : current mem usage 679 MB, max mem usage 968 MB
GAPM 0.18 679 MB info : triangle count (including retessellation) : 5713899
PHEN 0.18 651 MB info : Reflection rays skipped by threshold: 4542425

I don't know how to properly tonemap the samples_buffer in imf_disp so I just did a screen grab with the exposure set to -7.0 on both images. You can see there is a huge difference between the number of samples between the two. Really cool stuff!

The memory usage seems to have increased slightly, which I don't understand. Any ideas why that may be? But anyway, I'm extremely happy with these results. I'm definitely going to be using Unified on my next project.

-Justin

[Remydrh]

Yep, in fact you should probably reduce Quality on a few scenes for extra gains. The change between versions is pretty extreme. Depth of field and motion are also smoother (hence faster) than before.

The memory usage is a bit odd, we've generally found out memory usage decreased.

Used in conjunction with the new user_ibl_env and you should get some pretty nice/fast renders. This also improved Progressive. Hoping to see this working with IPR after a hotfix or two.

[JJJenkins]

I just added an error cutoff of .03 and reduced Quality down to 2.5. I got the render time down to 13:02.

This is what imf_diff had to say:

differing pixels: 1.999% (18426 of 921600)
average difference: 1.257%
maximum difference: 11.694%
Summary: Some pixels differ.

I could really only tell a very small difference with my eyes. A couple highlights changed slightly, but not by much. The rest of the image looked the same to me. With animation and some motion blur I think the time savings might be worth it to go with 2.5 for quality. Going from 30 min all the way down to 13 min is just awesome.

Getting progressive to finally work with IPR without destruction would be a most welcome change. I really hope they implement that in a hotfix in the near future.

-Justin

[JJJenkins]

I posted in cgtalk that I noticed more noise with user_ibl_env than Native IBL. Doing some more tests now. I thought that may be due to the whole texture re-baking thing. So basically the texture re-baking was helping to smooth out the result by decreasing detail with Native IBL. Any thoughts on that?

[Remydrh]

That's basically correct. Glossy reflection would benefit from less detail in the Native IBL if baked to a lower resolution.

But we found overall with Unified we can use very small samples settings with good results.

[rlevene]

Hey Justin,

Thanks for posting your tests. Really good to see.

We have finally come off a big job, some parts were rendered with unified so next week will get some time to run some comparison tests with 3.10 to see the difference.

Best,

Rich

[RobPhoboS]

I need to conduct some testing with unified as well, my first tests weren't impressive and in fact had long render times, haha. Oh well.

[Remydrh]

Take a look here for some more hints on how to get it to perform well: http://elementalray.wordpress.com/20...or-the-artist/

It's actually very simple.