Quick 3.3 Report....

[Alan Iglesias]

So I was in the middle of rendering out a short animation when 3.3 was released. I was using RT/CPU just because it was easy to adjust everything as I started the rendering. The frames had been rendering in the 10 min. range on my 5960x.

Previously, I did do a test using Production Vray with the file and using Progressive Sampler and BF/BF, cleaning up the images was also taking in the 10 minute range as well.

So I stopped the rendering to install 3.3.

After installation and bringing up the same file, the RT/CPU frames were about the same (which I expected), but then I re-tested the file using Production Vray, again using the Progressive Sampler and BF/BF.

Lo and behold, not only were the images cleaned up in half the time, five minutes, but they were cleaner than the 10 minute frames in Vray 3.2!

I guess I have this to thank:

(*) V-Ray: Better sub-pixel filtering for the progressive sampler;
(*) V-Ray: Dynamic noise threshold for the progressive sampler for more even noise distribution;

Then it was on to test the new Adaptive Sampler. After some balancing of min/max and the Color Threshhold amounts, I was able to get a clean image in ~3.5 minutes! Very cool.

Needless to say I am finishing the animation using the Production Renderer in 3.3!

Thanks again for all the hard work on 3.3!

-Alan

[king_max]

I just installed the SP3 in the middle of a project and the scene is much, much noisier with the same settings... It's a complex scene with lots of foliage and lit by a HDR in a dome light. I s there a way to override the subdivision just of the dome light? Yesterday I was using the method described by Vlado in the V-ray sampling tutorial and the images came out pretty clean, but today with the new V-ray build it's just too noisy. If I bump up the min. shading rate and/or reduce the clr. threshold, the render times gets really slow.

[vlado]

I just installed the SP3 in the middle of a project and the scene is much, much noisier with the same settings... It's a complex scene with lots of foliage and lit by a HDR in a dome light. I s there a way to override the subdivision just of the dome light? Yesterday I was using the method described by Vlado in the V-ray sampling tutorial and the images came out pretty clean, but today with the new V-ray build it's just too noisy. If I bump up the min. shading rate and/or reduce the clr. threshold, the render times gets really slow.

It's generally maybe not a good idea to update in the middle of a project; with that said, it would be great if I can take a look at your scene or some small part of it.

Best regards,
Vlado

[king_max]

I've attached the comparison of the 2 v-ray versions (small crop of the final 5k image) and a previous test render just to show the scene complexity. The 3.3 version is 5 time faster to render for obvious noise/sampling reasons. The point is, I didn't change anything in the render settings. How exactly do you control the render quality if using the automatic dmc sampling in 3.3? I've tried lowering the clr and dmc threshold to 0.003, but the image quality didn't change much. Same happened by using a min. shading rate of 64. The image was a bit cleaner, but still nothing near the 3.2 version. The only reason I've decided to switch to the new v-ray is to harness the performance gain (especially with Forest pro).

[king_max]

I've been messing around with the file and also tried the v-ray quick settings (BF+LC) all maxed out. The final render took more to finish than the 3.2 version and was still noisier. With the 3.3 I just can't get a clean image no matter how much I push the settings.... Switching to sun+sky doesn't help either. Using the universal method and lowering the thresholds to 0.002 still produces really noisy results.

[Jvincentfx45]

I have to say I have noticed the same issue or similar I cannot compare from 3.2 to 3.3 but can with the new default settings. This was done using my default scene set up, I use one for Sun and Sky and another for HDRI the material is a simple blinn for material override, 3.3 is faster than the file under 3.2 so that is already a plus but there was for sure some noise and rendering speed things that were interesting. I have noticed the noise issue mainly on the HDRI lighting, the defaults are no were close to being clean enough using the default settings. I notice the noise the most in the lighting pass, for the HDRI using the defaults you can see the lighting pass is super noisy. For this situation I have been going back to using Grants way of working it is interesting because its not only cleaner but faster as well. For the HDRI around 27 seconds faster and for the sun and sky 3 seconds faster. In this test out side of my ram settings everything was left alone except using the local subdivs and clr threshold is down to .005. The Sun and Sky produced very similar results and the speed was only slightly faster using Grants methodology. Its really nice to test this and easy with the local subdivs option. In any event its an amazing update and I am not going to complain about a few seconds here and there but just interesting talking points. I have attached a few images from the test, for the HDRI I am using Peter Guthrie series 2 1501 cirrus clouds. Thanks for the hard work its really amazing update. Best, J


Note: in this I said issue and that is not really true I guess. These default settings are there for a good starting point and to help people not kill there scenes with over sampling and what have you. So this was just really a comparison not in anyway a issue.

[vlado]

I've been messing around with the file and also tried the v-ray quick settings (BF+LC) all maxed out. The final render took more to finish than the 3.2 version and was still noisier. With the 3.3 I just can't get a clean image no matter how much I push the settings.... Switching to sun+sky doesn't help either. Using the universal method and lowering the thresholds to 0.002 still produces really noisy results.

Well, like I said, it would be best for me to check your scene, or a part of it. You can send it vlado@chaosgroup.com - I'm quite sure it's possible to render it with the same quality levels as before for shorter render times. If the quality doesn't change when lowering the noise threshold, it means most likely you don't have enough max. subdivs. But there might also be other things in the scene that I don't know about.

Bes regards,
Vlado

[Jvincentfx45]

Here is a sample with all the passes. Just for thoughts I think for me I am always looking for fast and the best quality anyway just sharing and if I am missing something that would decrease the render times and have a nice quality from just tuning a few setting that would be amazing and I am sure my students would love it as well. If i could just say hey look ignore this and you will have nice renderings and they will come out in a decent time then cool. But sometimes you have to work for things and thats ok to.

Image Set_A: is just allowing vray to do its thing I am only adjusting the clr threshold for quality starting at .01 and moving to .005 still the noise is really only a issue in that lighting pass and a bit in the spec pass, the rgb pass looks fine but I want to be able to control things later in post.

Image Set_B: is the same set up but this time I have taken over and am controlling the subdivisions of the light, material and the brute force sample. clr threshold is the same starting at .01 and moving to .005. you can see the light pass is cleaner and the spec pass if we are being super fine with detail is cleaner as well.

What I find interesting in all this I am still making a cleaner faster image with a lower clr threshold if I take over and adjust a few things 41.3 seconds at clr threshold of .01 as opposed to 1.51 min under clr threshold of .005 using vray method. I should say i could be doing something way wrong but does not feel like it. I guess as people have said I could throw more AA but that seems to be at a cost of rendering times which going up on this simple scene from 1.51 would not be great.
I can not get the same results if i switch to using a vray Sun and Sky it is a much tighter race in that case which is interesting and I wonder if it has to do with using a HDRI, anyway a fun day of testing while i was with the baby and the wife was out.

[^Lele^]

The only thing that stands out to me is the simplicity of the scene you are testing.
Try with a more complicated one, start wrangling individual subdivs, and see how far away you get from the sp3 defaults.
You'll soon discover that when all is not flat or empty, but rather detailed, things change about quite a wee bit.
But by all means, keep testing!

[Jvincentfx45]

Oh I agree it's super simple scene and that for sure could be the case when you start adding things in. I will see about digging one out and test on that but I think interesting all the same. I am still interested in the the noise in the light pass I can not reproduce this using the Vray sun and sky or regular lights, things clean up just fine with little time diff but using an HDRI just not happing but you are correct super lame scene and as I said just what I drop projects into so will be interesting to see.

Anyway still think the update is amazing and still faster than 3.2. If I get some time this weekend will try and test another scene that's actually a scene other wise maybe next week.

[^Lele^]

Your HDRI likely doesn't have the same float intensities in it common to light fixtures and the sun and sky.
Which is also part of why in your specific case you're getting odd results (as there is hardly any contrast in sight.).
Open an interior with textures, maybe some fur, lit with IES lights or sun and sky, then sp3 will clearly shine.

[Jvincentfx45]

Oh please don't get me wrong I think the update is amazing and already shining. I use the sun and sky sometimes hdri other times just depends so in any event it's still faster and I will say I could never clean up an image using a clr threshold of .01 in 3.2 the way it seems to clean up in 3.3 even in this super simple test so that alone is really amazing.

I am also ok with doing some small tweaks on things nothing usually comes with a simple click of the button. I really do appreciate the thoughts on the hdri it's interesting for and something to keep in mind specially if you are in a bind. Also something that I will share with my students.

Best, J

[^Lele^]

Oh, but i'm not offended, eheh.
I'm just stressing the point that you should try SP3 on a production-grade scene.

This is also because i really wouldn't want people to think lights/shaders/effects tweaking is needed, as it isn't anymore, to get ideal speeds out of the render, for a given cleanliness.
While it may look that for synthetic setups with one predominant characteristic (in yours it was the flatness of a third of the screen, in other synthetics i prepared, i had normals changing many times per pixel, across the screen, and so on.) some tweaking is beneficial (in yours, a raise of MSR would surely lead to lower renderimes.), when taken into the context of a "standard" scene, where geometric and shading characteristics are well mixed (fur, sss, area and ies lights, bright speculars and very dark corners) there really is no provable benefit in touching a thing.

Vlado posted in another thread (linked below) one of the spreadsheets of our tests, so that it's obvious where the optimal is (lowest point in the curves).
Notice how the optimal (ie. quickest for a given quality) is steady regardless of the scene type: without actually doing the tests, straying from the defaults is just as likely to slow you down than to speed you up.

http://forums.chaosgroup.com/showthr...245#post674245

While the graph shows MSR, it's just as valid, if not more so, for individual shaders and lights subdivs.
One of the most common causes of slowdowns when setting up a renderer (not just V-Ray!) is oversampling.
That is, in the effort to clean noise, one does so going well overboard with sampling, below the visibility threshold, without realising.
Since there is no visible noise either way, the oversampling often goes unnoticed: in other words, if your dome is visually clean for a given scene with 24 subdivs, and you set it to 64, you wouldn't see a difference, but end up oversampling.
With this approach, new to SP3, that risk is greatly reduced, as it's left to V-Ray to do the hardest of balancing acts.

Set a noise threshold, shoot.

[Jvincentfx45]

Well I must say you are 100 percent correct I did a quick test using the scene peter included with the hdri. I unloaded and reloaded vray so everything is set back to defaults. I did a test were I tweaked the same settings and yep it was junk 6.12min compared to the default at 3.27min so that is that super sweet I am more happy than i was and i was already quite happy. I am sure it would be more obvious with more complex textures ect and I am sure it could still be tweaked a little and the time may reduce.

Anyway super cool great to see, not much more to say. thanks for the input and pushing to do some more testing.

[king_max]

Well, like I said, it would be best for me to check your scene, or a part of it. You can send it vlado@chaosgroup.com - I'm quite sure it's possible to render it with the same quality levels as before for shorter render times. If the quality doesn't change when lowering the noise threshold, it means most likely you don't have enough max. subdivs. But there might also be other things in the scene that I don't know about.

Bes regards,
Vlado

I'm uploading the whole scene. It will take a while since it's huge. Anyway, there is a HDR/dome + sun combination as well as a sun+sky setup, both of which gave me unsatisfying results using any combination of subdivs/thresholds/quick v-ray production settings etc... in automatic mode. The noise is generated mostly from the GI (both with HDR or sun+sky). You can hide/delete the forest object for quicker scene preparations. Try working on the first house on the right since it's in shade and it's where the noise is most noticeable. I can render a clean scene with sp3 only when using manual mode, thus setting all the subdivs manually.