Micropoly detail

Started by Hannes, February 19, 2017, 05:24:48 AM

Previous topic - Next topic

Hannes

In my "Urban scane" thread I learned, that if all you see is objects, you can reduce micropoly detail to almost zero, let's say 0.01 because it only affects the terrain.
Sounds good, but accidentally I found out that TG renders faster with a higher micropoly detail value. In my test (default scene, one tree loaded, zoomed in on the tree and rotated the camera a bit, so that there is terrain behind the tree, AA 4) TG took 22 sec with MD 0.4, 32 sec with MD 0.01! and 33 sec with MD 0.8, so from a certain value on rendering times get higher again of course.
I thought that reducing MD would lower render times, but obviously it doesn't ??? ??? ???

archonforest

Interesting. Wonder what Matt have to say on this.
Dell T5500 with Dual Hexa Xeon CPU 3Ghz, 32Gb ram, GTX 1080
Amiga 1200 8Mb ram, 8Gb ssd

Oshyan

Interesting. I can't immediately think of a reason for this, but I don't know all the intricacies of the render process. I will say, however, that my first question is how this effect "scales". In your test the low micropoly detail render is taking something like 45% longer to render, which is a big deal. If that is maintained even on longer/more complex renders, that would be interesting and important to know. So for example if you have a similar scene that takes 5 or 10 minutes to render with micropoly detail at 0.8, would it then take 7-15 minutes with it at 0.01? I doubt it, but it's important to know. If instead it would render in, say, 5m10 seconds, or 10m10s, then it really comes down to a 10-20 second difference *on all renders*, i.e. it is not proportional but rather a set amount of overhead, then it probably doesn't much matter.

- Oshyan

Hannes

Thanks for your reply, Oshyan. I'll give it a go.

Hannes

Indeed interesting. And a bit disappointing. Unfortunately it seems to be (some sort of) proportional at certain values. I did a test with a heavy old scene of mine I dug out to try some settings. A few instances of a very detailed model with a lot of image maps and three populations of trees. Otherwise more or less default. See attached image.
Rendersettings: Micropoly detail (MD) - Anti-aliasing (AA):
0.4/6:     7.18 minutes
0.8/6:     7.59 minutes
0.01/6: 12.42 minutes!!!!!

I also did a test at
0.1/6:     7.13 minutes.

So, I'd say, there is no benefit in reducing micropoly detail when you're rendering just objects.
But I remembered me writing that exactly this works in my "Urban scene animated" thread:
http://www.planetside.co.uk/forums/index.php/topic,22211.msg224146.html#msg224146

Confusing...

Oshyan

Thanks for testing this. There was one thing that occurred to me after I wrote my last reply and your tests here confirm my suspicion: there is not so much a problem with "low" micropoly detail value as there is with *extremely low* micropoly detail. :D

Notice the results of your last test, which is funny, you sort of tacked on at the end but seems like the most important to me: 7m13s, faster than all the others. The only one that seems genuinely odd is that 0.01 at 12 minutes! My guess is the render engine is not good at dealing with that value, which is unnecessarily small anyway. It seems clear that reducing to 0.1 is beneficial though, right? Even if only to a small degree.

I wonder if others take the advice of "reduce micropoly detail with object-only scenes" to such "extremes". I always just set it to 0.1, which is why I probably have never seen this issue. But indeed it could be that others are running into this.

- Oshyan

zaxxon

Intriguing observation. I had a go at testing it as well. The tree in the image is fairly dense at 2 million tri's (432 meg). The surprising difference in the three MD settings is the obvious 57 minutes for the .01 setting. However as I watched it render, the actual terrain was rendered fairly quickly (comparable to the other settings), but the object took almost all of the render time. So it seems that the extreme low settings actually cause the object to render more slowly, and in my example, extremely so. This is probably a moot point as not many folks would consider such low settings. However, the render time difference between .1 and 1 in the example is interesting. A 10X's jump in MD quality for a 9% increase in render time. Curious if there is a happy trade-off somewhere in using high MD counts?

Dune

Interesting indeed. I wonder what would happen if you it to zero or a minus number? Render for infinite time?

Hannes

#8
Oshyan, you're absolutely right. In my last test I found that 0.1 is the best value for this. I was thinking that I had a MD of 0.01 in my "Urban scene animated" file, but I checked it, and it was 0.1.

Ulco, I did a test with zero out of curiosity with the default scene and a tree. It's incredibly fast, but unfortunately the tree doesn't show up at all  ;D.
I even tried -0.1. Looks the same (I never thought that it's possible to use negative values, but of course this is nonsense).

Zaxxon, thanks for testing it as well.

Hannes

I did some more tests and found out that you can get a little lower than 0.1 and get some more speed improvements.
I used the same scene like in Reply # 4 and tried a MD of 0.08, and it took 7.01 minutes to render. Twelve seconds faster than with 0.1. Not too much, but if this speed improvement would be growing proportionally it would be quite a boost for renders that take an hour or so. Have to check that out.

Hannes

Tested it with the same scene at 2.4k resolution.
0.08 is not remarkable faster than 0.1, so using 0.1 is probably the best.

Dune

Well, that's an excellent conclusion and a good number to keep in mind. Not too hard to remember either  :P Thanks for doing this, Hannes!

bobbystahr

I've never gone that low myself and often before I got higher AA with the paid program I would put MD as high as I could for better free rendering..I can remember 0.1 as well. Thanks for working thru this Hannes.
something borrowed,
something Blue.
Ring out the Old.
Bring in the New
Bobby Stahr, Paracosmologist

bobbystahr

#13
With the workstation down(TG4 lost a .dll and won't start so I shut the box down)I decided to give this a try on the dell with only 8G RAM so I certainly should have enough with 16G in the other machine?????. Bad RAM karma or something as 1/2 the original RAM I got(2x8 G) was bad as well.
The first is MD.6AA2 and the second is MD.1AA6. Although the 2nd one takes 4+ x the time it's a much nicer image overall.
something borrowed,
something Blue.
Ring out the Old.
Bring in the New
Bobby Stahr, Paracosmologist

Oshyan

Bobby, in a situation like your test case you have notable terrain elements visible. So reducing micropoly detail that significantly is obviously affecting your terrain rendering quality. I would be less extreme with it in such a situation. Most of the other tests here are, I think, referring to situations where little or no terrain is actually visible (for example when totally covered by plant objects). But I understand yours may have just been a test to see how it goes. I just want to make sure you do not intend to ultimately use .1 micropoly detail in an average scene where terrain might be visible.

- Oshyan