CHALLENGE - Show your skills (scene included)

[cecofuli]

Oh, yes, obviously =) DR of course for big still. BB for HD animation.

[LarsSonparsson]

Just two versions more. With and without Refractive Caustics.
I got render time down (under 1h) using interpolation in most glossy reflections. Not usually doing that, but I had to find some ways to make it faster.

[cecofuli]

Nice time/quality. I forget the existence of glossy reflection )) But I can see very bad GI on the right side of the dark furniture.
sub-pixel off or on?

[LarsSonparsson]

Allways sub-pixel off. That was the point right?
Yep some bad gi, but I think that can be fixed easily.

[cecofuli]

Oh, yes! Honestly I would liked to see if there was some possibility to make a better render time with BF+LC, without tips an trick.

[raymarcher]

I guess also without Vray 3.0

Oh, yes! Honestly I would liked to see if there was some possibility to make a better render time with BF+LC, without tips an trick.

[LudvikKoutny]

Okay, here we go:



This took 27m 40s on my old i7 870 which is performance-wise somewhere on the level of i5 2500k. I think i broke the rules of the game a bit, because i changed really lot of things, but for the sake of the optimization, i had to.

(DISCLAIMER: Any information presented under this line is based on my biased opinions and experiences, and are not claimed to be true nor verified)

First of all, it is rendered at larger res - 1280x800, so that's another plus for me

I never trust stranger's scene, so i merged everything into a new scene and redone from scratch. I tried to stay in original one, but i had really goofy things happening all over. For example GI was rendering some weird vivid saturated colors even when everything had gray MTL override.

The scene looks a bit different, as i moved camera (i will explain why later) and i redone all the materials from the scratch, so i was sure there wasn't some overshot value hidden anywhere in the scene.


Anyway, before i started to recreate the setup, i took a look at the original one, and took a notes of possible things causing problems.

The biggest problem overall i saw is that you have built a really difficult scene for the renderer, and you tried to make renderer crunch it all. If you want to play physics simulation, that's alright, but if you have to hand over renders to the client next day, then you usually have to do some compromises. You should always make your scene as easy for renderer as possible without sacrificing any significant amount of quality. I do not think there will be ever any renderer that will crunch any scenario in reasonable time if the user is not caring about what they are doing.

Now for the specific problems i found:

1, There were too many light sources. Vray is a regular raytracer, not any state of art path tracer, so increasing amount of lights will always slow it down. If you have large light surfaces like those ceiling floodlights it is a lot better idea to use emissive material rather than many light sources. While in something like Mental Ray you would not get away with that cause of it's crappy GI, Vray has very good GI solution(s) that can handle emissive surfaces quite well. If you want some extra quality on the contact areas of emissive surfaces when using Irradiance Cache, then you can try detail enhancement which works very well for such cases. Using light sources such cases is excessive. I have replaced the lights with one stripe of emissive VrayLightMTL.

2, You are using mesh lights. You used one for the light under the glass floor. I would avoid using those, as they do not seem to work well with Vray's importance sampling. I broke my teeth on them several times in production. Again, with a good GI setup, emissive materials (VrayLightMTL) will work just fine Like in the previous case, i replaced the mesh light with emissive material.

3, Your DMC max samples value is way too high. From my experience, DMC min/max value take care about mainly about eye rays, eg. what is seen on first hit. They define how precisely are objects sampled, not their secondary effects. Increasing max DMC samples pays off in scenes with very fine geometry, which is not the case of an average interior scene like yours, which is full of rather simple and large shapes. IIncreasing DMC samples is always at the expense of sampling quality of secondary effects, such as glossy reflections. In my opinion, ideal values for your scene are Min=1, Max=4, and those are values i have used.

4, You had disabled AA filter. Well, no wonder you were fighting fireflies. If you ever want to have correct antialiasing, you need to filter it, so always use some filter. If you do not wan't your image blurred too much, or if you don't want black outlines around highlights, then use VrayTriangle. It's probably the most subtle filter around there.

5, You had set your gamma at 2.0. Do not ever do that. Correct LWF is always 2,2, and you should never use gamma as a means of artistic control. You can add contrast in post, but don't touch gamma. Everytime gamma value is used for artistic control, god kills an adorable kitten.

6, You had sRGB button in framebuffer off. For correct LWF, you should have gamma set at 2,2, don't affect colors checkbox ON, and sRGB in VFB enabled. In my case, i disabled don't affect colors and sRGB button, because with correct LWF, color mapping controls do not work, and this scene was a nuclear apocalypse going on all around without color mapping exposure control. But once new service pack is out, you should really use the correct LWF. (Vlado, come on... you said any day now! )

7, You had refractive caustics on. Unless you use photon mapping, you are not going to get accurate caustics in scenes such as this one, and having refractive caustics on causes more trouble than benefits, so it's a good practice to turn it off, make sure that your glass materials have affect shadows checked on under refraction settings, and disable shadow casting and GI visibility for glass objects that cover key lights of your scene. In case of this scene, i did so for glass in windows and glass floor. It's one of those things, that have very little impact on final look, but very huge impact on rendering performance.

8, You had your Light Cache set mode set to world. Unless you do some animation, i would suggest to keep LC sampling in screen space, as it provides a lot more of adaptivity and flexibility. If you have leaks, there are better ways to control them than world space interpolation

9, You had LC retracing enabled. Retrace threshold has a huge impact on rendering performance, especially when using LC with IC. Turn that on ONLY when you experience leaks that can not be removed by any other methods (such as using check sample visibility in IC settings). Never ever use this feature preemptively, only use it when you encounter the problem that can not be solved any other way.

10, You had acc structure set to Auto. If you are sure your scene will fit into your RAM memory with ease, then always use static. It might eat more memory, but unless your are running short of it, it pays off when it comes to performance. Use Auto or Dynamic only in cases where your scene starts to eat significant amount of your free RAM (let's say more than 75% of it)

11, You had too small bucket size - 8. If you wan't your rendering to be efficient, i would not recommend going under 16x16 bucket size, unless you are doing very small region previews.

12, You had your DMC adaptivity set too low. In scenes such as these, adaptive distribution of sampling is very important. 0,5 as in your case is unnecessarily low. I would recommend something between 0,7-0,9. Default value of 0,85 is very, very wise.

13, Your noise threshold was too high - 0,015. If you render final output with high noise threshold, then you most of the times end up wasting a lot of samples and still having fairly noisy images. It's better to tweak sampling at very low noise threshold (about 0,005) and work your way with light/material/dof subdivs up until you hit the sweet spot between quality and speed. At least that's my workflow, and it proved quite efficient for me so far. It's more important how well you spend your samples, than how many of them you compute.

14, Your bucket sequence was set to spiral. Hilbert or Triangulation curves are not there just because they look cool. They are there to improve memory loading/unloading efficiency using bucket rendering. So unless you have some good reason to use different curve, i would keep it at default. The difference it makes is probably very small, but it still is a difference

15, For the amount of Max DMC subdivs you had (32) your light sampling was quite low (average of 12) You should keep in mind that light subdivs values are divided by Max samples value in DMC, so you might end up with very sparse light sampling and thus noise. You can keep light samples low and still get clean images, but for that, you have to keep DMC samples low as well (Like Min=1/Max=4 in my case )

16, You use Camera clipping. I have no idea how Vray handles it, but you should never ever do that unless absolutely necessary (like rendering 2*2m bathroom). Camera almost always causes some (in case of Vray probably only very slight) computational overhead, and sometimes might also cause geometry precision problems. The space you are rendering is huge, so there is a lot of room for a photographer to go stand with his camera. He does not need to make a hole inside of the wall to take a picture. This is the reason i moved the camera. so it's not inside the wall and i can disable the clipping.

17, Some of your materials were way too close to black. That makes DMC sampler's life a lot harder. And also, so dark materials almost never occur in real world. I never go under RGB 2 2 2. Not using superblack materials can help you avoid some sampling issues.

18, Glass in your windows had 3 faces O_O. WTF? Look at the bushes behind windows on your renders. Behind door, bush line looks a bit lower than behind windows. That's because it's being refracted. You have 3 faces in your window, and refractive media always needs enter and exit face. If you have 3, then you have enter - exit - enter, thus having all the area behind window full of glass medium, instead of air. You can have two, or 4 faced window glass, but always keep it in pairs Never 3 or 5

[PART 1/2]

[LudvikKoutny]

[PART 2/2]

Now for some other optimizations i did:

I have set portal lights to simple. It is one of those things that usually do not affect look much enough to justify rendertime increase. Simple mode in portal lights is a bit faster.


On materials with very low glossiness, such as that brown material you have on significant portion of your scene, it pay's off to set "treat glossy rays as GI" option to always. It will not impact look of material much, but it can potentially improve rendertime a lot. Especially when the material takes significant portion of the picture. Although, i would not use it for any material with glossiness above 0,5.

I have decreased per-object GI subdivs multiplier for foliage behind windows (via vray properties), as it is not a key element of this scene.

I have replaced some ambient lights with emissive materials as i already described above. I enabled check sample visibility in IC options to prevent light leaks.



And at last, the biggest problem of your scene is rendered DoF, in conjunction with a lot of overlapping refractions. For how small the DoF effect is, it does not pay off to be rendered, and in this case, it would be a bit smarter decision to do DoF in post. Again as Vray is not a path tracer but a regular raytraced, DoF usually adds a lot of rendertime. If i disable DoF in this scene, rendering speed more than doubles. This is very difficult scene for DoF.

On my picture, you can see some grain on chairs and in reflection. That is caused by DoF. My DoF samples are set only to 4, but raising them even slightly to 8 causes rendertime to almost double in this scenario. As even in perfectly focused areas, Vray wastes a lot more rays than it would regularly need to.

There are 3 options to handle it:

1, Do DoF in post, it will be lower quality, but you will have very fast renders

2, Render DoF with lower sampling and then denoise blurred areas in post. This is quite reasonable solution as removing grain in already blurred areas is not that difficult

3, Increase DoF sampling to get clean picture, but expect very high rendering time.

Usually, DoF will not slow down scene this much, but this very scene is quite a hard one, and often you simply have to do some trade offs. Especially when you have a deadline to race

Aaaaaand, i think that's all (Damn, i had a lot of coffee again )

PM me your mail adress and i will send you back the optimized scene. If you want, i can also try to do a BF+LC setup, which should not take that much longer, and could make illumination from those emissive materials even more accurate

[LudvikKoutny]

I arrived at work and did two more renders, because i have a bit faster machine here (Dual Xeon E5-2630 @ 2,3GHz)

First render is same setup as before only with increased DoF subdivs from 4 to 8, and noise threshold from 0,008 to 0,006.

Took 20m 11s at 1280x800 res:



Second image is exactly the same setup, but with DoF disabled.

Took 10m 44s at same res:



This should give you some idea on how much is DoF expensive in your scene

To get perfectly clean output, you'd have to raise DoF subdivs up to 20, and that would not double, but i think almost triple the rendertime. That's why i think that for such subtle effect as one you use in this scene, post processing DoF would be a bit better choice. Alex Roman's T&S is a living proof that post processing DoF does not always look bad

[cecofuli]

Recon442,
I haven't words. Your review is near to be perfect. O__O
About the glasses, I used a pre-model, I didn't know that the windows was done with two pieces of glass...
Yes, it would be nice if you can test with BF+LC method, just to see the difference, with DOF on and OFF.

[LudvikKoutny]

Recon442,
I haven't words. Your review is near to be perfect. O__O
About the glasses, I used a pre-model, I didn't know that the windows was done with two pieces of glass...
Yes, it would be nice if you can test with BF+LC method, just to see the difference, with DOF on and OFF.

I tried to reply to your PM, but it says your inbox is full

[KREATIVE_POINT]

Wow Recon.. excellent optimization and perfect explanation.. compliments

[cecofuli]

Oh, yes. I see. Now is ok

[cecofuli]

Recoon442, if you look your scene, it's less "vivid". No glossy on the black furniture, on the ceiling etc..
Also the transparent chairs are less... I don't know how explain, less "bright". Like with sub-pixel on (but I know, it's on)
And, IMO, your IM is low for 1280 ( -3 -2 , 80 20). You can loose GI details.
Anyway, I try to optimize my scene. Render at 1280:

(*) AA 1 16.
(*) No filter (save a lot of time).
(*) Gamma 2.2. Now the render is too much bright. I had to reduce exposure or light intensity a little bit
(*) sub-pixel OFF.
(*) BF 64.
(*) LC world 2000 - 10cm - Use glossy ray ON - retrace OFF.
(*) DMC default.
(*) subdivs MTL 128 - subdivs light 12.
(*) DOF on - 12 subdivs.
(*) Skyportal simple (I gained a lot of time).
(*) One big light for the ceiling.
(*) floor light as a VRayLight, instead VRayMesh.

The trees: I don't know why, but VRay2Side kill the rendering time.
Also I turn off the glossy (trace reflection OFF).

Now, the rendering time is 2h. I don't know if it's possible to reduce something more

[cecofuli]

One hours, but too much noise.