home
StrataCafe Forums
Username:
Password:
Save Password
Forgot your Password?

 All Forums
 StrataCafe
 Strata Design 3D CX
 Need Help Reducing Render Time
 New Topic  Topic Locked
 Printer Friendly
Next Page
Author Previous Topic Topic Next Topic
Page: of 2

Mike Lynch
A Regular

USA
171 Posts

Posted - 04/18/2006 :  11:46:55  Show Profile  Visit Mike Lynch's Homepage
OK here's the deal: I need to re-render this image for print... currently it's about 973 pixels wide and 72dpi. To get the image to the size required at 300dpi I need to render it at 5000 pixels wide at 72dpi (the image would then be resized/resampled to be 300dpi in Photoshop)

The problem is that using the render settings I created for my initial render means that I will miss my deadline... it's currently been rendering for 13 hours and it's less that 25% done. (for the sake of comparison: the original render took about 8 hours).

I'm looking for assistance in how to modify them to decrease my render time while retaining the quality of the image. Here are my current render settings.

Thanks guys!








My Gallery

Edited by - Mike Lynch on 04/18/2006 11:47:46

jpro
A Fountain of Information

USA
4991 Posts

Posted - 04/18/2006 :  12:09:00  Show Profile  Visit jpro's Homepage
Holy cow, Mike. Your oversampling and reflectivity bounces are gonna kill you. I'm looking other things over, but those two jump out as overkill right off the bat. I haven't looked at your image yet, but I seldom go over 3 for reflectivity bounces, or over 16 for max transparency.

Jean
Go to Top of Page

jpro
A Fountain of Information

USA
4991 Posts

Posted - 04/18/2006 :  12:11:43  Show Profile  Visit jpro's Homepage
okay, so I looked at the image, and see why the subject matter calls for all the reflectivity and transparency recursions. Ouch, you have some render time ahead of you. Did you try dialing those back a little to see if you lose anything? What machine are you rendering on? do you have more than one machine available?

Jean

Edited by - jpro on 04/18/2006 12:12:36
Go to Top of Page

Mike Lynch
A Regular

USA
171 Posts

Posted - 04/18/2006 :  12:59:36  Show Profile  Visit Mike Lynch's Homepage
Hey Jean, thanks for getting back to me. I did a test render and brought the Reflectivity Bounces down to 12 and the Transperancey Layers down to 24 and it looks OK... Not much in the way of a noticeable loss. We'll see how it goes (time wise) from there. :)

I'm on a mac - 1.8 GHz G4 - 1.5MB RAM - ATI Radeon 9800 video card - OS 10.3.9. And unfortunately I only have the one machine at my disposal.

...I'm open to further suggestions. :)

My Gallery

Edited by - Mike Lynch on 04/18/2006 13:03:51
Go to Top of Page

jpro
A Fountain of Information

USA
4991 Posts

Posted - 04/18/2006 :  13:30:38  Show Profile  Visit jpro's Homepage
well, I certainly wouldn't render at more than 9x oversampling. 4x will speed it up a lot, but it is up to you whether that image quality is acceptable. you could probably increase your octree height to 11 or so and kick it up quite a bit.

I don't feel so comfortable commenting on your radiosity settings, but the raytracing changes I have suggested should help a lot.

If you had another machine, I was going to suggest breaking the rendering into strips, but it doesn't sound like that's an option.

Jean
Go to Top of Page

MrChristopher
A Fountain of Information

USA
2220 Posts

Posted - 04/18/2006 :  14:16:36  Show Profile  Send MrChristopher an AOL message
Jean is right, your reflectivity recursions are mostly wasted processor cycles. You'll find diminishing returns in terms of visual appearance beyond just 4 reflectivity recursions. I would be tempted to set it to 4 and go from there. In fact, the few times I've needed more than 2 I've done a comparison render between 2 and 4 to see the difference. Beyond 4 it became more difficult to see the extra ray bounded in the reflections.

The transparency recursions of 36 just means IF there are a lot of transparent in front of transparent surfaces that the ray will continue IF there are that many transparencies. I doubt for your image you would need more than the default which I think is 12 or 16.

Jean is also right that at the resolution you're rendering this you're killing yourself with the oversampling rate so high. In fact at 5000 pixels wide I would just render it at 4x4. Make sure you set your pixel block detail size to 2 (or 3 even) and the subdivision limit to 6 as they are now. Setting the oversampling to 4x4 will automatically alter them.

Your RD settings look fine. Although for this high rendering you may be able to adjust the cache parameters a bit more losely to gain more speed.

Your front threshold of 16 produced a good image we're seeing in your current rendering. Because the cache parameters are tied into the actual resolution of the image being rendered, you can adjust those for larger renderings. So you can probably get away with setting your front threshold to 20 or 24 and the min/max values to 16 or 18 respectively.

Hope that helps,

Chris
Go to Top of Page

mwick
Regularly Helpful

Germany
847 Posts

Posted - 04/18/2006 :  14:27:14  Show Profile  Visit mwick's Homepage
I would first check to octree height: do a small render (very small) with octree of 9, of 10 of 11... up to 13, to see, which will be the fastest. i bet, 12 or 13 will be best. and will save time.
And, yes, think about reducing the oversampling to 9, better to 4, it may be enough when rendering at 5000 wide. And this will save LOTS of time.
Maybe reducing the max. diffuse lightning samples to 400 and the max. light/shadow samples to 50 - and check "rendomize light samples". this will save much time, too.
Im not shur if it works to calculate the render time by calculating 987 x 622 = 613914 pixels, neded size 5000 x 3151 = 15755000, thats 25.66 times the time for the render - means about 205 hours....
I think with octree of 12, 4 x oversamling, diffuse light 400, max light/shadow samples to 50 withrandomize, you will go down to 60 hours or less. try those hints with very small images, that may help you find out whats best.

and maybe mr. "the god" christopher my step in here... he may have the best tips for this.

And: If dont have one, buy a qud G5 :- )

Markus

Strata 3D CX 4.2 Mac and Windows, Render Pro 4.2 on Mac and Windows
Dual G4 1.25 GHz, OS X 10.3.9, 2 GB of RAM, 480 GB Harddisk, DVD-RW, CDRW, Radeon 9000
DualCore Athlon X2 4200+, 2 GB RAM, Win XP Pro SP2, 120 GB Harddisc, ATI Radeon 9600
Sempron 2800+, 1 GB RAM, 120 GB Harddisk, ATI 9200, Win XP Pro SP 1
Go to Top of Page

MrChristopher
A Fountain of Information

USA
2220 Posts

Posted - 04/18/2006 :  14:28:50  Show Profile  Send MrChristopher an AOL message
I should have also said that you may get a bit more dynamic image if you set the l ight handling to Clipped instead of Scaled. Scaled will attempt to bring very bright areas back down which can mute colors. You'll know if any areas are being compressed by testing parts of your scene out in Debug mode. Those areas will be rendered in red. in this case you can render with Clipped to get the better dynamic range.

Just a thought.

Also, have you looked at using fewer shadow samples? 80 may or may not be too high, hard to tell without doing a test. I will say this, in 5.0 you'll have the option of using randomization for JUST the shadow samples which allows you to use fewer samples than what are necessary for non-randomized soft shadows.

Chris
Go to Top of Page

MrChristopher
A Fountain of Information

USA
2220 Posts

Posted - 04/18/2006 :  14:35:03  Show Profile  Send MrChristopher an AOL message
As Mwick said, adjusting your octree may or maynot benefit this scene. The scope of this scene isn't broad unless you have large glow panels sitting well off the visual area we're seeing. I'm guessing you've got the Lightdome setup for reflections. This wouldn't alter the need for a higher octree setting.

However, it always worth spending a few minutes with a few test renders to see whether a higher octree will shave time off an image. In fact, you can do this without having raydiosity on. try a sample with 10 and 11 and see what you get timewise. You can render just part of the image to see.

As for diffuse RD samples, it really depends on how the scene is being lit. If the bulk of the light is coming from a light panel or Lightdome (especially if an HDRI image is used), you'll probably need more samples. but if the bulk of the bounced light is a result of secondary bounce from a spot or point light, then you can reduce your samples.

I would not recommend that you switch on randomization for your scene. I'ts likely to introduce artifacts.

Chris
Go to Top of Page

Mike Lynch
A Regular

USA
171 Posts

Posted - 04/18/2006 :  15:09:15  Show Profile  Visit Mike Lynch's Homepage
Thanks, guys. I really appreciate all of the input. I'll definitely modify the Oversampling and Cache. And I'll adjust Reflectivity and Transparency as you suggested, Chris and see where that takes me.

Chris, regarding your questions about Octree Height and the lighting: there are three very large glow panels outside the visible range of the camera in various positions (...essentially: one on either side of the camera, and one in front and slightly above). The global reflections do not mimic the lightdome. I used NYC Met. (HDRI) for the lightdome, but it was creating to many harsh reflections when used as the global reflection. I'm using "Space Balls" (believe it or not) for the global reflection. After playing around with numerous options that one actually worked the best in this scene - it distributed small amounts of color and offered some really subtle and interesting reflections to the various surfaces. The lightdome intensity is actually pretty low (35%) so the majority of the light in the scene is coming from the glow panels. There are no spot lights and no point lights used. ...So I'm assuming that in this case adjusting the Octree Height may not offer any great increase in speed.

EDIT: I just want to mention that the glow panels are using specific images in the Glow Channel. The glow panel to the left of the scene is using SoHo (HDRI) and the panels to the Right and in the back of the scene are using the pastoral image used on the glow panels in the tutorial for setting up a scene with three glasses in the "Art and Science of Strata" CD. So there are several different reflections going on in the scene, aside from the global reflection.

Thanks again - this is very helpful.

My Gallery

Edited by - Mike Lynch on 04/18/2006 15:18:16
Go to Top of Page

JeffrySG
Regularly Helpful

USA
543 Posts

Posted - 04/18/2006 :  15:33:42  Show Profile  Visit JeffrySG's Homepage
Mike, if I were you I'd also email your buddy Jeff and see if he can render it on his dual g5 2.7 at work! ;-) or if you want to break the image up into 3 sections he can render it on all three of his machines for you (prob. tonight) and email you the parts tomorrow morning sometime if they finish overnight! (not sure if they would)

drop me an email if you want me to... especially if you are going to adjust some of the settings first.

(to break a single image up into sections you can just put a black poly infront of the camera where you don't want to see that section of the camera view.)
Go to Top of Page

MrChristopher
A Fountain of Information

USA
2220 Posts

Posted - 04/18/2006 :  17:11:02  Show Profile  Send MrChristopher an AOL message
It would be interesting to see a screen grab of your setup, but FOR SURE if you're using large glow panels then you need to bump up the octree height. Otherwise the renderer is spending a lot of time searching through empty space for geometry intersections. By increasing the octree if makes this process more efficient and hence faster. The downside is that on scenes with a lot of heavy geometry it can take more memory to render.

I would try rendering your scene with 12 in this case for the octree.

Chris
Go to Top of Page

Dasch
Regularly Educational

Canada
1164 Posts

Posted - 04/19/2006 :  00:30:17  Show Profile  Visit Dasch's Homepage
Another tip; unrelated to render settings.
Do you NEED to go to 300ppi for your print job Mike?
If you are having laser prints made, maybe 240ppi would do it.
I heard at RRR (from a print specialist) that even lower resolution works well.
The 'Alice' poster I just did was rendered at 240 ppi and digitally printed and the detail is very sharp.

Dasch

Edited by - Dasch on 04/19/2006 00:31:00
Go to Top of Page

Mike Lynch
A Regular

USA
171 Posts

Posted - 04/19/2006 :  09:17:33  Show Profile  Visit Mike Lynch's Homepage
Thanks, Dale. I had actually thought of that myself when I started this, but unfortunately I do need to present the image at300dpi.

As it stands I modified my render settings taking in all of the advise given thus far - with the noted exception of Randomized Light (as Chris mentioned It was creating artifacts) I've reduced my Oversampling to 4, I bumped up my Cache to 24/18 as Chris had suggested, I increased my Octree Height to 12 as per Mwick and Chris, I reduced my reflections and transparency layers (4 and 16 respectively. The scene has been rendering for a little over 10 hours now and I'm already well beyond were I was with my previous settings after 13 hours. Thanks guys.

And I may still take you up on your offer, Jeff - depending on where I am in the render tonight, but right now things are looking pretty good.

My Gallery

Edited by - Mike Lynch on 04/19/2006 09:19:07
Go to Top of Page

Mike Lynch
A Regular

USA
171 Posts

Posted - 04/19/2006 :  10:18:55  Show Profile  Visit Mike Lynch's Homepage
Chris -

I just realized that I forgot to show you the scene set-up... I hope these give you a good idea of what's going on. The ground plane is pretty small - you can see it's proportions to the glow panels best in the custom view.

Top View


Right view


Left View


Front View


Custom View



My Gallery
Go to Top of Page

MrChristopher
A Fountain of Information

USA
2220 Posts

Posted - 04/19/2006 :  12:08:41  Show Profile  Send MrChristopher an AOL message
Oh boy, just looking at those pics I can see why octree 9 was killing you. I would even try 13.

Chris
Go to Top of Page

JeffrySG
Regularly Helpful

USA
543 Posts

Posted - 04/19/2006 :  12:23:25  Show Profile  Visit JeffrySG's Homepage
Just looking at the size of those panels, was there a reason they ended up so big? I kind of think of using the big glow panels like you would use a soft-lightbox in a photostudio... and those would be some MASSIVE softboxes!!! haha... If you were to rework it (at a later date) I'd be curious if you could make the panels much smaller and just bring them in much closer to get the same result but quicker?

If you need me to finish the render just let me know, Mike...

Jeff
Go to Top of Page

Mike Lynch
A Regular

USA
171 Posts

Posted - 04/19/2006 :  12:23:47  Show Profile  Visit Mike Lynch's Homepage
Thanks, Chris. As always, your input is greatly appreciated.

My Gallery
Go to Top of Page

Mike Lynch
A Regular

USA
171 Posts

Posted - 04/19/2006 :  12:31:28  Show Profile  Visit Mike Lynch's Homepage
Hey, Jeff. I suppose I could have gone with smaller and closer - that's a good point.

The size of the glow panels was based on a tutorial on Chris' "Art and Science of Strata" CD. Specifically the tut. dealing with ligting three glasses for a photostudio-esque render. I just used that as a reference point when building my scene. I'll have to give your suggestion a shot and see how it comes out.

My Gallery
Go to Top of Page

MrChristopher
A Fountain of Information

USA
2220 Posts

Posted - 04/19/2006 :  12:58:23  Show Profile  Send MrChristopher an AOL message
I should have said that this is a good lesson about the importance of properly setting the octree for different scene setups. It can massively impact the speed of rendering. The default of 9 is a modest value that works well for many average scene that are fairly compact in nature. But in this circumstance there is a lot of deadspace that is efficiently subdivided by the low octree value and as a result the renderer wastes a lot of time hunting through empty space looking for goemetry intersections. When I have scenes like this, I don't even try and do a rendering with 9 but automatically try 12, then I do a test on either side (11, 13) to see which direction gives me a faster rendering.

Now, I always have the process viewer open when doing these testings to see what memory is doing. For this scene, memory isn't likely to be an issue because the geometry count is low. However, for scenes where you may have many 100's of thousands of polys, or millions the higher octree levels can actually end up slowing the rendering down as the renderer is forced to start chugging through large amounts of virtual memory (hard drive). The key is knowing how much actual ram you have and watching the actual ram usage. If you suddenly see actual ram usage go way down and VM go way up, and the processor usage starts going way down (the cpu is waiting for data from the hard drive), then your octree is probably too high.

Incidentally you can also end up with a similar situation if your octree is too low for heavy scenes.

Moral of the story, always, always, always run a few tests with different octree values to see what gives you the best ballance between render speed and memory usage.

Chris
Go to Top of Page

MrChristopher
A Fountain of Information

USA
2220 Posts

Posted - 04/19/2006 :  13:05:16  Show Profile  Send MrChristopher an AOL message
On the newer version of the Art&Science CD I'm preparing for 5.0 I will be including a section on this octree setting which I neglected in the first one.

Jeff, the size of the glow panels is fine actually. Many times I've used panels of this size. If they're too small they may not efficiently reflect or render well if you're using them for diffuse illumination. But there is a large degree of size variation you can use.

As to what the octree is actually doing, it's subdividing the scene down using a single large bounding box for the entire scene (minus lights), this is level 1 octree. It then divides that box into 8 boxes that are a quarter the size of the original. that's 8 new bounding boxes, hence OCTREE (8). This is level 2. It progressively subdivides each box down into another 8 boxes (nodes) until it has done the number of iterations specified. If the boxe (node) size hits 9 subdivisions in the case of this scene, they may still be leaving a lot of empty space around geometry which the renderer must parse through wasting precious processor cycles hunting for the surfaces to render.

Chris
Go to Top of Page
Page: of 2 Previous Topic Topic Next Topic  
Next Page
 New Topic  Topic Locked
 Printer Friendly
Jump To:
StrataCafe Forums © 2009 StrataCafe Go To Top Of Page
Snitz Forums 2000