MENTAL RAY (4)

[noar]

Hey all....

here's my problem....
i can preview my scene just fine... but for some reason, when i go to RENDER the same scene (also using the same renderer and all the same settings (mental ray)) it takes WAAAAAAAAAYYYYY longer.

why is this?

when i hit preview, the image comes up nice and quick.
when i hit render (to test render my animation) it seems to get stuck on 'file parsing' eventually the image will come, but it's NOTICABLY slower than when using the 'preview'

any help?

thanks

[fredc]

Not sure how new you are to Soft 3D, so I will assume basic stuff.
1. Are you sure you are using the same renderer. The Soft renderer which is the default for preview, is usually faster than MR, although it does not support shaders and advanced lighting.
2. Is the ray trace depth more. This determines how many bounces a ray has once it hits something. For instance, a reflection needs 1 bounce, but a reflection of a reflection needs two. Preview defaults to 1, but renderer defaults to 3 or 4.
3. Is the anti-aliasing set the same. Preview has it off by default.
4. Is motion blurring on for render but not preview?

Not sure what else to suggest, but go through each setting and be sure it is the same.
Sorry if this is not helpful and too basic. Just guessing.

Urp

[hypersim]

I'm having exactly the same problem. When it comes to Render with Mental Ray, it says Image Parsing and it takes about 5 minutes to parse the scene. Once the scene is parsed, it only takes about ten seconds to actually render. I opened up task manager (im using Windows2000) and Mental Ray was taking 160Mb of RAM; Soft was also taking up 160Mb of Ram. The scene itself is no more than 10Mb in size. However, the scene does contain about 12 Nurbs objects that add upto about 400,000 polygons. Could this be the reason why it takes Mental Ray so long to parse the file? Does it have to convert the Nurb objects into polygons before doing the render and therefore, as there are twelve moderately high-detailed objects in the scene, take an awful long time to do so? When I render with the soft renderer it takes half the time. I need to sort this out as I am going to render about 30 seconds worth of work and the deadline is only weeks away. I have a Dual 800 with 512 megs of ram, so its not that I have a slow system. Help!

[fredc]

If the Softimage renderer does it in half the time, then it sounds typical. The soft ray tracer is the faster ray tracer I have ever scene (and it isn't even threaded). However, it doesn't support shaders or GI etc.

You might try reducing the step size of the NURBS. Use Info/Step for U or V. If you haven't modified the defaults, you may find a major reduction in polygons with this.

You can accelerate MR with BSP tree etc. Look up acceleration methods in documentation. Depending on what type of rendering you are doing, this can have a dramtic effect too.

Urp

[devdoka]

@If the Softimage renderer does it in half the time, then it sounds typical. The soft ray tracer is the faster ray tracer I have ever scene (and it isn't even threaded). However, it doesn't support shaders or GI etc.

I'm sorry, but I had to respond to this one. The Softimage renderer is NOT the faster raytracer when compared to mental ray. If you know how to use mental ray to any capacity you'll find it much faster than the SI renderer. The SI renderer is view dependent, mental ray is not. That is, SI renderer only gains speed when things don't take up much of the frame, mental ray will render a frame the same almost regardless of where the object is placed.

If you're having problems with mental ray, it's probably the scene tessellation, not the render. To reduce this problem, try parsing out mi files per frame instead of rendering from the interface. It'll take a few minutes to parse, but rendering will go much faster and mirror what you see in previews as you won't have the overhead of the SI3D interface.

Also, if you have 400,000 polygons worth of NURBS (Not sure why you need that), you can select the objects en masse and use Matter > Info > Mental ray to change the tessellation from "parametric" to "adaptive" with curvature and use pixels not system values. This makes tesselllation view dependent so if a 50,000 triangle NURBS object is waaaaayy back from the camera, it won't be triangulated to the full 50K, it'll be triangulated only enough to look nice...which may be 500 polygons, not 50,000. You'll notice the surfaces come out ultra smooth - if not, tweak the curvature and chord length values. Details in the manuals. I also advise you to put subdivision limits on the tessellation so it doesn't get out of hand. Especially useful if you're using displacement mapping which can run amok if you don't keep tabs on it.

A simple demonstration of SI vs. MR:

1) Get > Primitive > Sphere. Bspline.
2) Matter > Texture > 2D Local. Put any texture bitmap on the sphere using UV wrapping.
3) frame the sphere in the perspective view and preview (should be using SI renderer at this point). Notice it goes pretty quickly.
4) Matter > Preview > Setup. switch renderer to mental ray.
5) Preview the scene. Notice it also goes pretty quick, maybe not as quick as SI renderer, but still quick nonetheless.

6) Dolly the camera very close to the sphere's surface until it fills the entire frame.

7) Preview with mental ray, take note of the time - about the same as previous test from step 5.
8) Matter > Preview > Setup. Change back to SI renderer.
9) Preview the scene - notice how slow the renderer goes.

Now for a true test, we must adjust the renderers for final quality output:

10) Matter > Preview > Setup. Click the anti-aliasing button and set the SI renderer to use Bartlett at level 4. De-activate "back culling" to put in on par with mental ray.

11) Preview the scene again. Notice the render goes very slow.

12) Matter > Preview > Setup. Again, switch back to mental ray, then enter the anti-aliasing options. Set min samples to 0, max samples to 2. This will mimick a bartlett 4 from the SI renderer. However, we'll bump up the quality even more by reducing the contrast of the adaptive supersampling options from 0.2 to 0.05 for RGB channels (this will slow down rendering significantly for low contrast scenes). This tells mental ray to continue firing rays until the difference in contrast between samples for a pixel averages less than 0.05.

13) Preview the scene. Notice mental ray hardly slowed at all and is much faster than the SI renderer.

Conclusion: the SI renderer is good for very simple scenes, but when complexity rises and quality needs to be maintained, mental ray will win hands down. Don't believe me? Add more objects to the scene - even simple grids with textures - and compare results. You'll find the gaps in time widen, not narrow. Add raytraced lights and the difference becomes clear. The SI renderer will display tessellation artifacts while mental ray will display smooth surfaces.

Also keep in mind that the SI renderer has a known memory leak. Therefore, the longer you render, the slower it will go.

A few years back a colleague of mine had to render a 4000 frame cinematic sequence for a computer game. It was all low polygon counts with simple textures - maybe 2000 polygons for the scene. Each frame took 7 seconds to render at the beginning using the SI renderer (do the math and the prediction points to a 7 hour render session). The next day when he came in to see the results he was surprised to find that the renderer was still going and only on frame 2400, but now each frame was taking several minutes! He called me in to look at it and at the time we figured some overnight network traffic slowed it all down. So we restarted the render the next night. When we came back in, same problem. Render was not done and each frame was taking up to 15 minutes to render. After stopping the scene and randomly picking frames to test, we found that each frame on it's own rendered in 7 seconds just like the first few. But overtime, RAM had become fragmented and less memory was available for each subsequent frame. Eventually little or no memory was left so the renderer would be forced to virtual memory to do it's work - which explains the slow render times. Softimage confirmed our findings after a few phone calls. The SI renderer has not been updated since - so don't expect anything different.

---

To fix your problems with mental ray, check your textured objects to see if you have activated "roughness" which in mental ray is "displacement" mapping. Whenever you use displacement mapping, mental ray further tessellates the surface so the displacement renders smoothly (eg - it adds more triangles locally to smooth out the displacements). The object may only be 500 triangles, but when mental ray is done with it, it may have over 250,000 if you crank the settings high enough. This is probably where the extra time and memory consumption is coming from. Motion blur will also kill you, so only activate it for objects that require it. Use 2D blur if you can as it will go significantly faster.

If you don't need raytracing, then turn if off and go with scanline. You'll be amazed at the speed improvement.

Finally, make sure you have enough swap space on your machine. Your virtual memory should be at least 300MB on the drive in which mental ray is installed.

If you're still having problems, give me a shout. Also tell me what you're rendering so I can be of more help.

Matt

---

Matt Lind
Animator / Technical Director
Softimage certified instructor:
Softimage|3D
Softimage|XSI
speye_21@hotmail.com

[fredc]

Speye_21,
Chill out. I am on your side. I like Softimage too. We Softy's need to at least try to remain a little friendly to each other.
Urp

[devdoka]

I'm sorry, but this SI vs MR debate one of those age-old questions that keeps popping up along with "where's that photoshop pic importer" and "where can I find 3d models for free". I'm just amazed at how often people keep recommending SI renderer when it's just a booby trap waiting to spring - the SI renderer hasn't been updated in nearly 7 years. Sure, mental ray isn't perfect either, but it's much more manageable, less problematic, and constantly developing. When mr v3.0 makes it's debut later this year, you'll see a mental ray you haven't seen before.

I'm simply trying to provide the best answer to the problem while illustrating what potentially lies ahead. Too often, advice is dispensed without any forethought as to side effects when things go wrong. Situations like this is what causes studios the biggest headaches - they rely on information they haven't tested themselves and therefore paint themselves into a corner not knowing what the options are. It's fine to say that the SI renderer is good for smaller scenes, but long scenes - even if simple - will have problems on a large scale. Knowing what NOT to do is just as important, if not more so, than knowing what to do when near a deadline.

Being an instructor, it's kind of my job to correct a situation when noticed. I'm not going to correct each and every mail I see, but I will step in if I see something that is a common problem and it's not getting resolved correctly. When I teach the certified courses, mental ray vs. SI is 35% of the class because it's that important and majority of people don't understand the differences. In fact, many people are intimidated by mental ray because of all it's available options. Most people start with the SI renderer only because it previews faster for simple scenes and is easy to adjust since it hardly has any parameters. After a while, they stop considering mental ray thinking they'll never get the job done when in fact the opposite is true. The SI renderer is like a sprinter - it starts fast but won't go the distance without tiring out first. mental ray is like a bull and will march to the end passing the tired sprinter on the way.

I don't see this as a "me vs. you" debate. I see this as a "OK, time to set the record straight" issue as I encounter it frequently. I would just like to spread the word so I don't have to face it as often.

Matt

---

Matt Lind
Animator / Technical Director
Softimage certified instructor:
Softimage|3D
Softimage|XSI
speye_21@hotmail.com

[pomme]

I agree with you (Matt) that mental ray is faster in complex scene. However, setting up mental ray is a nightmare. Since mental ray effect dialog box isn't user friendly enough, I have to do many test render to see what different settings mean. Sometimes, i can't even tell the difference with differents settings. I scared to use mental ray on enviorment and textures effects like smoke and stuff. I hope they can fix their interface so that it's more ez to use. Thanks again Matt for rendermap definition.

[devdoka]

@I scared to use mental ray on enviorment and textures effects like smoke and stuff. I hope they can fix their interface so that it's more ez to use

Exactly my point.

The best thing you could do is learn mental ray a little bit at a time. Start with the Rendering manual then move onto the "programming with mental ray" companion book. You don't have to read the boring chapters on scene syntax or other low-level bits and bytes, but there are some diagrams and very precise explanations of mental ray's core. The book was written by mental images themselves, so it doesn't get more accurate than that. You probably won't see any more mental ray enhancements in SI3D - all the energy is on XSI where it has been integrated and improved greatly over SI3D.

Sooooooooo, what kinds of things are you finding to be nightmarish? What things don't make any sense?

Matt

---

Matt Lind
Animator / Technical Director
Softimage certified instructor:
Softimage|3D
Softimage|XSI
speye_21@hotmail.com

[hypersim]

Thanks guys for taking time to help me; I've been asking questions since I started using Soft and this site has always been a great resource. Cheers

My scene consists of a rocket ship and two launchpads, some 'gripper' objects that lock the ship into place and some large floodlights to illuminate the rocket at night. Its for my second year negotiated project: We are a pretend company called Nanotech who have terraformed Mars and are now trying to sell the idea of living there to the public. We have gone for an Art-Deco style from the 40's and 50's so everything is all round windows, long lines and sharp edges. The rocket design is your typical retro 40's comic book archetype and so I have had to trim about 40 portholes into it. This has taken the rocket up to 40,000 polys although it was more at first until I took the steps down of each trim. The two launch-pads and the four grippers have alot of added decor- art dec style- and so, there is no way I can further reduce the scene without litterally deleting objects or compromising the artistic style. I've now converted everything other than the Rocket to polygons so it parses quite quickly in the Softimage renderer but it still takes up to 10 minutes to parse in Mental Ray; its driving me insane!!! What is the cause of this? I tried to follow your advice Speye and selected the whole scene and went to info> mental ray but the tick boxes were greyed out and I couldn't click on anything other than adaptive; that really pleased me ;- ( I then selected objects seperately and went to render setup, turning on spatial (5 units) with pixels instead of system. I also took the maximum thingy (can't remember what it was) down from 10 to 7. I did this to all the nurbs in the scene but it made not the slightest difference.

Im really bugged about this. The irony is that I haven't even completed the scene yet as I have to put a boy in there aswell as a large building. AND I haven't begun to texture either; I wonder what other hidden goodies await me inside the black box of myth and wonder that is, THE MENTAL RAY RENDERER.

Sorry about that last bit

Knick Knackered

[devdoka]

Alright, the problem has shown itself:

The mistake was converting everything to polygons from NURBS. This is what's taking the extra parsing time as each an every vertex has to be explicitly written out to the mi file whether you render interactively or from the command line. NURBS will parse out significantly faster simply because there are fewer control points to write out (mental ray does the curve continuity stuff at render time). Consult the illustrations in "Programming with Mental ray" to see what I mean (or set themode to ASCII and parse out a simply mi file of 2 spheres - one polygonal, the other NURBS. Open the mi file in a text editor to see the results).

I also suggest that you find an alternative to trimming 40 holes in your rocketship. If you're attempting to recreate the 40's sci-fi look, then chances are that your windows will have lips around them to cover any trims. If that's the case, try using transparency mapping to create the hole and use the lip to cover the jaggy edges. The ship isn't going to bend to reveal these edges, is it? Simply make a UV map on the ship and use rendermap to plot the locations of the holes (make sure your texture has very good resolution - or break up into multiple maps scattered about the surface). If you create some discs and surface deform them to the rocketship (or just place them where the holes would be), you can use rendermap to plot the shadows of the discs onto the rocketship's body. The captured shadows can then be used as the transparency map.

sooooooo:

1) Leave everything as NURBS
2) Use transparency maps instead of trims.
3) Continue to use adaptive tessellation - it only works for surfaces (NURBS, curves, and Patches), not polygons. The boxes were greyed because you have to 'pick' which fields you plan on editting, then click the button at the top to allow editting to actually occur - a stupid 2 step process, but it's not too hard.
4) Decrease max subdivision to say 5, not 7. You can bump it up later if needed.

You should now see very dramatic improvements in speed. I encourage you to read "Programming mental ray" as it covers all the options available. Many of which are not available from within the SI3D interface, but only from the command line as flags. A few of them will save your butt. As I said earlier, you don't have to read each and every page about coding, but going over some of the chapters describing mental ray's internal guts can go a looooooong way in helping to optimize your scene.

This will at least get you started. When things get rolling, give me another shout and we'll see if there are other areas that can be tweaked such as shaders, anti-aliasing, and other tunable parameters.

Matt

---

Matt Lind
Animator / Technical Director
Softimage certified instructor:
Softimage|3D
Softimage|XSI
speye_21@hotmail.com

[noar]

ack...

woah.. i cant beleive my post blew up with such popularity! hehe i guess i struck a chord somewhere out there

hrm... i do realize that i failed to describe my situation well enuff.

i should have also mentioned that:

I'm running win2k
DUAL p3 800 with 512 ram (YIH!) -- great speeds --
Both the renderer and the preview have the exact same settings when im doing compare tests (using soft or MR it makes no diff.....) preview is ALWAYS faster than render. so what's up with that?
My scene has a few nurbs spheres
and LOTS and i mean LOTS
of meta balls.

does this 'new information' shed any light on the problem?

oh, and BTW, as a treat, here's the scene im working on... it's a 250k divX animation (still in testing)
http://members.home.net/cameostudios/nametag.avi

[devdoka]

@My scene has a few nurbs spheres and LOTS and i mean LOTS of meta balls. Does this 'new information' shed any light on the problem?

Yep, actually it does.

Metaclay has to be recomputed every frame and will kill you if your not careful. Just activate shader mode in your viewports and start doing things and you'll see what I mean (outside of just camera moves which is cached).

Back when I was in school in the days before Nurbs, particles, and mental ray, there weren't any tools to blend surfaces together seamlessly in SI and I resorted to metaclay to build the entire upper torso of my character.....big mistake. Although the end result looked nice, it was dog slow in the renderer and displayed many artifacts from tesselation. SI changed the behaviour of metaclay since then a little bit, but not much. You shouldn't be getting the artifacts anymore, but the tessellation times will kill you as you've already found out.

The best advice I can give you is to use metaclay sparingly and adjust the parameters to the least quality possible that doesn't degrade your renders. Since most of your objects appear to be just spheres, you may want to work on a scheme to substitute the metaclay balls with Nurbs spheres on the frames in which they don't blend. Metaclay isn't computed for a ball if it's hidden and a Nurbs sphere will tessellate significantly faster. This will speed things up a bit - maybe a lot.

As far as mental ray is concerned - as I said in a previous post, Nurbs will tessellate significantly faster than polygons. Since metaclay is converted to polygons at parse time, it makes sense to substitute whenever possible. Another speed trick may be to reduce metaclay smoothness until it's start getting a little jaggy and then crank up mental ray's anit-aliasing via post filters such as the gaussian at 3x3. The last resort would be to increase the scene size by scaling everything up by a factor of 10 so you don't have to run the tessellation into the ground as you can use a lesser metaclay rendering quality to get the same result.

You could also download the speedrender shaders from imagica japan which includes a mental ray shader pack for faster rendering of metaclay.

Hope this helps,

Matt

---

Matt Lind
Animator / Technical Director
Softimage certified instructor:
Softimage|3D
Softimage|XSI
speye_21@hotmail.com

[hypersim]

I've managed to dramatically decrease rendering time for now by deleting some 300,000 polygons worth of Nurb objects. I have found, however, a strange anomally which occurs when using layered fog. Apart from getting [censored] results, whenever I use Layer Fog, even if I take all the values down and just have the Layer Fog tick box marked, the Parsing of the scene becomes QUICKER! It actually halfs the parsing time and for the life of me I don't know why. Can you shed some light on this?

So you think I should keep my objects as Nurbs then? I'll try that, Speye, Cheers!!

Knick Knack

[janina-overley]

Hi Speye

I've been reading this discussion with great interest, and I was hoping you might be able to help with two Mental Ray problems I've had for a while - the first involves Displacement mapping, which you've already touched on :-

I've created a number of scenes which involve Displacement mapping our company's logo. Rendered this looks lovely but the pre-render parsing stage takes upwards of 6 minutes for every frame. Given that in many of these scenes, the camera and logo do not move or change in any way, is it possible to get Mental Ray to remember the parsing it has done from one frame to the next?

My second problem is with Mental Ray motion blur - I can't seem to get it to blur transparency maps.

Thanks!!
snowctrl

[devdoka]

@I've managed to dramatically decrease rendering time for now by deleting some 300,000 polygons worth of Nurb objects

OK, a quick lesson with NURBS. When using parametric surfaces the idea is to use the fewest number of curves possible to represent the surface details. If you need the surface to be smooth, use the step values from the tessellation controls to do this - DO NOT ADD CURVES!

Demonstration (Start with a fresh session of SI3D):

1) Get > Primitive > Sphere. Bspline with default settings.
2) Matter > Preview > Setup. Set renderer to mental ray and activate anti-aliasing. min samples=0, max samples=2
3) Matter > Preview. Notice the sphere looks fairly smooth, but if you look verrrrry closely, you'll see facets at the edges.
4) Info > selection. Take note of the triangle count, number of curves in U and V (8 x 11) and the step counts (3 x 3).

5) Hide the sphere.

6) Get > Primitive > Sphere. Set steps in U and V to 4 and 3 respectively.

7) Matter > Preview. Notice the sphere looks quite faceted.
8) Info > Selection. Click render setup (same as Info > Mental ray). Change tessellation to "adaptive" with the curvature setting. Set units to pixel length and activate subdivision limits with min=1, max = 5.

9) Matter > Preview. Notice the sphere looks ultra smooth again.

10) Info > selection. Take a look at the triangle count - quite a bit less than the first sphere (1152 vs 216).

When working with parametric surfaces, you have many options. This demonstration illustrates how you can create very simple geometry but still get highly polished results. This means you can work interactively with low polygon counts and get work done quickly and let the renderer do the work - not you.

When you build your rocketships and stuff, you should follow the same principles. If the rocketship is a torpedo like shape, then you should only need a cross section revolved at a 90 degree interval and possible 5 or 10 points max on the curve (I haven't seen your rocketship, so I'm guessing).

There should be no need for 300K of polygons from NURBS objects as you can see. However, you will have to be more conscious of your adaptive tessellation settings to keep render times from getting out of hand. Placing subdivision limits is the easiest first-step prevention, but it's not bulletproof. When performing test renders, place the object in view as it will be seen in the final render and adjust it's settings off of that. Tweaking the chord length and angle values can go a long ways to further optimize your scene and keep triangle counts down while not displaying any artifacts. Details on these parameters are in the rendering guide as part of your manuals.

As far as the layer fog issue is concerned, I'm not quite sure what to make of that. I'm guessing the parser looks at the scene and optimizes the output according to what the camera would see. That is, if your layer fog is opaque, for example, then the parser wouldn't export anything beyond the max visibility distance defined in your fog. Just a guess.

Matt

---

Matt Lind
Animator / Technical Director
Softimage certified instructor:
Softimage|3D
Softimage|XSI
speye_21@hotmail.com