Adaptive image sampler (vray 3.0)

[thomaskc]

Hello everyone,

I would love to hear from you all, basically trying to get my head around the min/max values of the adaptive image sampler.
Every time I think I understand the workings behind it, I speak to someone new, and get a totally new perspective on it which confuses me once again.

So back in the day Vray 1.5 and earlier, I think it was quite common to use values like 2 min / 8 max for production...

During the reign of Vray 2.0+ guides and tests have stated that a 1 min / 100 max is the way to go, gives you good results etc...

So the other week I went to a Vray 3.0 promo event with chaosgroup and to my surprise (I think it was Vassil ? sorry forgot your name), he said that people do a lot of different things, but he typically uses 2min / 8 max for production... *mind blown* haha

I'm sure there is no better or worse but I am curious to hear what you guys are using for a "final" render while using the adaptive image sampler, and why? I have tested a few things but haven't come to any conclusion to which approach makes more sense to use.

Thanks in advance!

/Thomas

[joconnell]

Right!

The thing is that there's no one set of settings that work really quickly and to high quality for every scene ever as every scene is totally different and thus will have different needs in terms of your lighting, materials, motion blur, depth of field and so on.

One of the biggest factors is if you're rendering with 3d motion blur or 3d dof. Both of these effects are about sampling the "position" or the "geometry" of the object, in the case of motion blur vray is trying to sample the position of an object, just at multiple positions over time. With depth of field, it renders the scene from multiple angles / viewpoints (it's a bit like the camera is on a path constraint tied to a circle) so it's doing the same "position" sampling of an object at different angles / locations. Both of these position / geometry type of sampling factors are both sorted out by your max AA sampling. There used to be motion blur and depth of field subdivs controls in vray previous to version 3 but they do the exact same thing as AA controls - fire more geometry / position samples into your scene so it was kind of adding in two more controls that might not have been needed. Bottom line, if you've got motion blur or depth of field you'll need high max AA to smooth it out so you're looking at a 1 / 50 or 1 / 100 type of situation.

Another situation that you need high AA in is if you've got very fine, thin detail. Depending on what industry you work in that could be blades of grass, hair, wire fences or the likes. Again if you've got small geometry there's a chance that when vray is firing AA samples it might miss, so you end up using way more AA rays to make sure your tiny details get detected sufficiently.

For the 2 / 8 example, if you're NOT using motion blur or depth of field and don't have really fine detail in your scene then the big factor might be lights and materials. If you're using really big soft area lights, they tend to get a little bit grainy at the soft edge of the shadows. This is an issue with light sampling - it's almost a little bit like the AA issue with fine detail except rather than geometry / position samples being fired from your camera into the scene, the geometry / position samples are being fired from the direction of your light. If you don't have enough of those, it might not be able to accurately detect and define the outline of objects that it's casting shadows with. This'll mean that the shape of the shadow will look more dotty and random or grain as we see it visually. Likewise if you've got a material that has very soft reflections on it, instead of our camera firing geometry rays, or our light firing shadow rays, this time it's the material firing out reflection rays from the surface of your object into the scene. Same thing as shadows and geometry, vray is firing out samples in semi random directions and taking back a sample of the object that the reflection sees. If there aren't enough samples as usual, you'll get a really dotty result since there aren't enough "hits" to really clearly define the object being reflected. The long and short of it is that if you don't have enough material samples, you'll get grainy looking reflections and a bad result. If you've got a simple scene with no motion blur, and no fine geometry but you're using complex lighting or materials with loads of glossy stuff, then you can use less AA (geometry / position sampling) and instead up your light or material samples to clean everything up.

Now we get to vray 3!

So in vray 3, and in general, Vlado's thought (quite rightly) is that human time is more important than machine render time. With the universal settings (1/100 and so on) everything was left to the anti aliasing controls to clean up the entire scene. The main benefit of this is that you don't have to go around your scene tweaking individual material or light settings, everything is handled in one single location. It isn't as quick as going in and optimizing your scene manually each time, but if you can get 5 shots on to the render farm in a day using the convenient universal settings, that might be far better than tweaking your scenes and only getting 3 shots on to the farm in a day. If they're rendering overnight then it might not make that much difference.

I personally at the minute use a mix of settings for doing vfx work and most of it is determined by whether I'm using 3d motion blur or not. If I am, it's generally about 1/50 and a noise threshold about 0.01 and then material samples will be about default 8. If I'm doing something that doesn't need 3d blur I'm normally around 1/12 for fine detail and then my material settings for the likes of reflections range between 40 up to 240. This is a guide that works for me but a thing of note is my scenes typically don't have a lot of lights and they don't have a lot of materials to manage either, so if I want to go around my scene optimizing renders, it would take a lot less time than someone doing a huge architectural scene with a lot of lights and different materials.

Unfortunately there is no rule for what your scene needs, they're all unique and beautiful - Give Vlado's approach of using min max of around 1 / 50 or 1 / 100 and then gradually raining the min shading rate (this is only in vray 3 by the way) to clean up your materials and lights. If it works nicely for you, great! It's a quick approach for most scenes in terms of getting a shot out.

[RockinAkin]

Great write-up John!

Thomas, some additional info that might help you out if you haven't already read it:
http://www.cggallery.com/tutorials/vray_optimization/
http://www.interstation3d.com/tutori...yfing_dmc.html

[Neilg]

During the reign of Vray 2.0+ guides and tests have stated that a 1 min / 100 max is the way to go, gives you good results etc...

for what it's worth, I always thought this method was a massive waste of time and you shouldn't use it as a reference of the right way to do things. it was a lazy way that gave renders with as little work as possible optimizing.

I've never changed the way i render scenes, I dont really understand why it's so complicated. make a scene with default samples everywhere. if there's noise, figure out where it's coming from and raise the subdivisions.

[vlado]

for what it's worth, I always thought this method was a massive waste of time and you shouldn't use it as a reference of the right way to do things. it was a lazy way that gave renders with as little work as possible optimizing.

It depends on what you are rendering I suppose. I had a case where in an animation, the user had done all the set up for a frame that had relatively little motion blur and only a certain part of the scene was visible. Only later on did they find out that the settings didn't work well for other parts of the animation where they had more blur. In this case, a more universal-like approach ensured that all frames are sampled adequately and ultimately reduced the render times.

Another similar approach involved texture baking for a scene with many objects. The users had done the set up for a few objects that had relatively large flat areas. However for other objects, that were much more detailed, the render times exploded. Again, switching to more universal-like settings ensured that all objects are adequately sampled.

Yet another case involved a scene where the client initially only had a building with relatively large flat areas, and they had set up their materials/lights according to that. However, later on it was necessary to add some trees with lots of detail, which caused the render times to balloon. Again turning to a universal-like approach ultimately gave better results for less render time and simpler setup.

Best regards,
Vlado

[thomaskc]

Wow!

Thanks guys! All of you, thank you for taking the time.

It all makes a lot of sense, and looking at the amount of time we bloody waste on trying to tweak and "solve", often I think it's easier to just drop in a number I know that works and leave the computers to it.
I will keep playing around, so thank you all again!

[Neilg]

It depends on what you are rendering I suppose....

this is true - I was a bit strong with my wording
i'm happy that this method is much faster with vray 3 - it means everyone can produce good clean renders right from the start without worrying too much, and given a couple years in cpu tech the time difference between it and a carefully optimized scene will become irrelevant.

[alexyork]

Wow!

Thanks guys! All of you, thank you for taking the time.

It all makes a lot of sense, and looking at the amount of time we bloody waste on trying to tweak and "solve", often I think it's easier to just drop in a number I know that works and leave the computers to it.
I will keep playing around, so thank you all again!

This is almost always my method as well. If I know I can afford overnight to render finals then I'd rather spend more time working on the other scenes or Photoshopping than spending hours tweaking the render settings for basically 0 net gain.

[jasonhuang1115]

Love and learned from the write-up. Thank you for sharing, John.

[bennyboy]

This is almost always my method as well. If I know I can afford overnight to render finals then I'd rather spend more time working on the other scenes or Photoshopping than spending hours tweaking the render settings for basically 0 net gain.

Sometimes the gains are huge when you optimize, sometimes it's like you say. I've found it to be rather frustrating to be upping my subdivisions and getting no speed gains. Gotta love that feeling though, when you increase the subdivs on something and it's like letting a race horse out of the gate...it's a counterintuitive miracle every time it happens to me. My difficulty is understanding more why and when...love the explanations above.
BTW, today I was rendering this car for an ad; using IBL mostly; test renders were really noisy; time for final - I started optimizing - everything was still pretty noisy no matter what I did. Had the Dome light and GI up to like 700, still too much noise. I was in a hurry with the deadline, so I just scrapped it and used Vlado's -- saved my butt!!! It took 1.5 hours, but was near perfection - and I got to have lunch with my kids. A good example of why the 1/100 method will continue to be a favorite. Human time is indeed way more valuable than machine time. But it's a two edged sword, cause sometimes that optimization will save your butt too! I have experienced where the test renderings took longer than my optimized renderings!
Rule of thumb: if you've spent half an hour on optimization without seeing major improvements, switch to 1/100, and bring the noise down to around .003-2 - then go live your life and let the machine be a machine.

[fewlo]

and I got to have lunch with my kids

If thats not THE reason to use the 1/100 method, I don't know what is.

[Neilg]

Rule of thumb: if you've spent half an hour on optimization without seeing major improvements, switch to 1/100, and bring the noise down to around .003-2 - then go live your life and let the machine be a machine.

I did that on the dusk image here to test - http://forums.chaosgroup.com/showthr...seum-courtyard
and my render time went from a few hours to getting cancelled by me at around 20hours in, 2/3rds done.
Do you live in a world without deadlines?

[fewlo]

and my render time went from a few hours to getting cancelled by me at around 20hours in, 2/3rds done.

If thats not THE reason to not use the 1/100 method, I don't know what is.

[vlado]

If thats not THE reason to not use the 1/100 method, I don't know what is.

It could also mean that there is a problem somewhere in the code. I would very much like to take a look at that scene if possible.

Best regards,
Vlado

[bennyboy]

But it's a two edged sword, cause sometimes that optimization will save your butt too!

Uhemm...look, they BOTH can save you time, it just depends on the scene and the amount of time you can spend on it...theoretically, optimization should save you time, every time...but that's just a theory...a theory based on some pretty thick knowledge (mixed with the inevitable trial and error) that we don't all always have time for because of........................(wait for it)........................DEADLINES.