Max subdivs not matching the amount of passes

[LudvikKoutny]

I've just discovered that setting Max. subdivs to 10 and setting noise threshold to 0.0 does not make rendering stop after 100 passes. I don't know what's going on under the hood, but it makes Max. subdivs value quite unpredictable.

Setting Max. subdivs to just 4 made render reach 95 passes

[LudvikKoutny]

Some more tests:

Max subdivs 5: 120 passes
Max subdivs 10: 368 passes

I am really curious where are the additional passes coming from.

[^Lele^]

That is entirely normal, and nothing you should worry about, engine-wise.
V-Ray knows best (hint: adaptive routines.), yet if there's any worry is why it ought to try and sample whatever setup you fed it that much.
Send a scene, and we'll know.

[LudvikKoutny]

It's not a scene I can share, but I sent it to Vlado earlier today, as there were some performance issues as well.

Seems like I am in a scenario where it pays off more to ignore noise limit and just render a limited amount of passes. But I can't really define the amount of passes I want, as Max. Subdivs value behaves quite unpredictably.

[vlado]

That's normal. The subdivs are per pixel. During rendering, some pixels might be deactivated and not sampled for a while, but then a change in a nearby pixel might cause them to get reactivated, and their sampling continues. They in turn may cause other nearby pixels to be reactivated and so on. This ensures that the final image quality is stable and predictable.

Best regards,
Vlado

[LudvikKoutny]

Alright then. It just makes hard to predict the rendertimes, especially since counter gets deactivated once noise threshold is at 0. Guess I will have to find some workaround.

[vlado]

What are you trying to do?

Best regards,
Vlado

[LudvikKoutny]

Just render for fixed amount of passes, so that noise is consistent across the frames, but it's not noise threshold what stops the rendering. It's regarding that scene I sent you, which renders too slowly. I am chased by deadline so I am trying all sorts of crazy stuff

[vlado]

Ok, I will give you more details a bit later tonight.

Best regards,
Vlado

[^Lele^]

That's what the fixed sampler is for: exact sample counts.
Anything else, is designed with optimal sampling in mind, and a "pass" does *not* equate to more samples everywhere, nor are those samples necessarily added to pixels which already reached the Max (although some spillage is necessary.).
Fair warning, same sampling across frames does *not* equate to same noise across the frames (or who'd have cared of a noise threshold as target in the first place.).

[LudvikKoutny]

Thanks Vlado!

Lele: Yes, I thought about fixed sampler, but I meant something more similar to what Corona does by default, it renders adaptively (which fixed sampler afaik does not) but at the same time render is stopped by simple pass limit, which progresses linearly. So you get predictable linear passes but samples in each pass are distributed according to some adaptive importance.

[^Lele^]

that is our old behaviour, which however had conceptual issues, fixed with the progressive noise threshold and adaptive sampler.
You can still have that behaviour back, of course, if you so wish, through a maxscript option.
I wouldn't advise that, however: your best bet is to set a generous max subdivs amount, and a high NT if you want a quick preview with the same noise level across frames (add a noiseLevel RE so you can actually measure it, and maybe infer rendertimes for the higher quality sequence to complete later).

[vlado]

but I meant something more similar to what Corona does by default, it renders adaptively (which fixed sampler afaik does not) but at the same time render is stopped by simple pass limit, which progresses linearly. So you get predictable linear passes but samples in each pass are distributed according to some adaptive importance.

This does not actually get you predictable results in terms of noise level in the image. There are quite a few reports on the Corona forum for animation frames coming out with different noise levels.

Best regards,
Vlado

[LudvikKoutny]

Well, I've done quite a few Corona animations with fixed passes count, and it never let me down. Most of the reports I have seen were due to the time limit set, and rendering on multiple machines with varying performance. It's really hard to get varying noise levels between frames if I render each frame with fixed count of AA rays and shading rays. Some other reports were due to render being stopped by adaptivity noise level rather than pass count.

Non the less, it's completely fine if V-Ray does not work the same way as long as there is valid reason for that. I am just simply trying everything I can to render my animation in some reasonable time.

I especially find it strange since I used to render out long animation of full CG forest scenes in V-Ray 2.5 in FullHD with 1hour-ish rendertimes, and that was usually around 0.02 noise threshold. I have no idea why is this particular scene giving me such a hard time.

[^Lele^]

Well, I've done quite a few Corona animations with fixed passes count, and it never let me down.

How did i not doubt that for one second, i wonder.

Most of the reports I have seen were due to the time limit set, and rendering on multiple machines with varying performance. It's really hard to get varying noise levels between frames if I render each frame with fixed count of AA rays and shading rays. Some other reports were due to render being stopped by adaptivity noise level rather than pass count.

How did you measure noise levels in the output?

Non the less, it's completely fine if V-Ray does not work the same way as long as there is valid reason for that. I am just simply trying everything I can to render my animation in some reasonable time.

I especially find it strange since I used to render out long animation of full CG forest scenes in V-Ray 2.5 in FullHD with 1hour-ish rendertimes, and that was usually around 0.02 noise threshold. I have no idea why is this particular scene giving me such a hard time.

Yes, i think a tour of the salient threads which explain the changes as of 3.3+ would help you, just as much as a tour of the Docs.
You can start from here, it's long, but worth reading:
http://forums.chaosgroup.com/showthr...image-with-SP3
Then there's the Podcast from 2 years ago, unfortunately with my voice in it, going into details, here:
https://labs.chaosgroup.com/index.ph...3-for-3ds-max/

Both, however, are longer winded way to say the same thing, which i suggested you a post ago: raise NT and render away. You have resumable rendering now, any other approach is looking for unnecessary trouble.