Daz Iterations & Denoiser Question

[Deleted member 4707538]

What is a good number of iterations for standard VN scenes when rendering at 1440p followed by postwork then downscaling to 1080?
I've seen some recommendations for a low number of iterations (100) and then use Intel denoiser. But I don't want to introduce any blurriness to the image.
Also seen recommendations of around 5,000 - 8,000 iterations. Not sure if using a denoiser is recommended in these cases.
So looking for recommendations on a "rule of thumb" for number of iterations, and use cases for a denoiser.

 

[MissFortune]

I've seen some recommendations for a low number of iterations (100) and then use Intel denoiser. But I don't want to introduce any blurriness to the image.

You won't get blurriness, just loss of detail in the skin/hair/clothes/etc. or anything that has some visible texture to it. You won't lose it entirely, just a quite a bit of it - depending on how much noise their is.

What is a good number of iterations for standard VN scenes when rendering at 1440p followed by postwork then downscaling to 1080?
Also seen recommendations of around 5,000 - 8,000 iterations. Not sure if using a denoiser is recommended in these cases.
So looking for recommendations on a "rule of thumb" for number of iterations, and use cases for a denoiser.

There is no rule of thumb. It's just whatever works for you and the quality you're looking for. Along with how optimized your lighting, textures, reflections, and so forth are. Along with complexity and amounts of figures. Then, of course, there's hardware.

These are my settings for most 4K renders:



For reference, most of them are done at around 3000 iterations (if not earlier, depending on the simplicity.), but I almost always never let it finish unless I'm busy doing other things. I usually stop at around 3500/4000 samples as if it's still noisy by then, then it'll probably be noisy at 5 or 6 thousand. Rendering Quality should always be off. 4K has a bit more leeway with noise as downscaling it condenses that noise even further.

At 1440/2K, my settings are about the same outside of max samples. 2500 should be all that you really need. If it's still noisy after that, then drop it into an external denoiser. A lot of this is dependent on hardware, though. I'm on a 3080, so it's going to be a bit faster for me than most.

As for a denoiser? The variance in noise on a render-to-render basis makes it tough to give specific use-cases for when to use a denoiser. There're stages, I guess. Needs it, needs it in places, and doesn't really need it.

That said, a rule of thumb for denoising renders/images is to use an external denoiser (like the Intel one) and the original 'noisy' render.

So, I'll start with a noisier image that needs denoising:



In my opinion, this one needs to be denoised. Outside of the figure closest to the left, most of it looks normal, but as soon as you start zooming in you see noise in both the skin and the bokeh(s) from the depth of field. But some of it still looks fine, especially since I'd be downscaling to 1080. So, what I'd do first is bring the noisy layer in and then put the denoised version on top.

So, now you have both renders in your editor of choice. What you'd do is mask away parts of the denoised version of the render that didn't need to be denoised. So, in the above example, the figure in the back right side of the render really didn't need any denoising on her upper chest. So, I'd mask away that part of the denoised render, revealing the older(/noiser) version of the render underneath it. This way you can keep the skin detail in the parts of the render that didn't need the denoising.



This would be a "needs it in places" scenario. The skin, hair, necklace, and shirt are fine for the most part. But if you look at the walls of around the windows, you see some noise. So, like the above, I'd bring the noisy version in first and then the denoised version on top of it. Then I'd just mask out the figure on the denoised version.



With this one, you can zoom in pretty close and find the noise (and the jacked texture of the cloth over his head), but it doesn't stick out as noticeable at a reasonable viewing size. So, you could safely not denoise a render like this. Especially if you're scaling down from a lower size.

So, again, it's almost always going to be a case-by-case basis on denoising. Even worse, is that your average consumer likely won't notice a massive difference unless it's sort of in their faces. Just gotta decide for yourself where it's necessary and where it isn't. Hopefully cleared up some of the

 

[Saki_Sliz]

In my experience, the number of iterations is mostly dependent on the lighting.
most of the noise comes from soft lighting, where surfaces are light indirectly by light bounces, which do not always reach a light source before hitting their max bounce count (movies like big hero six which went with an art style of very soft and colorful spotlighting and using environment lighting went as high as 24 bounces per ray, but the most I ever see get used is 16 bounces, I usually don't need more than 3 or 4 to be fast but still allow for soft indirect lighting). To minimize noise is to ensure a scene has a very simple light set up (in terms of how light reaches the subject in the least amount of bounces). IE avoid making rooms softly light by a window or only light by a door cracked open at night. Try to ensure that there is a good source of light, and take advantage of ghost lights if need be. The more a surface is directly light by light, the more often the light rays find a light source thus a less noisy image, thus less iterations needed for denoising.

For misfortune's images, the first one (I assume a mock up or WIP scene) I would say the scene is purely soft light by room lighting and would be a long render at high iterations, I would also say its not dynamic enough/high contrast. For speed and style control I would fake a soft light effect using ghost lights. In order to make a light source look soft, you make the light source physically big, as if its light was coming from a large soft box. typically A point or spot light which you have to change the shape from point to sphere and then adjust the diameter (also helps to turn the emitter off so the light source itself is invisible, because sometimes I make lights 10 to 15 feet big for really soft light, keeping it 20 feet away from the subject if the camera is focusing just on them). Similarly, for the second image, it shows a good example of having a good main light from what I guess is the street light. I would want there to be less black in the image, make it feel colder, by making another ghost light, this time a direction light, and set the angle of the light (direction light doesn't have a size but has a sort of spread angle) between 15 to 60 degrees but still make it point down from the sky. I often also do this same effect but pointing up to emulate light bouncing from the ground without relying on light bounces for this. But that's if you want soft lighting. If you don't want soft lighting you probably don't need this, but if you are trying to make a scene look less dark and less 3D add soft lighting using large ghost lights so that you aren't relying on noisy light bounces so you can keep the render iterations down. With this technique I go as low as 40 iterations for speedy character renders (tho this is with blender and not so much daz)

 

[MidnightArrow]

I've seen some recommendations for a low number of iterations (100) and then use Intel denoiser. But I don't want to introduce any blurriness to the image.

The Intel denoiser needs an albedo pass, which Daz Studio doesn't create. (Some poor SOBs on the Daz forums are seriously trying to fake it with 3delight.) Without an albedo and normal pass it's not going to be much advantage over Daz Studio's built-in denoiser.

 

[Deleted member 4707538]

You won't get blurriness, just loss of detail in the skin/hair/clothes/etc. or anything that has some visible texture to it. You won't lose it entirely, just a quite a bit of it - depending on how much noise their is.

There is no rule of thumb. It's just whatever works for you and the quality you're looking for. Along with how optimized your lighting, textures, reflections, and so forth are. Along with complexity and amounts of figures. Then, of course, there's hardware.
.
.
.
So, again, it's almost always going to be a case-by-case basis on denoising. Even worse, is that your average consumer likely won't notice a massive difference unless it's sort of in their faces. Just gotta decide for yourself where it's necessary and where it isn't.

Thanks for your super helpful and thorough reply, much appreciated!

 

[MissFortune]

The Intel denoiser needs an albedo pass, which Daz Studio doesn't create. (Some poor SOBs on the Daz forums are seriously trying to fake it with 3delight.) Without an albedo and normal pass it's not going to be much advantage over Daz Studio's built-in denoiser.

But it's still going to put out a separate, denoised version of the render, which you don't get with Daz. Which allows you to mitigate the loss of detail via denoising by mixing both. So, even if only for that separate version, the Intel/Nvidia denoisers are inherently better.

 

[MidnightArrow]

But it's still going to put out a separate, denoised version of the render, which you don't get with Daz. Which allows you to mitigate the loss of detail via denoising by mixing both. So, even if only for that separate version, the Intel/Nvidia denoisers are inherently better.

True.

But my point is more that OP should temper their expectations about denoising since Daz Studio half-asses it. Other, better programs do it properly and the results are much nicer without even needing to Photoshop anything.

 

[mickydoo]

Its also subjective to the scene and/or character. I ditched these one set of rooms cos even after 20 gazillion interations there was still noise on the character, any of them I tried. I have this character I use now that is a shit to render indoors close up with no denoiser (I suspect the hair) but put her outside and no issue.

 

[coffeeaddicted]

In my experience, the number of iterations is mostly dependent on the lighting.
most of the noise comes from soft lighting, where surfaces are light indirectly by light bounces, which do not always reach a light source before hitting their max bounce count (movies like big hero six which went with an art style of very soft and colorful spotlighting and using environment lighting went as high as 24 bounces per ray, but the most I ever see get used is 16 bounces, I usually don't need more than 3 or 4 to be fast but still allow for soft indirect lighting). To minimize noise is to ensure a scene has a very simple light set up (in terms of how light reaches the subject in the least amount of bounces). IE avoid making rooms softly light by a window or only light by a door cracked open at night. Try to ensure that there is a good source of light, and take advantage of ghost lights if need be. The more a surface is directly light by light, the more often the light rays find a light source thus a less noisy image, thus less iterations needed for denoising.

I always ponder.

I did a test with an office asset. I think it's FG office and it has mesh lights i believe. So when you load it, it already has lights all over.
Now, i added a sphere outside to act a the sun from the left.
Then i tried to switch off all lights and instead light the room with primitives. To be honest, i did not see any difference in speed. It took the same amount of time.
If i run it with 5000 iterations, that will be 1,15h or so.
Most of the time the problem i see, is the fact that when you over-light (like some atomic ray) it will render pretty fast but you lose all the nuances in the render. Well, i suppose you can post work on that.
For me, i like to have it done right there.
So if you want to have a more ambient lighting to get more contrast in the scene it will take longer. At least in my experience.

This is the scene, in case someone is interested.

 

[coffeeaddicted]

Rendering Quality should always be off. 4K has a bit more leeway with noise as downscaling it condenses that noise even further.

I need to ask my dumb question. Why? Because i don't know.

 

[MissFortune]

I need to ask my dumb question. Why? Because i don't know.

No question is dumb (for the most part), as long as you're learning from it. That said, I can't explain it very well. So, from the

You must be registered to see the links

:

Render Quality Enable:
This is a simple toggle on or off setting. Turns on or off the render quality parameters. When off the software will rely solely on the max samples and max time settings to complete an image. Otherwise the software will consider Render Quality as the primary parameter for determining a complete image. (At least that is my understanding)

Render Quality:
This one is sort of a mixed bag of information so bear with me.
As defined by the Daz documentation render quality is simply that. At a lower value, the software will render the equivalent of something you would see in a video game (OpenGL). At higher values, the software emphasizes the image quality over the time it takes to render. OpenGL is a very fast render type, so anything beyond the first setting of 1, does not use this. The render quality setting has a maximum value of 4. However, you can change this number to a crazy value such as one thousand. To my knowledge (and experiences) a value higher than 4 has no bearing on the render quality.

As defined by some general information I have dug up, render quality overrides the max time and max samples settings (I have not experienced this). The setting increases render time linearly each time the number is increased. Setting it to two doubles what it would take at one and so on and so forth.

Further defined by information found elsewhere, this setting determines how close any pixel has to be to its prior sampled value before it is considered “converged.” The higher this value, the closer in value any pixel must be to its prior value before it is considered “converged.” In other words, if a pixel in the center was sampled at being “Blood red” (simplifying this by the way), then the next x number of samples must also come up with “Blood red” to consider that pixel converged.

Personally, I use render quality and set it to 4 for final product images. Results may vary, but I have seen no issues. In my experience the Render Quality setting will render the image until either the rendering converged ratio is met, or the maximum samples or max time are met.

I've always had middling results with it, even with recommended settings and just straight up experimenting with it. I find it far easier to just set up a regular/base sample number with the max time off and just let it render until it looks good enough to stop.

 

[coffeeaddicted]

No question is dumb (for the most part), as long as you're learning from it. That said, I can't explain it very well. So, from the

You must be registered to see the links

:



I've always had middling results with it, even with recommended settings and just straight up experimenting with it. I find it far easier to just set up a regular/base sample number with the max time off and just let it render until it looks good enough to stop.

Well, i have to try.
Clearly that person put quality to 4. This must take a long time.
I think i tried it before without quality but i think the time saving wasn't there. Even though that isn't the main concern i have to run it again to see if the quality is lesser than with it.
Since i am using glasses it may not be apparent to me really.

 

[coffeeaddicted]

No question is dumb (for the most part), as long as you're learning from it. That said, I can't explain it very well. So, from the

You must be registered to see the links

:



I've always had middling results with it, even with recommended settings and just straight up experimenting with it. I find it far easier to just set up a regular/base sample number with the max time off and just let it render until it looks good enough to stop.

Ok, so i tried it.
Time elapsed around 40 minutes, which is of course faster. About half the time.

Now for this scene figure sits in a room with a window. Sun shines through the window on midday.
Nothing special.


rendered with 0 quality and 2500 iteration


after intel denoiser used.

There is of course a difference when a denoiser is used.
I mean i take it and i think no one will really know how the render was done anyway. As long it looks good.

Thanks

 

[Deleted member 1121028]

No question is dumb (for the most part), as long as you're learning from it. That said, I can't explain it very well. So, from the

You must be registered to see the links

:



I've always had middling results with it, even with recommended settings and just straight up experimenting with it. I find it far easier to just set up a regular/base sample number with the max time off and just let it render until it looks good enough to stop.

From my memories I don't think it works the way Daz user explains it. In absence of solid refs most of it is guess work, so put it right in the bin lol.

But why would pixel thresold (defined in render quality) shoud be set with openGL? That's my main grip, more likely use the first batch of Iray iteration ("canvas was written"). As render progress and more pixels are flagged completed, render will stop when XX% converged pixels are reached. IIRC, render quality is also near perfect linear, (so pixel thresold/render quality 1 = 30 min, 2 = ~1 hour).

"render quality overrides the max time and max samples settings", is not true but I may be wrong, far I remember. I would not shit on Render quality as it's quite an elegant approach for an impossible task (when my tracer render is finished on average?) and also a great benchmark. AI should push it further tho.

 

[noping123]

From my memories I don't think it works the way Daz user explains it. In absence of solid refs most of it is guess work, so put it right in the bin lol.

But why would pixel thresold (defined in render quality) shoud be set with openGL? That's my main grip, more likely use the first batch of Iray iteration ("canvas was written"). As render progress and more pixels are flagged completed, render will stop when XX% converged pixels are reached. IIRC, render quality is also near perfect linear, (so pixel thresold/render quality 1 = 30 min, 2 = ~1 hour).

"render quality overrides the max time and max samples settings", is not true but I may be wrong, far I remember. I would not shit on Render quality as it's quite an elegant approach for an impossible task (when my tracer render is finished on average?) and also a great benchmark. AI should push it further tho.

Just a note, since I *DO* use render quality settings:

yea, the render quality setting (1,2,3etc) is infact almost perfectly linear.

So, taking from that link, this is effectively how render quality works:

Further defined by information found elsewhere, this setting determines how close any pixel has to be to its prior sampled value before it is considered “converged.” The higher this value, the closer in value any pixel must be to its prior value before it is considered “converged.” In other words, if a pixel in the center was sampled at being “Blood red” (simplifying this by the way), then the next x number of samples must also come up with “Blood red” to consider that pixel converged.

There's a bit more that goes into it, but for the basics, it mostly works like that. Raising the render quality from 1->2, effectively doubles the number of samples necessary to consider it converged.


And you are correct - render quality does *NOT* override max time/samples setting. If it reaches the max time, or max samples, before reaching your target convergence, the render ends. If it reaches the max convergence early, the render ends. It's an and/or, not just an or.

IMO there really isn't any good reason to NOT use render quality setting. At some point, additional iterations won't accomplish a damn thing - if it reaches 99% convergence after 500 iterations, an additional 4500 won't actually do much of anything, but otoh if it gets to 5000 iterations and is only 36% converged, well.... then the render just ends anyway.

 

[coffeeaddicted]

It seems you have to set time and samples to max to get the convergence goal.
I will see how it will look. Man, i love playing with this software.

 

[Deleted member 1121028]

Just a note, since I *DO* use render quality settings:

yea, the render quality setting (1,2,3etc) is infact almost perfectly linear.

So, taking from that link, this is effectively how render quality works:



There's a bit more that goes into it, but for the basics, it mostly works like that. Raising the render quality from 1->2, effectively doubles the number of samples necessary to consider it converged.


And you are correct - render quality does *NOT* override max time/samples setting. If it reaches the max time, or max samples, before reaching your target convergence, the render ends. If it reaches the max convergence early, the render ends. It's an and/or, not just an or.

IMO there really isn't any good reason to NOT use render quality setting. At some point, additional iterations won't accomplish a damn thing - if it reaches 99% convergence after 500 iterations, an additional 4500 won't actually do much of anything, but otoh if it gets to 5000 iterations and is only 36% converged, well.... then the render just ends anyway.

I did also use quality setting a lot. Imho chasing pixels convergence is also really great tool to learn how to go fast (how my lightning shoud be done to be effective, what surface/shader layers are problematics, using DoF to your advantage instead of being detrimental, use of Iray section planes and so on).

I think problem is back in the day people misused % px convergence settings for "I will remove my hard noise from my terrible render if I push it hard". From the same logic "Surely 15000 more iterations will make that render better". There is no magic settings in any software that will change a bad built scene into a good one (not even the overuse of crappy denoisers).

 

[coffeeaddicted]

What i can say is this.
Convergence works great. And, the renders don't take longer in my case. Usually around 30 minutes.
I think this is the way to go, at least for me.
Not denoised

 

[Turning Tricks]

...
IMO there really isn't any good reason to NOT use render quality setting.
...

I'll give you two...

1. I've had several cases where I reached 100% convergence on a relatively simple scene, but there was still heavy noise in the shadowed areas; and
2. When running with Render Quality ON, you are stuck with that 100% convergence. There is no 101, 102 or 100-whatever percent convergence. So, if after rendering for 2 hours, you see that you will reach 100% but the image still has noise in it, you are sorta screwed. Whereas, with Render Quality OFF, you can easily increase either samples or time, mid-render, to allow it more time to get better.

In either case, generally speaking, you tend to know how many samples something is going to take. If for no other reason than you ran a quick test render already.

Using set values for time and sample with Quality OFF also makes it a lot easier to plan batches of over-night renders, so you know they will be done by the time you wake up.

 

[coffeeaddicted]

I'll give you two...

1. I've had several cases where I reached 100% convergence on a relatively simple scene, but there was still heavy noise in the shadowed areas; and
2. When running with Render Quality ON, you are stuck with that 100% convergence. There is no 101, 102 or 100-whatever percent convergence. So, if after rendering for 2 hours, you see that you will reach 100% but the image still has noise in it, you are sorta screwed. Whereas, with Render Quality OFF, you can easily increase either samples or time, mid-render, to allow it more time to get better.

In either case, generally speaking, you tend to know how many samples something is going to take. If for no other reason than you ran a quick test render already.

Using set values for time and sample with Quality OFF also makes it a lot easier to plan batches of over-night renders, so you know they will be done by the time you wake up.

Can't really confirm this.
Reason, i haven't had a case like that.
Though convergence seems to work for me, there are some indoor assets that could take more time.
But i switched to spotlights as it speed things up.

I do go for 5000 samples even with convergence on and if there is really some grain, i will use denoiser. I think i am not too concerned about some loss that may or may not anyone see.