Render times/quality [Daz 3D]

[blissy]

Hey everyone,
As for my first post on the forum I'd like to ask a question concerning render times.
I've been trying my hand at some simple Daz 3D stuff like creating a scene & posing. Things have been going pretty good. It's an intuitive program.
But when it comes down to rendering, I start running into trouble as soon as I try adding a 'scene' or an 'environment' to my project.
Please see the following image:

Spoiler: NFSW

You don't have permission to view the spoiler content. Log in or register now.

Resolution is Quad HD or 2560x1440. Quality set at 1. Scene is Tuscany Bedroom (pretty simple props and not too many of them - admittedly I did not delete anything that is not in view). GPU (GTX 1080) is being used at full capacity. VRAM is not maxing out. No fall-back to CPU.
This render took 7200 seconds or 2 hours of rendering (max. time) and then it timed out.
Set max. samples at 5000 (default), but it only got to about 2800 iterations before it timed out and therefore convergence was not even at 60%.
It's not like I'm unhappy with the result. If I downsample and put in some PS work it might turn out okay, but I find 2 hours for this result to be pretty abysmal. There's still quite some noise and the sharpness is completely off.
Any tips for a complete noob? Is this normal for my hardware?
I found this thread in which DrPinkCake posted a pretty insightful comment concerning downsampling. ( Iray Rendering Tips ) Advice is given to render @ 4K but with max. 700-1000 samples and then to downsample for much cleaner, sharper and faster images.
What is your take on this. How can I improve my render time & quality (or at least find a decent balance between both)?
Thanks for any input you might be able to give!
Greets

 

[Saki_Sliz]

Yeah, this seems normal for an iray render. For professionals it's not too uncommon to have multiple gpu's to speed things up. One thing you could do that people started to do in the past few years is AI Denoising. Intel has a common one used by many programs and can be found in a variety of places. This allows the artist to lower render quality (sample count) which speeds up renders, I've even seen some render 1:1 size or even lower, but then use an AI to upscale the image, and then use a normal image editng program to downscale back to final size to make all the details and edges cleaner without using a terrible sharpening filter.

 

[Rich]

There is a whole running feud over whether you should render at higher res and then downsample, or just render at the lower resolution. Some people swear by it, others say it's stupid. In my own experiments, it's a very "your mileage may vary" thing - depends greatly on the scene. Recognize that if you're rendering at twice the horizontal and vertical resolution that you intend to use, you're asking iRay to render 4 times as many pixels. So, in the same amount of time, it's not going to get as many iterations in. Yes, when you downsample, you'll eliminate some of the noise, because you're essentially averaging 2x2 pixel squares, but the new denoiser built into Daz Studio may be able to do just as well. So, if you're intending to produce 1280x720 final images, I'd probably try rendering the same image, same settings, same timeout at the lower resolution, and see how much convergence you get.

Also note that things not-in-view-but-still-in-the-scene CAN affect your render times, because iRay may have to evaluate whether any of them are casting reflections, are light sources, etc. So, if it doesn't affect your render and it's out of view, removing it certainly won't increase your rendering times.

Finally, you appear to have a reasonably well-lit scene. Most people who complain about long render times have poorly lit scenes. But how are you lighting it? Do you have ghost lights in the scene? They help fill in dark areas, but they also can noticeably increase the time it takes pixels to converge, because iRay doesn't "know" that they're lights the way it does with spotlights. Sometimes, a large cross-section spotlight with a wide spread angle can do just as well, and provides hints to iRay that help it converge faster.

Anyway, just my $0.02.

 

[mickydoo]

As I understand it, the theory on downscaling is thus -

An image rendered at 1000x1000 takes 1 hour and leaves noise, rendering for 2 hours eliminates the noise.
An image rendered at 2000x2000 does not take twice as long, say an hour and a half. then when it's reduced it removes the noise.

I have no idea if that's actually correct, I used to downscale and cannot remember either way on times.

 

[Rich]

As I understand it, the theory on downscaling is thus -

Yes, that's the theory. But it always relied on people's judgment of "whether it got rid of the noise sufficiently or not." I never heard of anyone doing a truly scientific test (truly quantitatively evaluating the amount of noise left). Nor did I see a really quantitative evaluation of what would have happened if that original image had been rendered for 1.5 hours (same as the oversampled one).

Downscaling an image is certainly going to tend to reduce random noise, but, depending on exactly how you do it, it's also going to introduce artifacts of its own. Slight fuzziness at sharp edges, for example, since downscaling is (essentially) a form of a low-pass filter. (Which is why it removes noise, since noise is, by it's nature, high frequency) Different programs use different algorithms for it, which will produce subtly different effects.

Plus, there's a big subjective difference between noise in the face of a character that's "front and center" and noise in a shadow off in a corner.

So, it always amused me to see how adamant people would get wrt one tactic or another, when they were probably not really apples-to-apples comparisons.

But, all of this might be moot now anyway, since I'm betting that the AI denoisers available now probably do a better job. (Again, no empirical evidence to back up that statement, however.)

 

[Deleted member 1121028]

One thing you could do that people started to do in the past few years is AI Denoising. Intel has a common one used by many programs and can be found in a variety of places.

the AI denoisers available now probably do a better job.

AI denoisers gonna completely wash lot of skins details as most people don't provide a Normal pass and Iray can't provide an Albedo one (you can fake one with 3DL but nobody gonna do that). Those 2 are required if you want to preserve a maximum of details, without that the cost in quality is actually quite huge. In fact if you plan to denoise, sharpening diffuse/SSS maps before rendering with that terrible filter might not be a bad idea if it helps to keep some details .

Maybe I do something wrong, but all the tests I made (at pure equal quality) with downscaling were not conclusive, many time even longer with no real benefice.

 

[GNVE]

I don't know if it's just a subjective feeling but I think I noticed that my GPU uses more VRAM when rendering on a higher resolution. For a scene that is right on the edge it might be the difference between GPU and CPU rendering. So if I'm right stepping down a resolution might at times be advantageous. Haven't tested this rigorously, just a feeling I got when rendering 4K wallpapers for my Patreon.

 

[Saki_Sliz]

most people don't provide a Normal pass and Iray can't provide an Albedo one

Really? That stinks. Is there any compositing support, or does it just provide a final render only?

I don't know if it's just a subjective feeling but I think I noticed that my GPU uses more VRAM when rendering on a higher resolution.

yep. I think I've seen people talk about using programs to add more virtual video memory to a gpu to try to get over this issue, but performance slows down a ton, but not as much as cpu rendering. I was thinking of moving to a threadripper CPU since I use my computer for a lot of heavy work, but when I checked, it still couldn't beat out my graphics card, and the only other benefit would be fast cloth simulation and better fluid simulations, but that wasn't worth the cost just yet.

 

[Rich]

I don't know if it's just a subjective feeling but I think I noticed that my GPU uses more VRAM when rendering on a higher resolution.

That's completely expected, since the pixel buffer to which iRay is rendering is kept in VRAM. It's only periodically copied out for "as we go along" storage.

 

[Deleted member 1121028]

Really? That stinks. Is there any compositing support, or does it just provide a final render only?

From Iray? Yes, you have the common canvas

Spoiler

You don't have permission to view the spoiler content. Log in or register now.

Or more pass via

You must be registered to see the links

(who should be able to provide an albedo pass, but nope). As to why albedo pass doesn't work with Daz studio, I have no clue (bug? bad LPE implementation? Low on the to-do list?), someone may know better.

If your talking about AI denoisers, nope you have zero control.

 

[Saki_Sliz]

From Iray? Yes, you have the common canvas

Spoiler

You don't have permission to view the spoiler content. Log in or register now.

Or more pass via

You must be registered to see the links

(who should be able to provide an albedo pass, but nope). As to why albedo pass doesn't work with Daz studio, I have no clue (bug? bad LPE implementation? Low on the to-do list?), someone may know better.

If your talking about AI denoisers, nope you have zero control.

At least it has normal, depth and material id.

 

[Deleted member 1121028]

A (very) quick "gotta go fast" guide with Iray :

1. What you need:
- iray interior camera (file attached in the post)
- any shitty HDRI (file attached in the post)

2. Load your scene:

Here mine, daddy sleep naked his head into the cushion.
(Notice I already loaded an Iray interior cam & use it)



3. Light you scene:

Slot the shitty HDRI like this (use dome rotation if needed) :



Just an HDRI gonna be bland so I add a rectangle spotlight like this :



3. Making a test render

1080p with render quality enabled and infinite time/samples :



2020-08-25 19:48:40.120 Finished Rendering
2020-08-25 19:48:40.177 Total Rendering Time: 4 minutes 9.21 seconds (I have a 2070)



A bit of post process (aka hipster filters) :



hop done under 5 min \o/

 

[DonJoe]

A bit of post process (aka hipster filters) :

View attachment 787089

hop done under 5 min \o/

May I ask about these mysterious hipster filters ?

Some people use them these days and I like the outcome (seems like tonal contrast or film effects from color efex) ?

 

[Deleted member 1121028]

May I ask about these mysterious hipster filters ?

Some people use them these days and I like the outcome (seems like tonal contrast or film effects from color efex) ?

Indeed, it's from

You must be registered to see the links

. Did it so fast I can't remember . Matte filter, a flat curve and a bit of sharp/vignette I think (didn't bother to balance anything). If it's not that you should find it anyway, there are few ones .

 

[rev2020]

Remove the ceiling for faster rendering of "indoor" scenes.

For high quality renderers i usually renderer them around 10 hours or 20K to 30K Iray iterations.
(RTX 2080)
for some indoor scenes you can get away with only 3K Iray iterations.
it depends on how many light sources you have.
more light sources takes more time.

i think this one took 10 hours


Add spectral renderering Natural for more realistic renderers
this takes away that flat look that you have in your render.
but downside is that spectral rendering takes even more time.

and for total light control and highest dynamic range use canvases.
add every light source in the scene to an "light group" so you can control the lights after in PS.
renderers from canvases have higher depth and better colors than the default jpg/png renderers.

but every renderer looks flat and boring compared to HDR10 renderers.
to get the real HDR brightness and colors you must use canvases and save renderers in HDR format.
then use win10 latest build with an compatible PC that can show HDR wide color gamut in win10 along with an HDR10 capable tv.
app for showing real HDR10 renderers is "HDR + WCG image viewer"
search the microsoft store.

the only problem when working with 32bit HDR10 renderers in PS is that its difficult to set NITs to light sources
you need to guess how "bright" an light source will be.
but the amount of details you get in these renderers before you start to clip bright parts is crazy.
you cant compare it to SDR.

example of an HDR10 renderer
headlights on the truck shines at around 1000 nits
MaxCLL is 1026 Nits to be exact

i add the finished HDR10 file in the RAR file here.
View attachment MonsterX-Canvas1-hdr1.rar

 

[recreation]

For high quality renderers i usually renderer them around 10 hours or 20K to 30K Iray iterations.
(RTX 2080)

i think this one took 10 hours

WTF man, that's awfull, 10 hours and there's even still some grain visible. Seriously dude, do yourself a favor and look into how to optimize render times... also 20k iterations is crazy, I barely go over 5k and have no grain.

 

[rev2020]

nah i dont have any rush.
skinny bitches goes faster to renderer btw.
less materia to calculate

 

[recreation]

nah i dont have any rush.
skinny bitches goes faster to renderer btw.
less materia to calculate

Suuuure xD
You do realize that it doesn't matter how fat you make your chars, the poly count and texture size will stay the same
Btw the op talks about rendering being slow, you having time to spend 10 hours per render is your thing, but what's the point in telling him to spend more time than he already does if that's already too much for him?

 

[DonJoe]

no__name

Thank you

 

[recreation]

Indeed, it's from

You must be registered to see the links

. Did it so fast I can't remember . Matte filter, a flat curve and a bit of sharp/vignette I think (didn't bother to balance anything). If it's not that you should find it anyway, there are few ones .

There's some nice asset that let's you do that in Daz with diverse camera filter, it's of course not as quick to set up as in lightroom, but it has it's benefits doing it directly in the scene with lights, volume etc.