Daz Daz Studio post denoiser vs MCJ

[Fatalmasterpiece]

I was genuinely curious, I feel like the only person who doesn't hate the built in denoiser in Daz Studio.
It seems to do a decent job and is built in so no need to run an extra script, it looks good to me and it just works
Some people say it slows down render times. I haven't experienced this myself. As you can see below, for some reason the render with DS denoiser was actually slightly faster.
Many others swear by MCJ denoiser, so I wanted to do a direct comparison. I didn't bother trying Nvidia or Intel denoiser because others have already tested them with less than stellar results.

I tried to set up a scene that would purposely have LOTS of noise so you could see the difference.

Model is Aurore Genesis 8 with 2021-09 hair, scene is "Art Deco Bath" with 1 ghost light on the ceiling, 8 candle lights and an HDRI. Section plane nodes on all sides of camera. Render Quality OFF, Max Samples 900, 1920x1080

No denoiser, render time 4 minutes 56 seconds


DS post denoiser set at 880, render time 4 minutes 52 seconds


MCJ denoiser, just a few extra seconds to run the script.


So how did they do? Well like most things, turns out the devil is in the details.


MCJ does a really good job but potentially looses more detail? On 4k images MCJ might be the better choice though. On 1080p images I don't think it makes that much of a difference.

 

[Empiric]

I use intel denoiser and it just performs better than daz in-built one. But even if they performed exactly the same, I'd not use the in-built one for one simple reason. That is, if you use the daz in-built one, you can't un-use it. Daz spits out the already denoised image at you. When I use the Intel one, I have the denoised and non-denoised image and I can just blend in the areas I want denoised and not lose detail where there was not as much noise in the first place.

 

[Fatalmasterpiece]

I use intel denoiser and it just performs better than daz in-built one. But even if they performed exactly the same, I'd not use the in-built one for one simple reason. That is, if you use the daz in-built one, you can't un-use it. Daz spits out the already denoised image at you. When I use the Intel one, I have the denoised and non-denoised image and I can just blend in the areas I want denoised and not lose detail where there was not as much noise in the first place.

That is a very valid point. I can see why for some people that would be a valuable asset.
Personally for me I haven't needed the unprocessed versions and the built in step saves me time and keeps my render library less cluttered. That wouldn't be the case for everyone.

 

[Empiric]

That is a very valid point. I can see why for some people that would be a valuable asset.
Personally for me I haven't needed the unprocessed versions and the built in step saves me time and keeps my render library less cluttered. That wouldn't be the case for everyone.

Yeah, you can basically control the strength of the denoising, by putting the denoised image over the original and reducing the opacity to reduce the grain enough but to keep the details. With in-built denoiser you go either 100% or 0%, which is the biggest downside.

 

[function2020]

Good comparison. I only use DAZ post denoiser, because it is simple, and effective. But I really don't know the difference between start Iteration 8 and 880, need another comparison.
Talking about details, its pixel filter Mitchell with 0.92 works fine. More over, render double size and resize in PS makes it more clear and sharp.

 

[Fatalmasterpiece]

Good comparison. I only use DAZ post denoiser, because it is simple, and effective. But I really don't know the difference between start Iteration 8 and 880, need another comparison.
Talking about details, its pixel filter Mitchell with 0.92 works fine. More over, render double size and resize in PS makes it more clear and sharp.

Good call, here are the results. I watched the render and denoising did indeed start at 8 iterations and looked terrible at first, until more iterations had passed. Your overall render iterations are what really seem to matter though, but there is a diminishing return on time. Honestly... I don't think denoiser iterations make a big difference. The change is nearly imperceptible.

The differences in render time here are probably just due to slight temperature differences and background process in my computer.

900 render iterations, denoiser set to start at 8 iterations. Render time 4 mins 37 seconds


900 render iterations, denoiser set to start at 899 iterations. Render time 4 mins 36 seconds


100 render iterations, denoiser set to start at 8 iterations. Render time 49 seconds


100 render iterations, denoiser set to start at 90 iterations. Render time 53 seconds


10 render iterations, denoiser set to start at 8 iterations. Render time 12 seconds

You can see at only 10 iterations, the details really do not come through at all and the denoiser kills all the fine detail.


You're absolutely right about oversampling, which is a good practice. I'll do it if its an important shot of someone's face or something like that. But if its just a render of the outside of someone's house or a landscape (which I also do a lot of) I'm not going to bother.

 

[function2020]

Good call, here are the results. I watched the render and denoising did indeed start at 8 iterations and looked terrible at first, until more iterations had passed. Your overall render iterations are what really seem to matter though, but there is a diminishing return on time. Honestly... I don't think denoiser iterations make a big difference. The change is nearly imperceptible.

The differences in render time here are probably just due to slight temperature differences and background process in my computer.

900 render iterations, denoiser set to start at 8 iterations. Render time 4 mins 37 seconds
View attachment 1636358

900 render iterations, denoiser set to start at 899 iterations. Render time 4 mins 36 seconds
View attachment 1636360

100 render iterations, denoiser set to start at 8 iterations. Render time 49 seconds
View attachment 1636371

100 render iterations, denoiser set to start at 90 iterations. Render time 53 seconds
View attachment 1636372

10 render iterations, denoiser set to start at 8 iterations. Render time 12 seconds
View attachment 1636373
You can see at only 10 iterations, the details really do not come through at all and the denoiser kills all the fine detail.


You're absolutely right about oversampling, which is a good practice. I'll do it if its an important shot of someone's face or something like that. But if its just a render of the outside of someone's house or a landscape (which I also do a lot of) I'm not going to bother.

Thanks, your job is perfect. Those render iterations and Max Samples are same thing which I recommend to set 200 in my post, 200 almost enough for most situations.
Through your comparison, it looks really no difference when choosing those post denosier iterations, believe that's the reason DAZ set 8 as default, small number can let people stop render at any time they satisfied, big number means they have to wait till denoiser start.

Today I tried to investigate the Advance render setting - Texture Compression Thresholds, raise up both to 8K, it tremendously extend the render time, but didn't see any difference under 2560 x 1440 resolution. I would like to search some knowledge for this in DAZ forum but its server seems broken in this weekend.

 

[Fatalmasterpiece]

Just for curiosity I did a super HD render of 900 iterations at 7111x4000 and sampled it back down to 1080p. Turned off Daz Denoiser. Render time 56 minutes


And here is 1920x1080 at 4000 iterations, no denoiser. ~20 minutes


So if you have the time, doing hi-res or many more iterations can be a decent way to cut down on noise.
Also the DAZ forums are always down it seems, and not very good even when they are up. I have learned so much more here than there.

 

[function2020]

Just for curiosity I did a super HD render of 900 iterations at 7111x4000 and sampled it back down to 1080p. Turned off Daz Denoiser. Render time 56 minutes
View attachment 1636520

And here is 1920x1080 at 4000 iterations, no denoiser. ~20 minutes
View attachment 1636519

So if you have the time, doing hi-res or many more iterations can be a decent way to cut down on noise.
Also the DAZ forums are always down it seems, and not very good even when they are up. I have learned so much more here than there.

yeah, these 2 pictures have more details (e.g handrail) than above, but still leave very fine noise grains all over. Seems like denoiser 'kills' some details to denoise?

However, more test, generates more questions, for example:
1. is Firefly filter on? I normally let it on. How about turn it off and only rely on Post denoiser?
2. what is the usage and effect of different 'Noise degrain filtering' values?
3. what is the usage and effect of 'Post Denoiser Denoise Alpha'?
4. regarding details, compare with picture size/resize, Max Samples value, and Texture compression threshold, which one is the most cost-effective way?

 

[function2020]

Found a weakness of the DAZ Post Denosier.
Yesterday I investigated detailed photo-real skin, no matter how I adjusted those parameters, camera focal, and maps, all pictures still looks same and very smooth. Finally I found it is because the Post Denosier.
After turn it off, all details appeared in the picture, together with noise, so it needs to raise up Max Samples.
Don't know if de-noise in PS can only remove noise but not make the skin so smooth.

 

[MissFortune]

Found a weakness of the DAZ Post Denosier.
Yesterday I investigated detailed photo-real skin, no matter how I adjusted those parameters, camera focal, and maps, all pictures still looks same and very smooth. Finally I found it is because the Post Denosier.
After turn it off, all details appeared in the picture, together with noise, so it needs to raise up Max Samples.
Don't know if de-noise in PS can only remove noise but not make the skin so smooth.

The Daz denoiser is far, far worse than the other AI denoisers out there. Intel and Nvidia AI denoisers are probably the best out there currently, with Topaz being a close third. The problem with the Daz denoiser is that no matter what you do, you're losing far more detail (especially in the skin) than is really necessary.

That being said, you're experiencing the inherent (or current) issue with AI denoising (or denoising as a whole, for that matter.). You/the computer are basically suppressing noise (which, in it's absolute simplest sense, is blurring the noise), which comes at the cost of losing small detail that you wouldn't notice from far away. You're going to lose stuff like strands of hair and skin pores in place of a more overall visually pleasing image.

 

[function2020]

The Daz denoiser is far, far worse than the other AI denoisers out there. Intel and Nvidia AI denoisers are probably the best out there currently, with Topaz being a close third. The problem with the Daz denoiser is that no matter what you do, you're losing far more detail (especially in the skin) than is really necessary.

That being said, you're experiencing the inherent (or current) issue with AI denoising (or denoising as a whole, for that matter.). You/the computer are basically suppressing noise (which, in it's absolute simplest sense, is blurring the noise), which comes at the cost of losing small detail that you wouldn't notice from far away. You're going to lose stuff like strands of hair and skin pores in place of a more overall visually pleasing image.

I didn't install any extra denoisers before, but yesterday I checked the

You must be registered to see the links

,

You must be registered to see the links

, in its first page, the comparison picture, my God, it even worse than DAZ's, every details gone, totally blur. So I didn't download it. Guess Nvidia's is the same.

I am thinking about using DAZ Canvas, to render objects which need super details separately by turn off the post denoiser, then turn on it for the whole picture, finally combine them together, this maybe much easier.

 

[Andreas Hardcastle]

The Daz denoiser is far, far worse than the other AI denoisers out there. Intel and Nvidia AI denoisers are probably the best out there currently, with Topaz being a close third. The problem with the Daz denoiser is that no matter what you do, you're losing far more detail (especially in the skin) than is really necessary.

That being said, you're experiencing the inherent (or current) issue with AI denoising (or denoising as a whole, for that matter.). You/the computer are basically suppressing noise (which, in its absolute simplest sense, is blurring the noise), which comes at the cost of losing small detail that you wouldn't notice from far away. You're going to lose stuff like strands of hair and skin pores in place of a more overall visually pleasing image.

Hint: The Nvidia AI denoiser you named as probably the best together with Intel... is the NVIDIA Optix Denoiser. Which is the same denoiser used by Iray, and therefore used in DAZ.

Idea: More time invested in doing research does help quite a bit when joining a technical discussion :-D

*
*
*

Getting back to the topic: I have tried pretty much every denoiser, sharpener and upscaler, from Intel, NVIDIA, various scientific projects like Real-ESRGAN and Waifu2x with various forks. In my experience, Topaz offers the best overall solution for image restoration/enhancement jobs. There are constellations where other programs have the upper hand, but for everyday stuff like cleaning up Daz renders, it always delivers.

In comparison with Intel Open AI Denoiser and Daz denoiser (aka NVIDIA Optix AI denoiser included in Iray), the Topaz DeNoise AI solution is in a completely different league. If you get worse results than Intel's and NVIDIA's denoisers using Topaz, please let me know how you did it - I actually consider that quite an achievement! :-D

Noise Degrain Filtering & Post Denoiser used by DAZ Studio and part of Iray:
Check out the

You must be registered to see the links

, freely available to browse and download as PDF on NVIDIA's Developer Website. It's an interesting read and written well enough that you don't have to be a rocket scientist to get it. Here's an excerpt about our subject at hand:

Spoiler: Noise Degrain Filtering

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

Spoiler: Post Denoiser

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

 

[MidnightArrow]

In comparison with Intel Open AI Denoiser and Daz denoiser (aka NVIDIA Optix AI denoiser included in Iray), the Topaz DeNoise AI solution is in a completely different league. If you get worse results than Intel's and NVIDIA's denoisers using Topaz, please let me know how you did it - I actually consider that quite an achievement! :-D

Are you testing Intel OID with albedo and normal passes though? Because in my tests in Cycles, Intel OID is excellent as long as you're providing the passes and not just using it as a glorified blur filter.

 

[felldude]

Daz uses Optix 5.0.1 for version 4.21.0.5

Optix is great for real time de-noising, it picks the best filter (Sinc, Lancoz etc) based off off the models they ran.
Its main use is things like up scaling games, or de-noising up scaled videos in real time.

For a single image there are defently much better options but I think for the new, or possibly even most daz users it is welcome and needed.

If your looking for the latest and greatest up scaling and de-nosing it is probably HAT_SRx4
But that hasnt even been fully implmented in pytorch yet.

 

[Fatalmasterpiece]

I can also recommend Topaz Gigapixel AI for upscaling. The best part is, it is easy to use and you can do a huge batch of images. I'm sure you can do that with others using scripting but the interface of Gigapixel is much more user friendly. Getting the settings just right can be a little tricky though.