Comparison: (25 sample / 250 Samples / 2500 samples) + Intel Denoiser

Hi everyone,

I was trying combinations of Intel Open Image Denoise with variable number of samples in Cycles Render. To do this, I pick a scene that I’m working on and I rendered with 25, 250 and 2500 samples to compare the looks and render time difference.
I used blender 2.82 with Cycles Render using GPU (AMD Radeon Rx 580 Dual, 8GB) . The render dimension was 2880x1620 pixels and, in all results, I used the Intel Open Image Denoise with the same post production nodes in Blender Compositig.

Render Results



Close-Ups


Render time:
25 samples - 00:08:30 minutes.
250 samples - 00:52:40 minutes.
2500 samples - 08:10:40 hours.

Conclusion

First, the Intel Open Image denoiser is really powerfull, it can get good results from really noised images and, in my opnion, can get a much better result then Blender Denoiser, NVidia AI Denoiser, V-Ray Denoiser and etc. I know that this topic is not comparing the denoisers, but it’s worth to talk about it.

It’s notable the difference between the Image with 25 samples and the others, but the visual difference betwen 250 to 2500 samples is very small, while the render time is so disparate! We know that most people will see the images that we made in a little screen, like a smartphone, therefore the difference will be even smaller. So, in my cases, 250 samples with the Intel Open Image Denoise it’s a best cost benefit.

Thank you for watching and, please, post your opnion about this comparasion!!

Behance
Instagram
[email protected]

4 Likes

Nice work on the comparison research, good meaningful numbers.

Makes you wonder if the “average person” would even notice the difference between 25 and 2,500 - potentially saving you (the artist) 8 hours to create some more !!

Thanks for your time.

1 Like

Thanks for your reply, @calpgrmr.
On really, I did asked for ‘avarage person’ in my contacts if they can see the difference, the most of them said no!

That’s really cool. 2 years ago and no one would have believed you. I wonder what A.I. will take next in blender.

3 Likes

I wonder if 3D artists could see a quality difference if you showed them the images without them knowing which one has how many samples.
I mean, if you switch from the 25 to the 2500 image (the whole scene imgage) you can see that it is different because some things are jumping but I wouldn’t say that one is better than the other.
Take the notebook in the lower right. In the 2500 it is a bit brighter and there is a bit more detail in the grime. Does this difference matter? Imo not really.

But I guess this scene is a very good scenario for denoising because of the strong dof. The detail in the plant for example doesn’t really matter because it is out of focus anyway. If the scene was without dof, objects as close to the camera as the plant and as detailed as the plant might not work that well with low sample count.

3 Likes

I legitimately can barely tell the difference even knowing that it’s here. These images are huge, so it’s a bit easier for denoiser to make sense of it. That’s on a smaller ones issues appear, since there’s not a lot information to extrapolate from.

2 Likes

Hey @Lumpengnom, thanks for reply!
Maybe this scene can be more approprieted scenary for denoisers, but the DOF in most cases are not a good place to denoise, specially when we use Bokeh effect in the camera (I used 6 blades).

1 Like

It is 2880 x 1620 pixels which is not that large. Most of my customers want 4k for animations these days. For stills they usually want 10 or 12K.

2 Likes

That’s interesting… maybe a another good comparison to do is ‘low samples + denoiser + larger image’ VS ‘low samples + denoiser + small image’ VS ‘high samples + small image’.

1 Like

Ah, ok. What I meant is that the viewer will accept some splurging caused by denoising in places where it is maybe not supposed to be because there is so much “correct” splurging caused by dof.

1 Like

I know the samples, but i still can’t see a significant quality difference that would justify the increased rendertime.
Guess i stick to 25 samples and the Intel Denoiser from now on.

1 Like

Even if it was the same rendertime, is one actually better than the other in any significant way?

1 Like

Between 25 to 250 yes! But 250 to 2500 no… Watch in the Luminary video the bokeh in the 25 samples… looks horrible!

Hm, that doesn’t really bother me. But I have to say that the vast majority of my jobs are on the smaller budget side where this kind of thing doesn’t really matter.

1 Like

Just for a sake of it - maybe, but what’s the point, if it’s clear that with the reduced rendertimes it’s worth to up a resolution, even if you don’t want need it, so picture would be more clear?

It depends on the picture, of course. If you have a ton of high frequency detail even on that scale, Denoiser can potentially mud it. In case of super sharp and detailed forest, grass or another pattern, for example.

1 Like

IDK about that, since even on the examples it’s clearly visible that the edge of the table is mudded by a denoiser, even considering DOF behind it. Another thing is that barely anybody notice it and it’s not worth it to put additional render times in it. That’s like Richard Schmidt using lost edges to make pictures better and cut a bit of work for himself.

1 Like

Could you reword that, I am not sure what you are saying.

Here is a image without denoiser


This is a preview version, without a sun lamp, but this is the one difference.
Look at the edges of the table.