Denoising vs higher sample count

Hi all,

I’m finishing off a scene at the moment, and while tweaking a few render settings noticed something I’d like to discuss with the Intel Open Denoiser, specifically compared to good ol’ rendering with a high sample count. What I noticed was that while all the noise is gone from the image as intended, it almost has a look about it like it is reducing the color depth? See the below images and I’ll discuss a bit more.

Close-up of a cabinet in my scene, 5000 samples:

Same shot, with the denoiser:

You can see slight color banding on the denoised image that isn’t present on the first one. It is more pronounced in the render window before saving. The conclusion I’ve drawn from this is I will probably continue to use just a high sample count for renders for a couple of reasons:

  1. I’m not a fan of the above ‘banding’ look where it looks like the color depth has been reduced

  2. I actually prefer the look of a little bit of noise in my image (not fireflies though). I know many disagree but I’m of the opinion the effects we add on that photographers try to remove - DOF, noise, chromatic aberration etc are good when tastefully present in a render

  3. Render times aren’t a big deal for me, I’m happy to leave it rendering overnight on a 1080ti

Now this doesn’t mean I see no place for the Intel Open denoiser. The quality it can produce for preview renders with next-to-no samples is astounding, and really improves a workflow. You’ll probably notice I have no question at the end, I’m actually just interested in the point of view of more experienced artists.

Thanks.

I see point 2. the same way and I am also of the opinion that A.I. denoiser are a very good help for working as I believe that all kinds of a.i. stories are (will be) a good help in 3d. But for the final result I use the denoiser (as well as for all other photography tasks) from Darktable.
Therefore I am not of the opinion that you need such a high number of samples. Also I can’t sleep when the computer is on. LOL

The banding is most likely just compression artifacts from the image format you used to save the image in.

As for the noise, most renderers usually implement denoiser as a blend amount, where you can blend denoiser with the original image. It’s exactly for this reason; Too much noise is unrealistic, but so is 0 noise. Yet, unwritten rule of Blender is to make things as complicated as possible, so as usually, you gotta do it yourself:

Just drop in a mix node to blend between original and denoised image, and dial in the right amount of noise using the Fac(tor) input.

Thanks for the input, I’m struggling to understand a bit so you do use a denoiser for final rendering?

The banding is present in all the viewport windows too, render window, compisiting window, actually after I save the image it seems to be less noticeable…

That’s a good idea actually. I’ve seen this method before with general compositing, because we tend to go overboard with all the effects, mix the composited version with the original rendered image to bring the subtlety down a bit.

Really suprised nobody has noticed the banding in their own tests, it’s quite noticeable on mine and isn’t present before denoising.

Yes, but not in Blender, but in Darktable and only as much as necessary as @rawalanche describes. But I use the viewport denoiser from LuxCoreRender all the time to estimate the scene. Of course I render the denoiser layer to use it in compositing (e.g. for intentional blur effects). But as I said, I find the look much more natural with the detour via Darktable.

Didn’t notice banding (in my images) and I don’t think it should be the case but what do I know? Anyway good point that orig/denoise mixing will try that out. @rawalanche

Generally, any non brute force methods and some brute force methods which converge to near perfect levels (cycles being brute force renderer) will produce gradients so perfect that with just 8 bit display color range, there will be apparent banding. So if the banding is not caused by the image compression, then it’s definitely lack of dithering.

Dithering is present in almost every renderer, as a method to cover up this apparent banding. Based on the evidence so far it appears that the dithering gets applied prior to denoising, and intel denoiser then cleans it up along with the brute force noise. Dither setting in Blender is present here:

Anyway, to combat this, the solution of Blending a bit of original noise onto the image using mix node should be sufficient.

1 Like

Interesting thought about the dithering, so if I understand what you’re saying cycles is dithering the image, then the denoiser is essentially undoing that when it does it’s job?

I will definitely give the idea a go of mixing the image. Something I’d heard about with general compositing effects but never thought of mixing a denoised and non-denoised one, hopefully should give the best of both worlds. Good info!