Is the Optix denoiser in Arnold 7 worse?

[phibonacciSequence]

When I updated to Arnold 7 and rendered an image I had previously with Arnold 6.2.1.1, using the GPU and Optix denoiser, I noticed the image looked blurrier. I hadn't changed the settings so are there any that I have to now that I've upgraded to 7? Since then I've downgraded back to make sure and it (at least to me) looks sharper.

I've uploaded the images (at different resolutions) to compare.

I've also uploaded the 2 renders in video format.

(I'm using Maya 2020.4 and NVIDIA GeForce RTX 2080 version 472.12 (Studio Driver)).

[phibonacciSequence]

I've uploaded a gif/mp4 in case the differences aren't clear enough: https://i.imgur.com/RLSrCdM.mp4

[thiago.ize]

I think the only change we made was to add specular reflections and refractions for denoising. This should have made it look sharper, not blurrier. Or more specifically, it should look more accurate.

Can you post an Arnold 7 image showing what it looks like if you turn off denoising but use the same number of samples? Also, can you post another image without denoising but this time with enough samples that it is converged? To speed things along, feel free to render just a sub-region that shows lots of the blurring.

It might be interesting to also try the imager_denoiser_oidn that is in arnold 7.

[phibonacciSequence]

Sure, I've uploaded the images to imgur.

I'm not sure what you mean by "enough samples that it is converged".

Also are the hyperlinks not working in my original post? I can't seem to open the Imgur and YouTube links on my end.

[thiago.ize]

Yeah, those links seem to be empty. Can you directly paste the URL instead?

As for converged, I mean increase the AA until you get a mostly noise-free render without the help of denoising. This will give us a correct reference image to go off of.

[phibonacciSequence]

I see. Here are the URLs from the original post:

I've uploaded the images (at different resolutions) to compare: https://imgur.com/a/dIEF6DB

I've also uploaded the 2 renders in video format: https://youtu.be/r6bPSUjzyPc

And here's the URL from my reply: https://imgur.com/a/L96fIBm

I'll get on the converged image now. Would uploading the scene file help? (It's mostly just a royalty free model and HDRi anyway).

[phibonacciSequence]

Here's the converged image (I set the Adaptive Sampling to 100 (AA)):

[thiago.ize]

Thanks for all the images. The 6.2 images do look better. It looks like the imager_denioser_oidn is doing a good job and doesn't have the optix issues (or did I miss it?). If this is the case, are you ok with using oidn instead of optix for this type of scene?

[thiago.ize]

If you can post here the scene or a link to it, I can try it myself, but if you want to do it, tell maya to render out the custom AOVs "denoise_albedo" and "denoise_albedo_noisy" and upload these images here. While you're at it, upload the N aov as well since that's used by oidn. Since the only change I'm aware of is the new denoise_albedo_noisy I'm curious what this will show.

[phibonacciSequence]

I'll probably just downgrade back to 6.2.1.1

[phibonacciSequence]

I can't find the AOVs "denoise_albedo" and "denoise_albedo_noisy" and "outputing denoising AOVs" hasn't worked since Arnold added imagers (Arnold 6.2). I can either output the image with denoise or without (unless I export it in the Arnold viewport).

Here are the images rendered:

Oidn: https://imgur.com/a/3p0lLWo

Optix: https://imgur.com/a/qKfWPIi

I'll upload the archived scene file to mediafire in a bit; it's too big to attach.

[phibonacciSequence]

Here is the archived scene file: https://www.mediafire.com/file/3vuxcljq4ukeirr/disco.mb.zip/file

Or just the scene file:https://www.mediafire.com/file/d7cvkowfdokbo40/disco.mb/file

The HDRi I used was from HDRihaven/Polyhaven - Studio Small 01 at 4K. https://polyhaven.com/a/studio_small_01

[thiago.ize]

You add these albedos using the custom interface. As long as you type the name exactly like this it should work.

The "output denoising AOVs" is for use with the standalone noice denoiser. You don't need it for imager denoisers.

[phibonacciSequence]

Sorry, I forgot I was using GPU to render (which causes visual bugs in the scene when in another computer). Here's the scene set to CPU rendering:

https://www.mediafire.com/file/h08j71wyh53io8i/disco_cpu.mb/file

[phibonacciSequence]

Thank you. That's been driving me nuts. (Strange they didn't add it to Maya)

[phibonacciSequence]

I downloaded a couple scenes from: https://docs.arnoldrenderer.com/display/A5AFMUG/Learning+Scenes and there wasn't much of a difference when I rendered using optix in either 7 or 6.2.1.1 (though in some cases like the Toy-Space-Pirate-Studio and Transmission-ray-depth_Diamond the denoiser did improve in 7)

My (unprofessional) guess is that there's an issue upgrading the renderer when optix was already enabled in a saved scene (or maybe it just doesn't like the chrome preset for aiStandardSurface).

[thiago.ize]

Thanks for providing the scene. Looking at the denoise_albedo which is used in Arnold 6:

and denoise_albedo_noisy from Arnold 7:

I think makes it clear what is going on. The 6 is missing albedo information in the mirror part, so it's black, and this combination of black mirror face and the grey mirror edges work out in your favor and the optix denoiser is able to detect that these edges are important and shouldn't be blurred.

The 7 albedo happens to have the mirror face and the mirror edges almost the same color (it's using the base and specular colors you assigned to these shaders), so the optix denoiser has a hard time telling them apart and thinks this is all a single feature and blurs it all together.

While this might be an unfortunate regression for your discoball scene, scenes that are mixing different materials (more specifically materials with different albedos) normally see an improvement. A great example is at https://docs.arnoldrenderer.com/display/A5ARP/7.0.0.0 where the sima model has blurry hair with Arnold 6 optix denoising and really nicely denoised hair with Arnold 7 optix denoising (scroll down to the "Higher quality AI denoising" section). There are other good examples when specular shaders and diffuse shaders are mixed together, such as a metal stove next to a wood table...

[phibonacciSequence]

So for this one instance, 6 happens to work in my favor.

Thank you for taking the time to answer this.

I guess I'll just stick to 6 for now.