Denoiser Workflow // Questions

[KeeWe]

Hey Guys,

I'm currently setting up a scene where I want to use the AI Denoiser. I set up the scene and was pretty impressed with the first result, unfortunatley I had to go back some steps due to some mistake happen before setting up the lighting and denoiser settings. Since then the results don't look the same, I get a very unsharp and "blotchy" image, se the attached screenshot.

I wonder what are the best settings or at least the correct workflow setting up the denoiser. So far I know:

- Activating it via the camera imager. Is there a difference between global and the Tag camer imager?

- Render Passes: set the tone map type to Tonemapped, Image Color Profile SRgb save the Denoised Beauty.

Whats with the Min Denoiser samples and the Denose interval? I didn't see a difference there.

So far I can see that the results using very few samples (under 100) are not useable. The scene clears up at about 3000 - 4000 PT Samples.

Can someone guide me in the right direction? What settings do I have to tinker with? Or is the solution to go up with the samples to, say 1000?

Goal would be to get the render times down to 90 seconds or so, coming from 6 minutes. is this possible?

Please see attached my current render settings and the sample image.

Current light setup only contains one HDR but I'll probably add some are lights.

Please notice the slightly off lighting as well. Is Srgb and tonemapped the wrong option?

Thanks in advance.

Edit: Octane version: V4.0-RC7

[KeeWe]

Just a quick follow up, please compare the attached images. First is the AI Denoiser, 450 samples. Second is the same image, 450 samples Beauty pass with applied denoiser insider AE.

AI Denoiser looks like a badly compressed jpeg, while the Denoiser in AE remains the details. Look at the clean plane on the right hand side comapred to the AI denoiser one.

Is this normal? I'm still pretty confused about that first testing where everythings seems to look fine.

[frankmci]

I'm no expert, so take all of this with a pinch of salt. A few items on your list:

Min Denoiser samples are only used if "Denoise on Completion" is turn off. It tells the denoiser not to bother running before that threshold has been passed. The Denoise Interval is how often it runs on longer renders. It can be handy for images like a high res stills, but for animation you'll probably just stick with Denoise on Completion.

Yes, under 100 samples is pushing it, with all the depth of field/motion blur you're using. Depending on the context, I usually start at 128 and increase as necessary. 3000 sounds excessive to me. For that image, I'd guess 300-400 should do the trick.

It seems quite reasonable to render that image in less than 90 seconds (depending on your hardware, of course. Even a GTX 980 should handle this scene pretty well.) It might render faster with simple three point lighting, depending on what kind of geometry you've got going on outside the frame. Sometimes an HDRI is overkill, especially when most of it is blocked by geometry. As a general rule of thumb, only emit light from surfaces that have an obvious impact on the visible geometry. In an enclosed/confined space, most of that HDR surface is often going to waste and just eating up render time spitting out minimally useful rays.

I think 14 samples for diffuse and specular are overkill. Yes, this is all indirectly illuminated, but still, 14 bounces is an awful lot. I usually start with 6 diffuse, 2 specular, and increase as necessary.

Adaptive sampling, wise, I usually set my noise threshold a bit lower, at 0.02, but also reduce min samples to 32.

Tonemapped sRGB is fine if you are happy with direct output, but you will find people who will argue about high bit depth linear color space 'till they turn blue. I try to avoid religious arguments on the forums, though.

[Ferdinand13]

Can I quickly chime in here?

I'm currently using the Denoiser in an Octane network render environment, got a PC with three GTX 1070s as the master and a laptop with another GTX 1070 plus a GTX 1080 Ti connected to that laptop as an eGPU.

Now on the surface, everything looks fine. I managed to cut my former render times in half and I'm happy. However, I had a closer look at the Denoiser passes that this network render produced and I'm not so sure anymore.

Here's the Denoised beauty pass. It looks good:

However, the Denoised reflection indirect pass is a different story:

The plane behind the phone is noisy as hell. What the heck happened?

And even though it's hard to see, the Denoised remainder is very noisy as well:

Now, this only seems to happen within a network render environment (and I vaguely remember someone else reporting this issue on the forum). What's the fix? The render times I'm hitting are super sweet and finally allow me to implement Octane Render in a faster turnaround, but if I can't use Denoiser in combination with network rendering, it's kind of pointless (as a lot of the speed gain comes from saving max samples by using the Denoiser).

Any help would be much appreciated!

[Ferdinand13]

Hey,

Any updates on this one? I've reverted to standard multipasses for my current project (rendered via Pathtracing with 1500 Samples), as the Denoiser creates a blotchy mess in the Denoised Reflection Indirect and the Denoised Remainder passes. Not sure why this is, but using 'classic' noisy passes and then running noise reduction in After Effects (with the Neat Video plugin) looks waaaaaay cleaner (first and foremost it avoids the 'blobs' that the Denoiser creates).

Am I the only one who has these issues? It would be great to get some feedback as to how to make the undoubtedly powerful Denoiser work in my workflow.

Cheers!