Hi, I've recently needed to do some tests to reduce indirect noise and decided to post the results here for some feedback.

This is a small rendered section of the indirect_reflect image plane where I could see the most noise. I used default mantra rop settings as a baseline but did change two things;
1. The Raytracing Bias was reduced to match scene size. (This a cropped section of a USB port)
2. Constrain to Maximum Roughness is enabled. My brief test showed it reduced noise at minimal accuracy loss. (Does what it says on the tin)

I've repeated results on the first and last columns to make it easier to compare results.

- I first started with 3 pixel samples and increased the Reflection Quality until it had no affect on quality. (14 was the sweet-spot)
- I then increased the pixel samples to 4 but it produced more noise. (Compare 1st row, 2nd image to 1st row, 3rd image)
- Increasing pixel samples further did reduce noise but time taken was increasing too much.
- At 8 pixel samples and 14 reflection quality (last column), noise is almost completely gone but time taken is too high.
- I then started with 8 pixels samples, and different reflection quality settings to compare times/noise.
- I looked over results and tested some combinations.

To test further, I'd like to understand why there is an increase in noise from 3 to 4 pixel samples. (Any info/suggestions appreciated)
If I say that 3 pixel samples and 14 reflection quality (2nd image) is my baseline for time vs quality, is there anything other than pixel samples I should look at to further reduce noise? (Increasing the pixel samples will make other (non metallic) materials in the scene take longer and I want to avoid oversampling)

Rob

Thanks for this post exploring the various image quality settings. It is interesting to see the interplay of variance quality and pixel samples. Have you experimented with min/max rays and noise level yet? These settings, although less intuitive, may allow finer grained control over sampling than the convenient quality level controls. To optimize the balance of min/max and noise level, I recommend enabling the utility passes “Indirect samples” and “Direct samples” to evaluate their effect.

Good luck on your endeavors.

Hi, I forgot to mention that I did actually set the noise level lower for these tests. (0.0001 for all)

I did have a brief look at min/max but decided to keep the initial tests focused on reflect quality and pixel samples. I did one test sample of min/max and felt the quality gain vs time wasn't as effective compared to reflection quality.

ie. 8 pixel samples, 1-24 min/max, 1 Reflection quality, took 3m57s and produced slightly more noise than using 8 pixel samples, 4 Reflection quality, 3m 47s. I didn't do further tests as the times would only get worse and felt it would also contribute to oversampling the rest of the scene.

I was impressed by the amount of noise reduced by using the reflection quality compared to oversampling with pixel samples. My next step would usually have been to slightly increase pixel samples (ie. from 3 to 4) and then look at the min/max. However, my results produced an anomaly which I should try to understand before continuing.

Next, I may either use the 1m46 (2nd image) as the baseline and work with min/max or use the anomaly (3rd image) to see if min/max will fix it. I'll also look at the Indirect Samples passes too.

thanks
Rob

Hi, do you have an example .hip we can look at? (for increased noise when pixel samples go from 3 to 4) Thanks.

Hi, I created the model inside Houdini but it is actually for a commission so I can't share it publicly. I'd like to keep this open to other users so I've created a quick scene using similar techniques, materials & settings.

Although the noise isn't quite as noticeable overall, it does increase from 3 to 4 pixel samples. I'm sure it's something simple I'm overlooking so I'd be grateful for the insight.

thanks
Rob

This seems to be RNG issue - it seems to be generating correlated numbers and causing bias and increased noise (particularly when reflection quality setting is set high). This will be addressed in the next major release.

Thanks for confirming the issue dlee, that's something I'm unable to fix myself.

When you say next major release, do you mean next production build or the next version like 16.5 or 17?

If it's production build, I'll probably wait until then to continue render tests. The job itself is going to take a long time and this is just a small part of it. (I have a lot of modelling to do before I do the final renders)

If it's 16.5/17, I'll have to either revert to higher pixel samples or postwork noise reduction. (ie Photoshop)

This is assuming (guessing) that the next production build will be within a month or so and the next version will be much further away.

Thanks
Rob

We don't want to introduce look changes for productions that are using 16.0, so it will have to be the next version.