V-ray 6 seems slower than V-ray 3.6. am i doing wrong?

[hong_seungsoo]

hi. our company used v-ray 3.6 and now doing some test with v-ray 6 for update.
i’m not english speaker so write body text with AI. if there is something akward, please understand.

What users expect from a renderer update is faster render times for the same noise quality, or better quality for the same render time.
However, in my tests, Vray 6 seems to be slower than 3.6. I was expecting a dramatic improvement since it's a major version jump, but the results are confusing.

Test Environment

i tested in two environments

• Maya 2018 + V-ray 3.6
• Maya 2022.5 + Vray 6

Test Scene

I've read on the forums here that the renderer may not work properly when importing a scene that was worked on in an older version into a newer version,
so I imported a simple test model with a single sphere and cube in a small room into FBX in each version and set the materials/lighting/render exactly the same in each.
here is google drive links of scene files.

• testscene_3 → Maya 2018 + 3.6 (google drive link)
• testscene_6 → Maya 2022.5 + 6 (google drive link)

Maya 2022.5 + V-ray 6

With 1 min/48 max Subdivs and a Threshold of 0.002, I got satisfactory results. The total render time was 35 minutes and 9 seconds.

Maya 2018 + V-ray 3.6

With 1 min/8 max Subdivs, a Threshold of 0.001, and a 200 of Local subdivs in Dome Light, Reflection, and Brute Force. the render finished in 26 minutes and 28 seconds.

Result

vray3.exr (google drive link)
vray6.exr (google drive link)

Aside from the slightly darker output in 3.6, The noises are visually almost same to me, or at least 3.6 seems to have better noise levels, while taking 25% less time to render. especially in the GI pass, 3.6 looks better.
I was thinking of doing a little more testing with different passes, but I thought I'd post here on the forum.
I am quite confused by these results. Is there something I'm doing wrong?
I've seen in some of developers posts that use internal numerical data to compare noise levels, and I'd like you to keep in mind that most V-Ray users are artists, not scientists or mathematicians.

Additional Question About Render Setting Optimization

About 6 years ago, when I was a student, I got really into V-Ray render optimization theory.
The idea was to increase the Local Subdivs of noisy elements rather than increasing the Max Subdivs of the Image Sampler.
In this post, ^LeLe^ says "As for local subdivs, we're very much in line with everyone else: there is no one left on the market which has manual AA/Sampling controls.".
Similarly, this post says that optimizing these methods is unnecessary since Vray 3.3, but I've been able to save a lot of time in Vray 3.6 with the above optimizations.
if the time spent trying to find the right render settings is included in the render time, I agree with ^LeLe^ that the current Vray6 approach (adjusting Max Subdivs and Threshold) is much easier and faster to get results in a project like Archviz that ends up with single image render.
However, I work in the animation industry. If I make an hour-long animation at 30 fps, the total number of frames is 108,000.
If I save a minute in rendering individual frames, that's like saving 108,000 minutes of time. So it's always worth taking the time to find the right settings for optimization.
Furthermore, the V-ray documentation states that you don't need to increase Min Subdivs unless it's a special case.
However, even in general situations that aren't the special cases the documentation talks about, increasing Min Subdivs will often give me faster results than increasing Max/Local Subdivs or lowering the Noise Threshold.

All in all, I'm confused about everything. Please let me know if I'm misunderstanding anything.

[Neilg]

Your test scene isn't really testing anything, you're pushing settings way beyond what is needed in production.

When you say you had to go down to a noise threshold of 0.002, are you zooming in to 500% on still frames of plain clay materials and reducing the noise threshold until the surface is perfectly clear? You need to check that in real production jobs and as part of a total pipeline. I've never gone below 0.01 for a project, and for most animation work i'll even pull that up to 0.02 or higher and denoise in post.
When you let vray 6 leave a little noise in the render, it's much more consistent noise - easier to remove in post, and more photographic. The noise you'd get in vray 3 was unpredictable and some areas of the image would have a lot and others none. This more even noise is a lot more desirable and easier to work with, so you have to adjust your workflow accordingly. But trust me, nobody is rendering animations with a noise threshold of 0.002.

[hong_seungsoo]

I understand what you're saying, but I think denoising in post is beside the point. My purpose in writing this post is simply to compare the performance between V-ray 6 and V-ray 3.6.
Also, I'm a rigger, so I don't know much about denoising in post. However, I have studied the inner workings of V-ray in the past and what each number in the render settings means in the render settings,
so I've been running some tests on my own because I think my team's render settings are inefficient.

The 0.002 threshold is just for testing purposes. I know it's unnecessarily high. I just set it this low because it's easier to "visually" compare the amount of noise.

what I wanted to talk about Is it really effective to leave everything to Adaptive Sampling and only adjust the max subdiv and threshold?

There are two reasons for this doubt

• Increasing the Min Subdivision is quite effective.
  • The official documentation recommends leaving Min Subdivision at 1 except in some cases. This is because V-ray's Adaptive Sampling determines the optimal subdivision between Min/Max, and it's wasteful to put in extra samples for pixels where only one sample is sufficient.However, in our actual production scene, not the test scene, i found that when i rendered the scene with min 8/max 24, 0,01 and left the threshold the same, but changed it to min 1/max 24, the render time was reduced, but the noise was very high, and i had to set it to min 1, max 48, threshold 0.005 to get similar results to min 8/max 24, leaving everything to Adaptive Sampling. Unsurprisingly, the render time more than doubled. Note that this test scene consists mostly of flat objects, so I don't think it's the special case mentioned in the docs. (This test was done with V-ray 6.)
• Adjusting the Local Subdiv was also quite useful.
  • According to the post I linked to in the text, adjusting Local Subdiv has become an outdated approach since V-ray 3.3, but I was still able to save quite a bit of render time with this approach in V-ray 3.6. However, since this is not possible in V-ray 6, I wanted to compare the "outdated" approach in 3.6 with the approach in V-ray 6. The result is that the old school approach in V-ray 3.6 still seems to be more valid, hence this post.

[Neilg]

My point is that vray isnt optimised around getting noise to a threshold as low as 0.002 in a near empty cube.

When you compare a slightly higher consistent level of noise across an entire render of a much more complex scene, you'll find 6 faster at getting more consistent results.

[glorybound]

I use all defaults, and I don't have any issues. I use denoiser, but at a shallow setting (.2-.4). I was rendering using the progressive mode and would let it render for an hour and kill it. It has noise, but nobody cares. I upgraded my hardware and now use bucket mode and let it render for 3 hours, and there is no noise. Nobody has commented on the new lack of noise, so I think we care about noise more than the client.