Area Lights and Spotlights very noisy near source in Volumes #87565

New Issue

Joe Williamsen · 2021-04-16T18:07:53+02:00

Joe Williamsen commented

2021-04-16 18:07:53 +02:00

System Information
Operating system: Windows-10-10.0.19041-SP0 64 Bits
Graphics card: NVIDIA GeForce RTX 2070/PCIe/SSE2 NVIDIA Corporation 4.5.0 NVIDIA 465.89

Blender Version
Broken: version: 2.93.0 Beta, branch: master, commit date: 2021-04-15 17:14, hash: fa8d566c3b
Worked: Unknown

Short description of error
If you have area lights or spotlights in a volume, they render very "noisy" near the light source.

Exact steps for others to reproduce the error
Open the attached blend file, render with F12.

Volume Scattering Bug.rar

**System Information** Operating system: Windows-10-10.0.19041-SP0 64 Bits Graphics card: NVIDIA GeForce RTX 2070/PCIe/SSE2 NVIDIA Corporation 4.5.0 NVIDIA 465.89 **Blender Version** Broken: version: 2.93.0 Beta, branch: master, commit date: 2021-04-15 17:14, hash: `fa8d566c3b` Worked: Unknown **Short description of error** If you have area lights or spotlights in a volume, they render very "noisy" near the light source. ![volume scattering.jpg](https://archive.blender.org/developer/F9936152/volume_scattering.jpg) **Exact steps for others to reproduce the error** Open the attached blend file, render with F12. [Volume Scattering Bug.rar](https://archive.blender.org/developer/F9936149/Volume_Scattering_Bug.rar)

Joe Williamsen commented

2021-04-16 18:07:53 +02:00

Added subscriber: @Joe_W

Alaska commented

2021-04-18 03:55:24 +02:00

Added subscriber: @Alaska

Alaska commented

2021-04-18 03:55:24 +02:00

Changed status from 'Needs Triage' to: 'Archived'

Alaska closed this issue

2021-04-18 03:55:24 +02:00

Alaska commented

2021-04-18 03:55:24 +02:00

Testing on my computer, the noise pattern is the same as the above render in the versions I tested (2.90, 2.91, 2.92, 2.93, 3.0). The area light was different in older versions (due to the area light spread introduced in 2.93).

The behavior being the same between these versions suggestions this is not a regression and is thus technically not a bug. In theory, noise can be improved with a different sampling methods, however this would be a feature request which is not tracked on developer.blender.org. Please use other channels for user feedback and feature requests: https://wiki.blender.org/wiki/Communication/Contact#User_Feedback_and_Requests

In the meantime, increasing the sample count and using denoising should resolve this issue.

Testing on my computer, the noise pattern is the same as the above render in the versions I tested (2.90, 2.91, 2.92, 2.93, 3.0). The area light was different in older versions (due to the area light spread introduced in 2.93). The behavior being the same between these versions suggestions this is not a regression and is thus technically not a bug. In theory, noise can be improved with a different sampling methods, however this would be a feature request which is not tracked on developer.blender.org. Please use other channels for user feedback and feature requests: https://wiki.blender.org/wiki/Communication/Contact#User_Feedback_and_Requests In the meantime, increasing the sample count and using denoising should resolve this issue.

Joe Williamsen commented

2021-04-18 05:31:08 +02:00

As far as increasing samples and adding denoising, the example scene (and render provided) was set to use 2048 samples. Initially, I, too, thought that I'd have to increase samples to get this to render cleanly, but even after going up to 6000 samples, the same "chunky" appearance occurs around the light source.

What DID seem to reduce the blotchy noise pattern was increasing the number of Light Path>Max Bounces>Volume from the default 0 to 1 or more ... but this introduces problems of it's own (specifically, an overall "mushy/blotchy" result to the entire volume).

I'm not sure how this is a feature request as opposed to a bug. On it's face it seems like undesirable behavior. Are you suggesting that the proper approach would be to put in a request for "better volumetric lights" - or something to that effect?

As far as increasing samples and adding denoising, the example scene (and render provided) was set to use 2048 samples. Initially, I, too, thought that I'd have to increase samples to get this to render cleanly, but even after going up to 6000 samples, the same "chunky" appearance occurs around the light source. What DID seem to reduce the blotchy noise pattern was increasing the number of Light Path>Max Bounces>Volume from the default 0 to 1 or more ... but this introduces problems of it's own (specifically, an overall "mushy/blotchy" result to the entire volume). I'm not sure how this is a feature request as opposed to a bug. On it's face it seems like undesirable behavior. Are you suggesting that the proper approach would be to put in a request for "better volumetric lights" - or something to that effect?

Alaska commented

2021-04-18 10:46:29 +02:00

Doing a quick test, I was able to get a relatively "clean" result with 8,192 samples, and it gets cleaner as the samples are increase. However this was with adaptive sampling off. Having adaptive sampling on causes issues with the 8,192 sample scene. This means the blotchy noise pattern comes from a combination of two things.

A low sample count, which is to be expected. Things are naturally noisy in Cycles with complex shaders (especially volumes) and small lights (E.G. The small cone of light from the area and spot lights) without a large number of samples.
Adaptive sampling pre-maturely stopping sections of a render when it shouldn't.

So, this is a bug with adaptive sampling then? No. Adaptive sampling determines whether or not something should continue to be sampled based on the noise in certain areas of the render. As it just so happens, in this scene with 2,048 samples, enough samples occur without actually sampling the light in certain areas, resulting in no noise, and as a result the adaptive sampler stops sampling that area. This results in "blotchs" around the light that is black from a lack of samples of a light source. This is expected behavior. The adaptive sampler is detecting noise properly and stopping at the right time, it's just that Cycles hasn't sampled the scene how you'd expect and noise in these regions hasn't been generated before the adaptive sampler decides to stop it.

As a result of these types of issues occurring, the Blender developers exposed an option called Minimum samples that allows you to determine when adaptive sampling should make its first check for noise. Leaving this Minimum samples value at 0 lets Blender pick a number automatically (I believe it's just the square root of your samples, which for a 2048 sample scene is roughly 45 samples), but you can change it to something better suited for your scene (E.G. 512, or even 1024), but keep in mind that increasing this number will increase your render times.

Now, you might still be getting small issues, even with a minimum sample count of 1024. And this is simply down to the fact that cycles isn't that great at sampling this type of scene, so it's still hard to generate noise, even at a large minimum sample count. So it might be best for you just to disable adaptive sampling in this scene. Yes, you'll end up with longer render times, but that's the trade off you need to make for quality from a path tracer like Cycles.

So, the bug is technically in Cycles then, and how it samples lights? Yes. But it's not considered a bug, more of a known issue. Path tracing these types of scenes with a high level of quality and quickly is hard. As a result, various path tracing algorithms exist. Blender even has two path tracing options, Path tracing and Branched Path Tracing, both excelling in different scenarios. Many of the shaders and lights in Blender also have options that change little things about how they're sampled to reduce noise. The issue you're experiencing is that one of the path tracing algorithms isn't efficient enough in this scene to generate noise for the adaptive sampler. However, the fact that algorithm is working means this is not a bug. And as such, it technically makes your report a "request for a better sampling algorithm".

In the meantime, may I make some suggestions for resolving this issue?

You can go down the route of increasing your sample count and Minimum sample count for the adaptive sampler.
The other option is to switch to CPU rendering. CPU rendering produces much better results in a fewer number of samples, and as a result generates more noise early on allowing the adaptive sampler with Minimum samples set to be automatic to work as expected. You'll find an explanation for why switching to CPU helps in this case below:

Volumetric objects have support for Multiple Importance Sampling which allows Cycles to automatically pick better sampling patterns depending on a bunch of factors. This feature allows for things such as better sampling of lights in a volume (what's happening in your scene). However, this feature is currently NOT supported on GPUs and thus you must use your CPU for rendering to take advantage of it. All you have to do is change the render device to CPU and select Render. Multiple Importance Sampling is enabled by default for all objects.

Doing a quick test, I was able to get a relatively "clean" result with 8,192 samples, and it gets cleaner as the samples are increase. **However** this was with adaptive sampling off. Having adaptive sampling on causes issues with the 8,192 sample scene. This means the blotchy noise pattern comes from a combination of two things. 1. A low sample count, which is to be expected. Things are naturally noisy in Cycles with complex shaders (*especially volumes*) and small lights (*E.G. The small cone of light from the area and spot lights*) without a large number of samples. 2. Adaptive sampling pre-maturely stopping sections of a render when it shouldn't. So, this is a bug with adaptive sampling then? No. Adaptive sampling determines whether or not something should continue to be sampled based on the noise in certain areas of the render. As it just so happens, in this scene with 2,048 samples, enough samples occur without actually sampling the light in certain areas, resulting in no noise, and as a result the adaptive sampler stops sampling that area. This results in "blotchs" around the light that is black from a lack of samples of a light source. This is **expected** behavior. The adaptive sampler is detecting noise properly and stopping at the right time, it's just that Cycles hasn't sampled the scene how you'd expect and noise in these regions hasn't been generated before the adaptive sampler decides to stop it. As a result of these types of issues occurring, the Blender developers exposed an option called `Minimum samples` that allows you to determine when adaptive sampling should make its first check for noise. Leaving this `Minimum samples` value at 0 lets Blender pick a number automatically (*I believe it's just the square root of your samples, which for a 2048 sample scene is roughly 45 samples*), but you can change it to something better suited for your scene (E.G. 512, or even 1024), but keep in mind that increasing this number will increase your render times. ![Minimum samples - adaptive sampling.png](https://archive.blender.org/developer/F9952012/Minimum_samples_-_adaptive_sampling.png) Now, you might still be getting small issues, even with a minimum sample count of 1024. And this is simply down to the fact that cycles isn't that great at sampling this type of scene, so it's still hard to generate noise, even at a large minimum sample count. So it might be best for you just to disable adaptive sampling in this scene. Yes, you'll end up with longer render times, but that's the trade off you need to make for quality from a path tracer like Cycles. So, the bug is technically in Cycles then, and how it samples lights? Yes. But it's not considered a bug, more of a known issue. Path tracing these types of scenes with a high level of quality and quickly is hard. As a result, various path tracing algorithms exist. Blender even has two path tracing options, `Path tracing` and `Branched Path Tracing`, both excelling in different scenarios. Many of the shaders and lights in Blender also have options that change little things about how they're sampled to reduce noise. The issue you're experiencing is that one of the path tracing algorithms isn't efficient enough in this scene to generate noise for the adaptive sampler. However, the fact that algorithm is working means this is not a bug. And as such, it technically makes your report a "request for a better sampling algorithm". In the meantime, may I make some suggestions for resolving this issue? 1. You can go down the route of increasing your sample count and `Minimum sample` count for the adaptive sampler. 2. The other option is to switch to CPU rendering. CPU rendering produces much better results in a fewer number of samples, and as a result generates more noise early on allowing the adaptive sampler with `Minimum samples` set to be automatic to work as expected. You'll find an explanation for why switching to CPU helps in this case below: > Volumetric objects have support for `Multiple Importance Sampling` which allows Cycles to automatically pick better sampling patterns depending on a bunch of factors. This feature allows for things such as better sampling of lights in a volume (what's happening in your scene). However, this feature is currently NOT supported on GPUs and thus you must use your CPU for rendering to take advantage of it. All you have to do is change the render device to CPU and select `Render`. `Multiple Importance Sampling` is enabled by default for all objects.

Joe Williamsen commented

2021-04-23 18:16:55 +02:00

Alaska - very sorry about the late reply - I didn't see an email informing me of your response.

Thank you very much for the detailed explanation. I don't think I've ever rendered with 8K samples, but I guess it's good to know that might be what is required. I've played with the two variants of Path Tracing and had mixed results. My first response with most renders that are too noisy is to throw more samples at them - or up the specific number of samples of the particular function/feature (Glossy, Diffuse, etc).

I'll also have to play more with the Minimum Samples setting. I can hear my GPU screaming in agony already .... ;)

Thanks again for the detailed response - I appreciate it.

(And after the Live-Stream today .... we have Cycles-X to look forward to ....)

Alaska - very sorry about the late reply - I didn't see an email informing me of your response. Thank you very much for the detailed explanation. I don't think I've ever rendered with 8K samples, but I guess it's good to know that might be what is required. I've played with the two variants of Path Tracing and had mixed results. My first response with most renders that are too noisy is to throw more samples at them - or up the specific number of samples of the particular function/feature (Glossy, Diffuse, etc). I'll also have to play more with the Minimum Samples setting. I can hear my GPU screaming in agony already .... ;) Thanks again for the detailed response - I appreciate it. (And after the Live-Stream today .... we have Cycles-X to look forward to ....)

Sign in to join this conversation.

No Label

Download

What's New

Blender Studio

Manual

Developers Blog

Documentation

Benchmark

Blender Conference

Development Fund

One-time Donations

Area Lights and Spotlights very noisy near source in Volumes #87565