Why is my rendered image so noisy compared to the example in the Mitsuba participating media documentation?

[kiramimi77]

Hi all,

I’m trying to render a participating media scene using Mitsuba 3.
The left image is my rendering result using the original scene files from mitsuba-data,
while the right image is from the Mitsuba documentation/example on participating media:

As you can see, my image contains a lot of noise (possibly Monte Carlo noise).
I’m wondering:

Why is my rendering so noisy compared to the example?

Are there recommended techniques or settings to reduce the noise in this type of scene?

[njroussel]

Hi @kiramimi77

These two images, don't look the same. Have you changed the emitter? I've rendered it with 4k samples (~3 minutes on a 4090) and it barely has any visible noise.

The examples are usually rendered with a fairly high sample count. The spp value in the XML scene description most likely does not match what was used when generating the image for the documentation.

Getting a lot of MC noise in volume rendering is pretty much expected, in part due to the fact that you'll have long light paths. In general, you'd want to have as "uniform" interactions, i.e diffuse BSDFs, isotropic phase functions, uniform emitters, etc. Of course, this strongly alters the actual appearance of the scene, so you have to find some middle-ground. You can also use denoisers (Mitsuba ships with the Optix denoiser), if you're ok with having a little bit of bias.

[kiramimi77]

i got that, appreciate for you help. And I have one more question, this 100k result for SSS cost 4 hour on GPU 4070,

path = "whole_scene.xml"
import numpy as np
import mitsuba as mi
import matplotlib.pyplot as plt
mi.set_variant('cuda_ad_rgb')
scene = mi.load_file(path)
params = mi.traverse(scene)

spp = 4000
npass = 25
for i in range(npass):
if i == 0:
image = mi.render(scene, spp=spp, seed=i) / npass
else:
image += mi.render(scene, spp=spp, seed=i) / npass
plt.figure()
plt.axis("off")
plt.imshow(image ** (1.0 / 2.2)); # approximate sRGB tonemapping
plt.title()
plt.show()

and for the scene setting

<default name="spp" value="128"/>
<default name="resx" value="864"/>
<default name="resy" value="648"/>
<default name="rr_depth" value="64"/>
<default name="max_depth" value="64"/>
<default name="integrator" value="volpathmis"/>
<default name="scale" value="350"/>

<integrator type="$integrator">
	<integer name="rr_depth" value="$rr_depth"/>
	<integer name="max_depth" value="$max_depth"/>
</integrator>

<rfilter type="mitchell"/>

<sensor type="perspective" id="elm__1" name="elm__1">
	<string name="fov_axis" value="x"/>
	<float name="fov" value="30.000004"/>
	<float name="principal_point_offset_x" value="0.000000"/>
	<float name="principal_point_offset_y" value="-0.000000"/>
	<float name="near_clip" value="0.100000"/>
	<float name="far_clip" value="100.000000"/>
	<transform name="to_world">
		<rotate x="1" angle="-45.00000466733367"/>
		<rotate y="1" angle="-3.5418673217217693e-06"/>
		<rotate z="1" angle="-179.999991348578"/>
		<translate value="0.000000 3.000000 -3.000000"/>
	</transform>
	<sampler type="multijitter" name="sampler">
		<integer name="sample_count" value="$spp"/>
	</sampler>
	<film type="hdrfilm" name="film">
		<integer name="width" value="$resx"/>
		<integer name="height" value="$resy"/>
	</film>
</sensor>

<bsdf type="roughdielectric" id="object_bsdf">
    <string name="distribution" value="beckmann"/>
    <float name="alpha" value="0.02"/>
    <string name="int_ior" value="bk7"/>
    <string name="ext_ior" value="air"/>
</bsdf>

<medium id="medium1_blue" type="homogeneous">
    <float name="scale" value="$scale"/>
    <phase type="hg">
		<float name="g" value="0.4"/>
	</phase>
    <float name="sigma_t" value="1"/>
    <rgb name="albedo" value="0.85,0.8,0.98"/>
    <boolean name="has_spectral_extinction" value="false"/>
    <boolean name="sample_emitters" value="false"/>
</medium>

<medium id="medium1_red" type="homogeneous">
    <float name="scale" value="$scale"/>
    <phase type="hg">
		<float name="g" value="0.4"/>
	</phase>
    <float name="sigma_t" value="1"/>
    <rgb name="albedo" value="0.98,0.8,0.8"/>
    <boolean name="has_spectral_extinction" value="false"/>
    <boolean name="sample_emitters" value="false"/>
</medium>


<medium id="medium1_white" type="homogeneous">
    <float name="scale" value="350"/>
    <phase type="hg">
		<float name="g" value="0.4"/>
	</phase>
    <float name="sigma_t" value="1"/>
    <rgb name="albedo" value="0.98,0.98,0.98"/>
    <boolean name="has_spectral_extinction" value="false"/>
    <boolean name="sample_emitters" value="false"/>
</medium>

<emitter type="point" id="elm__999" name="elm__999">
	<point name="position" x="1.0" y="0.0" z="-1.0"/>
	<rgb value="18.0 18.0 18.0" name="intensity"/>
</emitter>

<shape type="ply" id="elm__4" name="elm__4">
	<string name="filename" value="meshes/Plane.ply"/>
	<boolean name="face_normals" value="true"/>
	<bsdf type="diffuse">
		<texture type="checkerboard" id="checker_texture" name="reflectance">
			<rgb name="color0" value="0.1, 0.1, 0.1"/>
			<rgb name="color1" value="0.9, 0.9, 0.9"/>
			<transform name="to_uv">
				<scale x="10" y="10"/>
			</transform>
		</texture>
	</bsdf>
</shape>


<shape type="ply" id="elm__5" name="elm__5">
	<string name="filename" value="meshes/model.ply"/>
	<ref id="medium1_white" name="interior"/>
    <ref id="object_bsdf"/>
</shape>
<shape type="ply" id="elm__7" name="elm__7">
	<string name="filename" value="meshes/model_001.ply"/>
	<boolean name="face_normals" value="true"/>
	<ref id="medium1_blue" name="interior"/>
    <ref id="object_bsdf"/>
</shape>
<shape type="ply" id="elm__9" name="elm__9">
	<string name="filename" value="meshes/model_002.ply"/>
	<boolean name="face_normals" value="true"/>
	<ref id="medium1_red" name="interior"/>
    <ref id="object_bsdf"/>
</shape>

the result still exists lots of monte carlo noise especially on the shadow area, do you think it's caused by complexity of scene or there is something wrong with my xml setup? Your insights will help a lot.

[njroussel]

I didn't spot anything particularly wrong with your XML, so I'd say that it's simply your scene.

[kiramimi77]

Thanks. I will simply and take a look at re-render results.

[kiramimi77]

Hi @njroussel ,
I have a question regarding the difference in rendering results between two images. The first image was rendered without a point emitter, while the second one includes a point emitter—all other scene settings are exactly the same.
However, after adding the point emitter, I noticed that the subsurface scattering scene becomes significantly noisier, even though no other parameters were changed. I'm wondering why the presence of this point light source introduces so much noise in the rendering.

<emitter type="point" id="elm__999" name="elm__999">
	<point name="position" x="1.0" y="0.0" z="-1.0"/>
	<rgb value="5" name="intensity"/>
</emitter>


<emitter type="envmap" id="Area_002-light">
	<string name="filename" value="museum.exr"/>
	<transform name="to_world">
		<rotate y="1" angle="-180"/>
		<matrix value="-0.224951 -0.000001 -0.974370 0.000000 -0.974370 0.000000 0.224951 0.000000 0.000000 1.000000 -0.000001 8.870000 0.000000 0.000000 0.000000 1.000000 "/>
	</transform>
	<float name="scale" value="1"/>
</emitter>

[njroussel]

Hi @kiramimi77

For each light path that connects the sensor to the point light, let's just consider the last segment (surface to emitter). This segment can create very noisy BSDF weights. Think about any light path that is exiting one of the sub-surface scattering media: each light path wants to be scattered towards a certain direction (dictated by the BSDF lobe of the enclosing shape). However, by inserting a point light, you effectively force that direction to point to the emitter in order to create a valid light path. This "forced" direction might have a very small probability according to the BSDF. It's the large variance in these weights that you're most likely seeing. Some light paths barely have any contribution whereas other contribute a lot. I would imagine that making your BSDFs rougher should reduce the noise level.