The Arnold denoiser is a stand-alone program post-process denoiser executable. It works on EXR files with multiple layers and outputs an EXR file with the denoised layers. It takes into account multiple frames and multiple light AOVs. It requires variance information for all AOVs and optionally uses normal, depth, and albedo. Like most denoisers, it considers a neighborhood around the current pixel and looks for similar neighborhoods inside a given search area to blend together.
Required Input Layers
The required absolute minimum inputs are the following:
- main layer to be denoised (typically RGBA also known as "beauty")
- variance layer that specifies the per pixel variance
It's recommended to set "preserve layer name" to ON in the output drivers so that the Arnold denoiser can automatically detect what each AOV corresponds to.
Optional Feature Layers
These optional layers are used to guide denoising and will greatly improve the result:
- Normal (N)
- Depth (Z)
- Albedo (denoise_albedo)
Optional Light AOVs to be Co-Denoised
Additional light AOVs can be present in the original EXR. If requested (
-l light_aov_name) these will be co-denoised with the beauty layer. This main denoising effort will be spent on beauty, so in certain areas, the light AOVs might have increased noise.
Optional Additional Frames
Multiple additional frames can be specified to increase temporal stability (for instance
-i render.002.exr -i render.001.exr -i render.003.exr)
Instead of specifying all the input images, it's also possible to use the following flags:
-f n or
--frames n) to denoise image sequences. Another argument (
-ef n or
--extraframes n) specifies how many additional source frames before and after the current one should be used, for improved stability in animation sequences.
In this example, the Arnold denoiser will run for 10 consecutive frames starting at frame 5, taking into account two frames before and two frames after each source frame (e.g. source frame number 5 will take into account frames 3, 4, 5, 6, and 7):
Pixel Neighborhood Patch Radius
This increases the softness of the denoising (while always preserving the features). The default is 3 (on the big side), but a low value would be maybe 0 or 1, middle 3 and high 5.
For every pixel, the Arnold denoiser will consider its neighborhood patch and look for other pixels with similar neighborhood patches. The radius of this neighborhood can be controlled with the
-pr) command-line argument. The default value is set to 3, which gives a 7x7 square neighborhood.
Pixel Search Radius
noicewill search a square area with a radius set with the command line argument
-sr). The bigger this area the bigger the denoising stability and the higher the chance that similar neighborhoods to be considered will be found. The default value is 9, which gives a 19x19 square neighborhood. Setting it to 21 (a search window of 42 x 42) will look over a pixel area equivalent to loading 5 frames.
-v). The default value is 0.5, higher values will make the denoising more aggressive by considering similar neighborhoods that have bigger color disparities.
noicecan be specified with the command line argument
-i). Multiple frames can also be specified around the frame to be denoised in the following way:
-i <current_frame> -i <additional frames in order>
-i render.0010.exr -i render.0008.exr -i render.0009.exr -i render.0011.exr -i render.0012.exr
-l). for instance:
-l back_lights -l RGBA -l key_light
Custom Feature Layers
--features) specifies which AOVs will be used as features to guide denoising.
Ass File Example
These additional AOV are optional, but the variance info for RGBA and additional AOVs to be denoised is not. The denoiser will work much better if N, Z, and denoise_albedo are provided:
outputs 7 1 STRING
"RGBA RGBA filter driver"
"RGB RGB variance driver RGB_variance"
"LPE_name RGBA filter driver"
"LPE_name RGB variance driver LPE_name_variance"
"denoise_albedo RGB filter driver albedo"
"N RGB defaultArnoldFilter@gaussian_filter driver"
"Z RGB filter driver"
Notice that the AOVs that have the same source also have an EXR layer name specified at the end. The name of the layers is not important, but appending "_variance" seems like a clearer way to handle this.
The filters should be set up like this (here we use
gaussian_filter but it could be any other filter as long as it is matched in