This repository contains our implementation of our research paper "Statistical Error Reduction for Monte Carlo Rendering" [Sakai et al. 2025]. Our implementation is based on StatMC, itself built on pbrt-v3, and offers the following:
- a new unidirectional path-tracing integrator
StatPathIntegrator, which tracks the required statistics for multi-transform denoising, variance denoising, adaptive sampling, and linear explained-variance correction, described in our paper, - OpenCV integration for buffer management and CUDA abstraction,
- support for our CUDA denoiser implemented on top of OpenCV (hosted on a separate repository),
- albedo lookup tables for faster and more accurate albedo queries (compared to pbrt-v3's own
rho()function), and - support for the tev image viewer.
The extensions are mostly implemented in src/statistics/ and src/display/.
With the focus on research, this code is not intended for production. We appreciate your feedback, questions, and reports of any issues you encounter; feel free to contact us!
We developed our denoiser using CUDA 12.3 and OpenCV 4.8.1. Note that later CUDA versions (>= 12.4) are incompatible with OpenCV 4.8.1.
For reproducing the results presented in our paper, we recommend using Clang 16.0.6 on Ubuntu 22.04 LTS or Linux Mint 20 (as used for the paper). While we have successfully tested GCC 11.4.0, it produces slightly different results (mostly due to differences in random number generation).
In the following, we describe two alternative ways to build our code: an automatic approach tested for Ubuntu 22.04 LTS and a manual approach, which we recommend if you want to retrace the steps of the build process or use another operating system.
-
Clone this repository (OpenCV will be cloned automatically as a submodule):
git clone --recursive https://github.com/cg-tuwien/StatER.git cd StatER/ -
Install dependencies:
sudo ./scripts/_install-dependencies.sh
-
Build our code:
./scripts/_build.sh
Our version of the pbrt binary should now be located in
build-asm+mcl/pbrt-v3/.
We have prepared files required to run our project within a Docker container here.
Skip this if you have used the automatic approach above.
-
Clone this repository (OpenCV will be cloned automatically as a submodule):
git clone --recursive https://github.com/cg-tuwien/StatER.git cd StatER/ -
Make sure that CUDA 12.3, as well as the packages
cmake,clang,libstdc++-12-dev, andzlib1g-dev, are installed. The packages may vary depending on the operating system.
-
In the root directory of the repository, create the directories for building OpenCV:
mkdir build-asm+mcl cd build-asm+mcl/ mkdir opencv cd opencv/
-
Build OpenCV according to these instructions using the directories
../../src/ext/opencvand../../src/ext/opencv_contribfor<opencv_source_directory>and<opencv_contrib>. We build OpenCV and pbrt separately to have better control over the individual builds. -
Change to the
build-asm+mcl/directory for building pbrt in the next step:cd ../
-
In the previously created
build-asm+mcl/directory, create the build directory for pbrt-v3:mkdir pbrt-v3 cd pbrt-v3/ -
Create the CMake buildsystem:
cmake \ -DOpenCV_BUILD_DIR_PREFIX="build-asm+mcl" \ -DCMAKE_C_COMPILER=clang \ -DCMAKE_CXX_COMPILER=clang++ \ -DCMAKE_CUDA_COMPILER=/usr/local/cuda/bin/nvcc \ -DCMAKE_CUDA_HOST_COMPILER=/usr/bin/clang++ \ -DCMAKE_C_FLAGS="${CMAKE_C_FLAGS} -march=native" \ -DCMAKE_CXX_FLAGS="${CMAKE_CXX_FLAGS} -march=native" \ ../../
-
Build pbrt-v3
make -j 16
The directory scenes/ contains configurations and scene description files for reproducing the results in our paper.
We do not include the complete scenes; they can be downloaded by running the scripts/_download-scenes.sh shell script from the root directory of the repository:
./scripts/_download-scenes.shNote that the scenes are hosted on creator's or licensee's websites and are subject to being changed or taken down without prior notice: we are not responsible for (the availability of) the content hosted on other websites (except for pbrt-v3's measure-one scene, which we rehost for your convenience).
Once the scenes are downloaded, you can reproduce the results from our paper by simply running the shell scripts for the corresponding figures in the scripts/ directory. E.g.,
./scripts/1-veach-bidir.shResult images will be saved to the out/ directory.
The prefixes (here, 1) refer to the figure number in the paper and supplementary document.
The images generated by these scripts should generally match our paper results, which we have published in lossless format here (including references).
Image filenames indicate the number of samples per pixel (SPP) used for rendering.
To reproduce a figure, pbrt must be run up to the corresponding SPP value.
By default, the generated images correspond to the "MCL" variant from the paper.
To enable adaptive sampling ("ASM+MCL"), pass the as option to the script:
./scripts/1-veach-bidir.sh asYou can also build alternative variants of our code:
./scripts/alternate-builds/build-asm+m.sh # "[ASM+]M" built into build-asm+m/
./scripts/alternate-builds/build-asm+mc.sh # "[ASM+]MC" built into build-asm+mc/
./scripts/alternate-builds/build-asn+mcl.sh # "[ASN+]MCL" built into build-asn+mcl/
./scripts/alternate-builds/build-mt.sh # "MT{1-5}" built into build-mt/
./scripts/alternate-builds/build-mt-jb.sh # "MT{1–5}-JB" built into build-mt-jb/Refer to our paper and supplementary document for detailed explanations of these variants.
Once an alternate build has been created, you can specify it as an additional parameter. Here are a few examples:
./scripts/12-furball.sh as build-asm-mc # ASM-MC
./scripts/13-car.sh "" build-asm-m # M
./scripts/S6-lamp.sh mt1 build-mt # MT1 (MCL)
./scripts/S8-house.sh mt5-as build-mt-jb # MT5-JB (ASM-MCL)Feel free to experiment with different scenes and configurations. For a starting point, refer to the quick reference further below.
We reproduced the results in our paper by following the instructions on this page using two machines:
- a desktop PC equipped with an AMD Ryzen 9 5950X CPU, an NVIDIA RTX 3080 Ti GPU, and running Linux Mint 20, as well as
- a virtual machine hosted on a computing cluster equipped with an AMD EPYC 7413 CPU, an NVIDIA A40 GPU, and running Ubuntu 22.04 LTS.
Despite following our instructions for reproducibility, differences in hardware, operating systems, compilers, and other factors may still lead to minor variations in the generated images.
In this repository, we do not include implementations of the neural denoisers we compared against in our paper. For those comparisons, we have used the commits linked here:
For Moon et al.'s confidence-interval approach [2013] and StatMC [Sakai et al. 2024], we used the implementation and instructions provided in the StatMC repository.
With the scripts provided above, it is possible to generate all main results of our publication. If you wish to reproduce the results of our ablation studies, we have documented the required compilation settings here. Please also note the comments for the individual compilation flags here. If you need assistance with these steps, we are happy to help—please contact us!
Our version of the pbrt executable extends the original with the following options:
| Option | Description |
|---|---|
--writeimages |
Write images to disk. |
--displayserver <socket> |
Write images to the specified network socket (format <IP address>:<port number>). |
--baseseed <num> |
Use the specified base seed for RandomSampler. |
--denoise |
Skip rendering and use prerendered images on disk instead (useful for performing multiple denoising passes without rerendering). |
--warmup |
Perform a warm-up iteration (useful for consistent performance measurements). |
Most of the configuration is done in the scene description files. In the following, we provide an overview over our extensions to the original scene description format.
We have extended the original format with options for our StatPathIntegrator.
To illustrate, here is an example configuration, which utilizes all relevant options:
Integrator "statpath"
"integer maxdepth" [65]
"bool expiterations" ["true"]
"bool outputexpiterations" ["true"]
"integer iterations" [13]
"bool adaptivesampling" ["false"]
"bool denoisefilm" ["true"]
"bool calcstats" ["false"]
"bool calcstatmcstats" ["false"]
"bool calcprodenstats" ["false"]
"bool calcmoonstats" ["false"]
"bool calcgbuffers" ["false"]
"float varcizvalue" [2.80703]
"float filtersd" [10]
"integer filterradius" [20]
"string varfilterbuffers" ["albedo" "normal"]
"float varfilterbuffersds" [0.02 0.1]
"string meanfilterbuffers" ["albedo" "normal"]
"float meanfilterbuffersds" [0.02 0.1]
"string outputregex" ["film|filmD"]
The following table summarizes all available options for our StatPathIntegrator:
| Type | Name | Default Value | Description |
|---|---|---|---|
| integer | maxdepth |
5 |
Same as in the original: "Maximum length of a light-carrying path sampled by the integrator." |
| integer[4] | pixelbounds |
(Entire image) | Same as in the original: "Subset of image to sample during rendering; in order, values given specify the starting and ending x coordinates and then starting and ending y coordinates. (This functionality is primarily useful for narrowing down to a few pixels for debugging.)" |
| float | rrthreshold |
1 |
Same as in the original: "Determines when Russian roulette is applied to paths: when the maximum spectral component of the path contribution falls beneath this value, Russian roulette starts to be used." |
| string | lightsamplestrategy |
"spatial" |
Same as in the original: "Technique used for sampling light sources. Options include 'uniform', which samples all light sources uniformly, 'power', which samples light sources according to their emitted power, and 'spatial', which computes light contributions in regions of the scene and samples from a related distribution." |
| bool | expiterations |
true |
Our integrator operates iteratively, with each iteration comprising a rendering and denoising pass. true enables exponential growth of the total number of samples per pixel for rendering (e.g., 4, 16, 64, etc.), while false enables linear growth (e.g., 4, 8, 12, etc.). The (initial) number of samples per pixel (4 in the examples) is specified via the pixelsamples option of the Sampler. |
| bool | outputexpiterations |
expiterations |
true only outputs images for exponential iterations (e.g., 4, 16, 64, etc.), even if expiterations is false. |
| integer | iterations |
16 |
Total number of iterations |
| bool | adaptivesampling |
false |
true enables adaptive sampling. |
| bool | denoisefilm |
false |
true enables denoising of the rendered image. |
| bool | calcstats |
false |
true enables the calculation of G-buffers and statistics required by our denoiser. Use this option to precompute everything required for denoising without performing the denoising itself. |
| bool | calcstatmcstats |
false |
true enables the calculation of G-buffers and statistics required by StatMC [Sakai et al. 2024]. |
| bool | calcprodenstats |
false |
true enables the calculation of G-buffers and statistics required by ProDen [Firmino et al. 2022]. |
| bool | calcmoonstats |
false |
true enables the calculation of G-buffers and statistics required by Moon et al.'s confidence-interval approach [2013]. |
| bool | calcgbuffers |
false |
true enables the calculation of G-buffers required by NVIDIA's OptiX denoiser and Intel's OIDN. |
| float | filtersd |
10.0 |
Standard deviation of the denoising filter kernel |
| integer | filterradius |
20 |
Radius of the denoising filter kernel (limiting the kernel to a finite number of pixels) |
| string[] | varfilterbuffers |
["albedo" "normal"] |
G-buffers for variance denoising; possible options are materialid, depth, normal, albedo. materialid refers to unique numbers that are assigned to different materials by the renderer. For fair comparisons, we used albedos and normals only. |
| float[] | varfilterbuffersds |
[0.02 0.1] |
Standard deviations associated with the G-buffers for variance denoising ( |
| string[] | meanfilterbuffers |
["albedo" "normal"] |
G-buffers for mean denoising; possible options are materialid, depth, normal, albedo. materialid refers to unique numbers that are assigned to different materials by the renderer. For fair comparisons, we used albedos and normals only. |
| float[] | meanfilterbuffersds |
[0.02 0.1] |
Standard deviations associated with the G-buffers for mean denoising ( |
| string | outputregex |
film.* |
Regular expression specifying the buffers to output (to disk or network socket as determined by the --writeimages and --displayserver command-line options); buffers whose unique names match the specified regular expression are output. This way of specification provides a high degree of flexibility, e.g., film.*|t0-.* matches all buffers whose name begins with film or t0-. We provide a complete list of buffers below. |
Similarly to pbrt-v4, our scene description format supports file includes:
Include "../_active.pbrt"
We have implemented this feature to quickly switch between rendering and denoising configurations without changing the scene description file itself.
We provide the following configurations in the scenes/ directory:
| Configuration File | Description |
|---|---|
render-denoise*.pbrt |
Render and denoise using our denoiser |
As can be seen in the scripts for reproducing the figures, a configuration file is activated by overwriting scenes/_active.pbrt with it.
Once a configuration is activated, pbrt can be run normally, supplying the desired scene description file as parameter, e.g.,:
./pbrt ../../scenes/bathroom/scene-stat.pbrtThis section provides an overview of the buffer system in StatPathIntegrator, which enables working with various inputs for and outputs from our denoiser.
Note that a more detailed description goes beyond the scope of this overview; for more details, refer to the code itself or contact us!
There are five buffer types:
| Index | Name | Box-Cox Transformation | Description |
|---|---|---|---|
| 0 | Radiance |
applied | Monte Carlo radiance estimate |
| 1 | StatMaterialID |
not applied | Material ID G-buffer |
| 2 | StatDepth |
not applied | Depth G-buffer |
| 3 | StatNormal |
not applied | Normal G-buffer |
| 4 | StatAlbedo |
not applied | Albedo G-buffer |
The Box-Cox transformation of radiance samples in our multi-transform setting makes our approach more robust to non-normality; details can be found in our paper.
These types are enabled as required by the StatPathIntegrator configuration.
In particular, filterbuffers determines the enabled G-buffer types.
Each enabled type is assigned a consecutively numbered ID (for performance reasons). For instance, if denoising, normals and albedos are enabled, IDs would be assigned as follows:
| ID | Name |
|---|---|
| 0 | Radiance |
| 1 | StatNormal |
| 2 | StatAlbedo |
For each enabled type, a set of buffers is created. Based on the configuration and these rules, the following buffers are potentially created:
| Type | Name | Description |
|---|---|---|
| RGB | film |
Noisy rendered image |
| RGB | filmD |
Denoised rendered image |
| RGB | tX-b0-mean |
Sample mean of transformed samples for type X |
| RGB | tX-b0-m2 |
Sum of squared deviations of transformed samples for type X (division by the number of samples gives the second sample central moment) |
| RGB | tX-b0-m3 |
Sum of cubed deviations of transformed samples for type X (division by the number of samples gives the third sample central moment) |
| RGB | tX-b0-m4 |
Sum of fourth powers of deviations of transformed samples for type X (division by the number of samples gives the fourth sample central moment) |
| RGB | tX-b0-c2X |
Cumulative covariance of transformed samples and pixel positions along x direction for type X (division by the number of samples minus one gives the Bessel-corrected covariance) |
| RGB | tX-b0-c2Y |
Cumulative covariance of transformed samples and pixel positions along y direction for type X (division by the number of samples minus one gives the Bessel-corrected covariance) |
| RGB | tX-b0-m1 |
Total absolute deviation of transformed samples for type X (division by the number of samples gives the mean absolute deviation) |
| RGB | tX-b0-jb |
Jarque–Bera score of transformed samples for type X |
| RGB | tX-b0-lnS2 |
Bonett's ln(s2) of transformed samples for type X |
| RGB | tX-b0-se2 |
Bonett's squared standard error se^2 of transformed samples for type X |
| RGB | tX-b0-m2C |
LEV-corrected sum of squared deviations of transformed samples for type X (division by the number of samples gives the second sample central moment) |
| RGB | tX-b0-varD |
Denoised variance of transformed samples for type X |
| RGB | tX-b0-varCD |
Denoised LEV-corrected variance of transformed samples for type X |
| RGB | tX-b0-discr |
Welch discriminator of transformed samples for type X (used to cache per-pixel discriminators for mean denoising) |
All tX-b0-* buffers above are also available for untransformed samples with the prefix tX-b0-film-*.
In addition, the following buffers are additionally used for untransformed samples:
| Type | Name | Description |
|---|---|---|
| integer | tX-b0-film-n |
Number of samples taken for type X |
| RGB | tX-b0-film-meanVar |
Sample variance of untransformed samples for type X |
| RGB | tX-b0-film-meanD |
Denoised mean of untransformed samples for type X |
Depending on the configuration, some buffers may be disabled, and GPU-computed buffers must be downloaded to the CPU before they can be output.
For mt[-jb] builds, only non-film buffers are used for storing the transformed statistics, with b1 to b8 specifying the index of the transformation.
Covering the effects of every possible configuration is beyond the scope of this README.
For more information, please check the code directly or reach out to us.
The outputregex option provides a convenient way to select output buffers.
StatPathIntegrator supports the BoxFilter only.
We thank Thomas Auzinger for providing LaTeX plugins, José Dias Curto for support with confidence intervals, and Markus Schütz for assistance with the CUDA implementation. We also thank the creators of the scenes we used: Benedikt Bitterli for "Veach, Bidir Room" (Figs. 1, S12), "Cornell Box" (Fig. 2), and "Fur Ball" (Fig. 12); Jay-Artist for "Country Kitchen" (Figs. 4, 5, 10, S2, S10, S16); Mareck for "Contemporary Bathroom" (Figs. 7, 14, S13); thecali for "4060.b Spaceship" (Fig. 9); piopis for "Old Vintage Car" (Fig. 13); Cem Yuksel for "Straight Hair" (Fig. S4) and "Curly Hair" (Fig. S5); UP3D for "Little Lamp" (Fig. S6); axel for "Glass of Water" (Fig. S7); MrChimp2313 for "Victorian Style House" (Fig. S8); NovaAshbell for "Japanese Classroom" (Fig. S9); and Beeple for "Zero-Day" (Fig. S11). Statistical simulation studies were conducted using the Austrian Scientific Computing (ASC) infrastructure. This work has been funded by the Vienna Science and Technology Fund (WWTF) [Grant ID: 1047379/ICT22028]. This research was funded in whole or in part by the Austrian Science Fund (FWF) [10.55776/F77]. For open-access purposes, the author has applied a CC BY public copyright license to any author-accepted manuscript version arising from this submission. The authors acknowledge TU Wien Bibliothek for financial support through its Open Access Funding Programme.