Alain Galvan ·Sep 10, 2019 · Updated 2 years ago
A review of different techniques for denoising monte-carlo images with machine learning. Learn about autoencoders and currently available implementations such as OIDN and the Optix Autoencoder.
Tags: blogdraftwipmachine learningneural networksautoencodertheorygraphicsdenoisingoptixodin
Machine Learning has had years of literature focused on denoising. While it may have started as a solution to the problem of classifying noisy text [Gallinari et al. 1987] [G. Mixon et al. 2018], eventually the problem domain was expanded on to include more computer graphics focused problems such as denoising monte-carlo path traced images [Khademi Kalantari et al. 2013] [Bako et al. 2017] [Vogels et al. 2018] [Laine et al. 2019], real time ray tracing denoising [Khademi Kalantari et al. 2015] [R. Alla Chaitanya et al. 2017], improving ray tracing sampling [Hasselgren et al. 2020] [Muller et al. 2021], general image reconstruction [Mildenhall et al. 2017] [Lehtinen et al. 2018], and upscaling images while maintaining detail through super-resolution/super-sampling [Dong et al. 2015] [Ledig et al. 2016] [Wang et al. 2018] [Xiao et al. 2020].
Industry leaders such as Intel and NVIDIA have sponsored research in machine learning based denoisers, Intel Open Image Denoise and the NVIDIA Optix Autoencoder both use a denoising autoencoder to denoise images to great success.
The difficulty in using these lies in their computational cost. Without dedicated hardware or trade offs in quality and performance, it's is difficult to achieve real time high quality denoising using autoencoders.
Thus there's been research in the area of algorithmic techniques to denoise images in real time using spherical harmonics to encode low frequency data, guided filters to blur while avoiding sharp changes in normals or albedo information, and spatio-temporal reprojection to reuse data that isn't view dependent such as global illumination or shadows. [Schied et al. 2019]
You can find more information about these techniques in my post on Realtime Ray Tracing Denoising.
 1spp Input |  Optix |
|---|---|
 Ground Truth (10K spp) |  OIDN |
There's even ways of combining these two methods, with [Hasselgren et al. 2020] using spatio-temporal methods to feed a sample map estimator network to help dictate how many rays a renderer should dispatch. This helps solve the issue of denoising salt/peppering highlights/shadows in real time. It's also possible to use a different denoising scheme depending on the number of samples in your input such as in [Meng et al. 2020] where 64 spp images were used as input for one of their networks.
Nevertheless It's clear that with images with little history information or with sparse low frequency information, using machine learning results in significantly better quality outputs than algorithmic alternatives such as A-SVGF or custom implementations such as Quake 2/Minecraft RTX's use of split output buffers, upscaling, and spherical harmonics encoding for low frequency data.
It's likely that neural network based denoising solutions will either replace algorithmic techniques or the two will be used in tandem depending on user requirements.
We'll be reviewing machine learning techniques for denoising monte-carlo images, and looking at the state of the art in this area of research.
An Autoencoder is a neural network that is trained to attempt to imperfectly copy its input to its output. A Denoising Autoencoder take one step further from copying, attempting to learn how to undo the corruption of its input. [Goodfellow et al. 2016]
[R. Alla Chaitanya et al. 2017] makes use of a noisy LDR input, 3 component signed normals, depth, and roughness to train the autoencoder, with its lack of an albedo channel most likely the cause of its weakness in estimating the final color of its output. More recent techniques such as Intel Open Image Denoise and NVIDIA Optix using a noisy HDR input, 3 component signed normals, and albedo.
 |
|---|
| Comparison of an autoencoder trained without and with a GAN by Xu et al. Note sharpness on right. |
The most recent state of the art [Li 2020] points to a network being passed the diffuse and specular terms separately like Quake 2 RTX's A-SVGF, and providing auxiliary buffers (Normals, Albedo, Depth) with element wise biasing and scaling of weights. There's a variety of approaches that can be taken to both train and execute the model. Recent literature points to the use of generative adversarial networks (GANs) [Xu et al. 2019] in combination with autonomous reinforcement learning.
Google TensorFlow, next to Facebook PyTorch, is the most popular neural network framework, so we'll be using TensorFlow to quickly design and train our neural network, then converting our model to DirectML's hardware accelerated machine learning operations, and loading the same weights we calculate with TensorFlow.
GPU accelerated tensor operations isn't necessarily new, with industry leaders like Apple exposing a Metal based Machine Learning API in Metal 2, projects like TensorFire and the paper Growing Neural Cellular Automata showcasing the execution of tensors with WebGL, and thesis papers on writing compute shaders to train and execute models on the GPU. With the advent of dedicated tensor cores on modern GPUs, this is now even easier to do in real time.
We'll be building an autoencoder that takes in the noisy input texture, albedo, view depth, and view normals (with 1st bounce), metalness and gloss (for a total of 12 components), and outputs a single RGBA32 output. This will then be trained using supervised learning on a dataset of monte carlo ray traced images to produce a trained output.
These pixels with 12 components are passed to a series of 2D convolutional neural network layers connected with Leaky Rectifier Linear Activation Units (Leaky RLU) which connect with chunks of the input until it's reduced to a small tensor, then the reverse occurs where convolutions are scaled to a single output.
For instance, Intel OIDN opts to:
This design isn't exclusive to OIDN, NVIDIA's recurrent autoencoder also has 5 downsample convolutions and 5 upsample convolutions, as does Xu et al and Hasselgren et al.
Let's discuss the design of the network, its primitives, and why they're used where they are.
You can find out more about convolutional layers and primitives in Machine Learning in the TensorFlow Guide.
# 🍱🚿 TensorFlow
import tensorflow as tf
from tensorflow.python.keras import layers
layers.Conv2D(filters, kernel_size, ...)// ❌🧠 DirectML
struct DML_CONVOLUTION_OPERATOR_DESC
{
const DML_TENSOR_DESC* InputTensor;
const DML_TENSOR_DESC* FilterTensor;
_In_opt_ const DML_TENSOR_DESC* BiasTensor;
const DML_TENSOR_DESC* OutputTensor;
DML_CONVOLUTION_MODE Mode;
DML_CONVOLUTION_DIRECTION Direction;
UINT DimensionCount;
_In_reads_(DimensionCount) const UINT* Strides;
_In_reads_(DimensionCount) const UINT* Dilations;
_In_reads_(DimensionCount) const UINT* StartPadding;
_In_reads_(DimensionCount) const UINT* EndPadding;
_In_reads_(DimensionCount) const UINT* OutputPadding;
UINT GroupCount;
_In_opt_ const DML_OPERATOR_DESC* FusedActivation;
};A node that attaches to all subsequent nodes. A Convolution is a function that takes in all the inputs of the previous tensor and connects them with all your specified outputs according to each layer's shape.
This is analogous to a guassian filter in computer graphics. We use this operation to compute weights across a large kernel, similar to how guassian filters compute the average color of a kernel. Operations such as bluring and sharpening an image are convolutions, a matrix operation around a central pixel.
/**
* GLSL Guassian Blur Convolution Example
* By csblo from 🌎 Wikipedia (ShaderToy)
* https://en.wikipedia.org/wiki/Kernel_(image_processing)#Concrete_implementation
* 🎓 Slightly modified for this example:
*/
#define gaussian_blur mat3(1, 2, 1, 2, 4, 2, 1, 2, 1) * 0.0625
// 🟧 Find coordinate of matrix element from index
vec2 kpos(int index)
{
return vec2[9] (
vec2(-1, -1), vec2(0, -1), vec2(1, -1),
vec2(-1, 0), vec2(0, 0), vec2(1, 0),
vec2(-1, 1), vec2(0, 1), vec2(1, 1)
)[index] / iResolution.xy;
}
// 🟥 Extract region of dimension 3x3 from sampler centered in uv
// sampler : texture sampler
// uv : current coordinates on sampler
// return : an array of mat3, each index corresponding with a color channel
mat3[3] region3x3(sampler2D sampler, vec2 uv)
{
// ⚪ Create each pixels for region
vec4[9] region;
for (int i = 0; i < 9; i++)
region[i] = texture(sampler, uv + kpos(i));
// 💠 Create 3x3 region with 3 color channels (red, green, blue)
mat3[3] mRegion;
for (int i = 0; i < 3; i++)
mRegion[i] = mat3(
region[0][i], region[1][i], region[2][i],
region[3][i], region[4][i], region[5][i],
region[6][i], region[7][i], region[8][i]
);
return mRegion;
}
// 🕸️ Convolve a texture with kernel
// kernel : kernel used for convolution
// sampler : texture sampler
// uv : current coordinates on sampler
vec3 convolution(mat3 kernel, sampler2D sampler, vec2 uv)
{
vec3 fragment;
// 💉 Extract a 3x3 region centered in uv
mat3[3] region = region3x3(sampler, uv);
// 🔴🟢🔵 for each color channel of region
for (int i = 0; i < 3; i++)
{
// 🔴 get region channel
mat3 rc = region[i];
// ∏ component wise multiplication of kernel by region channel
mat3 c = matrixCompMult(kernel, rc);
// Σ add each component of matrix
float r = c[0][0] + c[1][0] + c[2][0]
+ c[0][1] + c[1][1] + c[2][1]
+ c[0][2] + c[1][2] + c[2][2];
// 🔴 for fragment at channel i, set result
fragment[i] = r;
}
return fragment;
}
void mainImage( out vec4 fragColor, in vec2 fragCoord )
{
// 📈 Normalized pixel coordinates (from 0 to 1)
vec2 uv = fragCoord/iResolution.xy;
// 🕸️ Convolve kernel with texture
vec3 col = convolution(gaussian_blur, iChannel0, uv);
// ✍️ Output to screen
fragColor = vec4(col, 1.0);
}
# 🍱🚿 TensorFlow
import tensorflow as tf
from tensorflow.python.keras import layers
layers.LeakyReLU(filters, alpha=.3, ...)// ❌🧠 DirectML
struct DML_ACTIVATION_LEAKY_RELU_OPERATOR_DESC
{
const DML_TENSOR_DESC* InputTensor;
const DML_TENSOR_DESC* OutputTensor;
FLOAT Alpha;
};A Leaky Rectifier Linear Unit is an activation function that reduces the magnitude of its input below a threshold value, but otherwise returns a scaled value if it's above the threshold.
This is analogous to:
float x = x < threshold ? x * alpha : x;We use this to help the neural network find what features it needs to filter, that being areas of high variance in monte carlo images. ReLUs are very good at finding features in stochastic data, but can result in dead neurons in your network, that being, neurons that do not activate.
# 🍱🚿 TensorFlow
import tensorflow as tf
from tensorflow.python.keras import layers
layers.MaxPooling2D(pool_size=(2, 2), ...)// ❌🧠 DirectML
struct DML_MAX_POOLING_OPERATOR_DESC
{
const DML_TENSOR_DESC* InputTensor;
const DML_TENSOR_DESC* OutputTensor;
UINT DimensionCount;
_In_reads_(DimensionCount) const UINT* Strides;
_In_reads_(DimensionCount) const UINT* WindowSize;
_In_reads_(DimensionCount) const UINT* StartPadding;
_In_reads_(DimensionCount) const UINT* EndPadding;
};
The size of the pooling operation or filter is smaller than the size of the feature map; specifically, it is almost always 2×2 pixels applied with a stride of 2 pixels.
A Max Pooling is a tensor operation that checks neighboring features and outputs the maximum of that area.
This is analogous to downsampling in graphics, where neighboring pixels are often averaged then written to an output smaller than the original as in super-sampled anti-aliasing.
The pattern of convolutions followed by a nonlinear function, followed by pooling is quite common to neural networks. [Brownlee 2019]
 Ground Truth |  1spp |
|---|---|
 Direct Normal |  Indirect Normal |
 Indirect Albedo |
It's worth noting what data the autoencoder should expect and in what format so it's better tuned for denoising.
Should the autoencoder be able to denoise direct ray data (so direct normals), or direct and indirect (such as reflected normals, reflected albedo, etc.). Intel Open Image Denoise opts for training for both cases, but prefers indirect data as it results in more information for the network to make better decisions with. The decision of which to use can also be ambiguous, as the visibility of indirect rays can depend on the incidence of the first bounce along with their BRDF behavior (imagine capturing normals under a 🚗 car windshield for instance.)
How should you make sure your denoised output is temporally coherent? It's possible to feed motion vectors as an additional feature buffer to help with this.
Is Monte-Carlo biased noise (sobol noise, importance sampling) a factor when training? What about renderers? (Blender Cycles vs. Mitsuba vs. PBRT, etc.)
Should all inputs be anti-aliased or not? Most commercial autoencoders suggest all inputs be anti-aliased, and it is easy to anti-alias ray traced images anyways though offseting samples so this is understandable.
When measuring error, should the variance of the output vs ground truth be used as a metric as error loss, or should the color cosine difference of the output and ground truth be used? It's possible to compare both and determine the best choice by visually comparing models. Xu et al recommends using Wasserstein distance, a probability based distance common to recent GANs.
//https://gist.github.com/mjdietzx/a8121604385ce6da251d20d018f9a6d6
import tensorflow as tf
def em_loss(y_coefficients, y_pred):
return tf.reduce_mean(tf.multiply(y_coefficients, y_pred))Autoencoders have had great success in denoising images, filling in missing details, and in general getting an image closer to its ground truth.
Here's a few additional resources when developing your own machine learning denoisers:
In the University of Pennsylvania's CIS565 course, Dewang Sultania & Vaibhav Arcot implemented a Pytorch version of the paper Interactive Reconstruction of Monte Carlo Image Sequences using a Recurrent Denoising Autoencoder.
Riko Ophorst (@sanctumed) made a project that
uses DXR and NVIDIA's official autoencoder to denoise ray traced images in real
time.
Rajesh Sharma (@xarmalarma) of Disney Research hosted a Siggraph 2020 Course on Machine Learning for Computer Graphics.
Xiaoxu Meng released the source
of their Neural Bilateral Grid paper here.
Ingo Wald (@IngoWald) released a Optix course
providing examples for using the Optix denoiser here.
Martin-Karl Lefrançois (@doragonhanta) and
Christoph Kubisch (@pixeljetstream) of NVIDIA
released VKDenoise, an example showcasing
how to denoise a Vulkan scene with the Optix Denoiser.
NVIDIA has released an SDK to itegrate Deep Learning Super Sampling here. AMD's Fidelity FX Super resolution is also available for developers to integrate into their applications.
DirectML has released a few resources for using it in combination with TensorFlow and the ONNX Runtime.
| [Gallinari et al. 1987] |
| [G. Mixon et al. 2018] |
| [Khademi Kalantari et al. 2013] Removing the Noise in Monte Carlo Rendering with General Image Denoising Algorithms Eurographics 2013 research.nvidia.com |
| [Bako et al. 2017] Kernel-predicting convolutional networks for denoising Monte Carlo renderings ACM Transactions on Graphics 2017 cvc.ucsb.edu |
| [Vogels et al. 2018] Denoising with Kernel Prediction and Asymmetric Loss Functions ACM, ACM Transactions on Graphics 2018 doi.acm.org |
| [Laine et al. 2019] |
| [Khademi Kalantari et al. 2015] A Machine Learning Approach for Filtering Monte Carlo Noise ACM Transactions on Graphics (TOG) 2015 cvc.ucsb.edu |
| [R. Alla Chaitanya et al. 2017] Interactive Reconstruction of Monte Carlo Image Sequences Using a Recurrent Denoising Autoencoder Chakravarty R. Alla Chaitanya, Anton S. Kaplanyan,  Christoph Schied (@c_schied), Marco Salvi (@marcosalvi), Aaron Lefohn, Derek Nowrouzezahrai and Timo AilaACM, ACM Transactions on Graphics 2017 doi.acm.org |
| [Hasselgren et al. 2020] |
| [Muller et al. 2021] Real-time neural radiance caching for path tracing Association for Computing Machinery (ACM), ACM Transactions on Graphics 2021 dx.doi.org |
| [Mildenhall et al. 2017] |
| [Lehtinen et al. 2018] |
| [Dong et al. 2015] |
| [Ledig et al. 2016] Photo-Realistic Single Image Super-Resolution Using a Generative Adversarial Network CoRR 2016 arxiv.org |
| [Wang et al. 2018] |
| [Xiao et al. 2020] |
| [Schied et al. 2019] Real-Time Path Tracing and Denoising in 'Quake 2' Christoph Schied (@c_schied) and Alexey PanteleevGDC 2019 gdcvault.com youtu.be |
| [Meng et al. 2020] |
| [Goodfellow et al. 2016] |
| [Li 2020] |
| [Xu et al. 2019] Adversarial Monte Carlo Denoising with Conditioned Auxiliary Feature Modulation ACM, ACM Transactions on Graphics 2019 adversarial.mcdenoising.org |
| [Brownlee 2019] A Gentle Introduction to Pooling Layers for Convolutional Neural Networks 2019 machinelearningmastery.com |