The following API documentation of Intel Open Image Denoise can also be found as a pdf document.

Intel Open Image Denoise API

Intel Open Image Denoise provides a C99 API (also compatible with C++) and a C++11 wrapper API as well. For simplicity, this document mostly refers to the C99 version of the API.

The API is designed in an object-oriented manner, e.g. it contains device objects (OIDNDevice type), buffer objects (OIDNBuffer type), and filter objects (OIDNFilter type). All objects are reference-counted, and handles can be released by calling the appropriate release function (e.g. oidnReleaseDevice) or retained by incrementing the reference count (e.g. oidnRetainDevice).

An important aspect of objects is that setting their parameters do not have an immediate effect (with a few exceptions). Instead, objects with updated parameters are in an unusable state until the parameters get explicitly committed to a given object. The commit semantic allows for batching up multiple small changes, and specifies exactly when changes to objects will occur.

All API calls are thread-safe, but operations that use the same device will be serialized, so the amount of API calls from different threads should be minimized.

To have a quick overview of the C99 and C++11 APIs, see the following simple example code snippets.

C99 API Example

#include <OpenImageDenoise/oidn.h>
// Create an Intel Open Image Denoise device
OIDNDevice device = oidnNewDevice(OIDN_DEVICE_TYPE_DEFAULT);

// Create a denoising filter
OIDNFilter filter = oidnNewFilter(device, "RT"); // generic ray tracing filter
oidnSetSharedFilterImage(filter, "color",  colorPtr, 
                         OIDN_FORMAT_FLOAT3, width, height, 0, 0, 0);
oidnSetSharedFilterImage(filter, "albedo", albedoPtr,
                         OIDN_FORMAT_FLOAT3, width, height, 0, 0, 0); // optional
oidnSetSharedFilterImage(filter, "normal", normalPtr,
                         OIDN_FORMAT_FLOAT3, width, height, 0, 0, 0); // optional
oidnSetSharedFilterImage(filter, "output", outputPtr,
                         OIDN_FORMAT_FLOAT3, width, height, 0, 0, 0);
oidnSetFilter1b(filter, "hdr", true); // image is HDR

// Filter the image

// Check for errors
const char* errorMessage;
if (oidnGetDeviceError(device, &errorMessage) != OIDN_ERROR_NONE)
  printf("Error: %s\n", errorMessage);

// Cleanup

C++11 API Example

#include <OpenImageDenoise/oidn.hpp>
// Create an Intel Open Image Denoise device
oidn::DeviceRef device = oidn::newDevice();

// Create a denoising filter
oidn::FilterRef filter = device.newFilter("RT"); // generic ray tracing filter
filter.setImage("color",  colorPtr,  oidn::Format::Float3, width, height);
filter.setImage("albedo", albedoPtr, oidn::Format::Float3, width, height); // optional
filter.setImage("normal", normalPtr, oidn::Format::Float3, width, height); // optional
filter.setImage("output", outputPtr, oidn::Format::Float3, width, height);
filter.set("hdr", true); // image is HDR

// Filter the image

// Check for errors
const char* errorMessage;
if (device.getError(errorMessage) != oidn::Error::None)
  std::cout << "Error: " << errorMessage << std::endl;


Intel Open Image Denoise supports a device concept, which allows different components of the application to use the Open Image Denoise API without interfering with each other. An application first needs to create a device with

OIDNDevice oidnNewDevice(OIDNDeviceType type);

where the type enumeration maps to a specific device implementation, which can be one of the following:

Supported device types, i.e., valid constants of type OIDNDeviceType.
Name Description
OIDN_DEVICE_TYPE_DEFAULT select the approximately fastest device
OIDN_DEVICE_TYPE_CPU CPU device (requires SSE4.1 support)

Once a device is created, you can call

void oidnSetDevice1b(OIDNDevice device, const char* name, bool value);
void oidnSetDevice1i(OIDNDevice device, const char* name, int  value);
bool oidnGetDevice1b(OIDNDevice device, const char* name);
int  oidnGetDevice1i(OIDNDevice device, const char* name);

to set and get parameter values on the device. Note that some parameters are constants, thus trying to set them is an error. See the tables below for the parameters supported by devices.

Parameters supported by all devices.
Type Name Default Description
const int version combined version number (major.minor.patch) with two decimal digits per component
const int versionMajor major version number
const int versionMinor minor version number
const int versionPatch patch version number
int verbose 0 verbosity level of the console output between 0–4; when set to 0, no output is printed, when set to a higher level more output is printed
Additional parameters supported only by CPU devices.
Type Name Default Description
int numThreads 0 maximum number of threads which the library should use; 0 will set it automatically to get the best performance
bool setAffinity true bind software threads to hardware threads if set to true (improves performance); false disables binding

Note that the CPU device heavily relies on setting the thread affinities to achieve optimal performance, so it is highly recommended to leave this option enabled. However, this may interfere with the application if that also sets the thread affinities, potentially causing performance degradation. In such cases, the recommended solution is to either disable setting the affinities in the application or in Intel Open Image Denoise, or to always set/reset the affinities before/after each parallel region in the application (e.g., if using TBB, with tbb::task_arena and tbb::task_scheduler_observer).

Once parameters are set on the created device, the device must be committed with

void oidnCommitDevice(OIDNDevice device);

This device can then be used to construct further objects, such as buffers and filters. Note that a device can be committed only once during its lifetime. Before the application exits, it should release all devices by invoking

void oidnReleaseDevice(OIDNDevice device);

Note that Intel Open Image Denoise uses reference counting for all object types, so this function decreases the reference count of the device, and if the count reaches 0 the device will automatically get deleted. It is also possible to increase the reference count by calling

void oidnRetainDevice(OIDNDevice device);

An application typically creates only a single device. If required differently, it should only use a small number of devices at any given time.

Error Handling

Each user thread has its own error code per device. If an error occurs when calling an API function, this error code is set to the occurred error if it stores no previous error. The currently stored error can be queried by the application via

OIDNError oidnGetDeviceError(OIDNDevice device, const char** outMessage);

where outMessage can be a pointer to a C string which will be set to a more descriptive error message, or it can be NULL. This function also clears the error code, which assures that the returned error code is always the first error occurred since the last invocation of oidnGetDeviceError on the current thread. Note that the optionally returned error message string is valid only until the next invocation of the function.

Alternatively, the application can also register a callback function of type

typedef void (*OIDNErrorFunction)(void* userPtr, OIDNError code, const char* message);


void oidnSetDeviceErrorFunction(OIDNDevice device, OIDNErrorFunction func, void* userPtr);

to get notified when errors occur. Only a single callback function can be registered per device, and further invocations overwrite the previously set callback function, which do not require also calling the oidnCommitDevice function. Passing NULL as function pointer disables the registered callback function. When the registered callback function is invoked, it gets passed the user-defined payload (userPtr argument as specified at registration time), the error code (code argument) of the occurred error, as well as a string (message argument) that further describes the error. The error code is always set even if an error callback function is registered. It is recommended to always set a error callback function, to detect all errors.

When the device construction fails, oidnNewDevice returns NULL as device. To detect the error code of a such failed device construction, pass NULL as device to the oidnGetDeviceError function. For all other invocations of oidnGetDeviceError, a proper device handle must be specified.

The following errors are currently used by Intel Open Image Denoise:

Possible error codes, i.e., valid constants of type OIDNError.
Name Description
OIDN_ERROR_NONE no error occurred
OIDN_ERROR_UNKNOWN an unknown error occurred
OIDN_ERROR_INVALID_ARGUMENT an invalid argument was specified
OIDN_ERROR_INVALID_OPERATION the operation is not allowed
OIDN_ERROR_OUT_OF_MEMORY not enough memory to execute the operation
OIDN_ERROR_UNSUPPORTED_HARDWARE the hardware (e.g., CPU) is not supported
OIDN_ERROR_CANCELLED the operation was cancelled by the user


Large data like images can be passed to Intel Open Image Denoise either via pointers to memory allocated and managed by the user (this is the recommended, often easier and more efficient approach, if supported by the device) or by creating buffer objects (supported by all devices). To create a new data buffer with memory allocated and owned by the device, holding byteSize number of bytes, use

OIDNBuffer oidnNewBuffer(OIDNDevice device, size_t byteSize);

The created buffer is bound to the specified device (device argument). The specified number of bytes are allocated at buffer construction time and deallocated when the buffer is destroyed.

It is also possible to create a “shared” data buffer with memory allocated and managed by the user with

OIDNBuffer oidnNewSharedBuffer(OIDNDevice device, void* ptr, size_t byteSize);

where ptr points to the user-managed memory and byteSize is its size in bytes. At buffer construction time no buffer data is allocated, but the buffer data provided by the user is used. The buffer data must remain valid for as long as the buffer may be used, and the user is responsible to free the buffer data when no longer required.

Similar to device objects, buffer objects are also reference-counted and can be retained and released by calling the following functions:

void oidnRetainBuffer(OIDNBuffer buffer);
void oidnReleaseBuffer(OIDNBuffer buffer);

Accessing the data stored in a buffer object is possible by mapping it into the address space of the application using

void* oidnMapBuffer(OIDNBuffer buffer, OIDNAccess access, size_t byteOffset, size_t byteSize)

where access is the desired access mode of the mapped memory, byteOffset is the offset to the beginning of the mapped memory region in bytes, and byteSize is the number of bytes to map. The function returns a pointer to the mapped buffer data. If the specified byteSize is 0, the maximum available amount of memory will be mapped. The access argument must be one of the access modes in the following table:

Access modes for memory regions mapped with oidnMapBuffer, i.e., valid constants of type OIDNAccess.
Name Description
OIDN_ACCESS_READ read-only access
OIDN_ACCESS_WRITE write-only access
OIDN_ACCESS_READ_WRITE read and write access
OIDN_ACCESS_WRITE_DISCARD write-only access but the previous contents will be discarded

After accessing the mapped data in the buffer, the memory region must be unmapped with

void oidnUnmapBuffer(OIDNBuffer buffer, void* mappedPtr);

where mappedPtr must be a pointer returned by a call to oidnMapBuffer for the specified buffer. Any change to the mapped data is guaranteed to take effect only after unmapping the memory region.

Data Format

Buffers store opaque data and thus have no information about the type and format of the data. Other objects, e.g. filters, typically require specifying the format of the data stored in buffers or shared via pointers. This can be done using the OIDNFormat enumeration type:

Supported data formats, i.e., valid constants of type OIDNFormat.
Name Description
OIDN_FORMAT_UNDEFINED undefined format
OIDN_FORMAT_FLOAT 32-bit single-precision floating point scalar
OIDN_FORMAT_FLOAT[234] … and [234]-element vector


Filters are the main objects in Intel Open Image Denoise that are responsible for the actual denoising. The library ships with a collection of filters which are optimized for different types of images and use cases. To create a filter object, call

OIDNFilter oidnNewFilter(OIDNDevice device, const char* type);

where type is the name of the filter type to create. The supported filter types are documented later in this section. Once created, filter objects can be retained and released with

void oidnRetainFilter(OIDNFilter filter);
void oidnReleaseFilter(OIDNFilter filter);

After creating a filter, it needs to be set up by specifying the input and output images, and potentially setting other parameter values as well.

To bind images to the filter, you can use one of the following functions:

void oidnSetFilterImage(OIDNFilter filter, const char* name,
                        OIDNBuffer buffer, OIDNFormat format,
                        size_t width, size_t height,
                        size_t byteOffset,
                        size_t bytePixelStride, size_t byteRowStride);

void oidnSetSharedFilterImage(OIDNFilter filter, const char* name,
                              void* ptr, OIDNFormat format,
                              size_t width, size_t height,
                              size_t byteOffset,
                              size_t bytePixelStride, size_t byteRowStride);

It is possible to specify either a data buffer object (buffer argument) with the oidnSetFilterImage function, or directly a pointer to shared user-managed data (ptr argument) with the oidnSetSharedFilterImage function.

In both cases, you must also specify the name of the image parameter to set (name argument, e.g. "color", "output"), the pixel format (format argument), the width and height of the image in number of pixels (width and height arguments), the starting offset of the image data (byteOffset argument), the pixel stride (bytePixelStride argument) and the row stride (byteRowStride argument), in number of bytes. Note that the row stride must be an integer multiple of the pixel stride.

If the pixels and/or rows are stored contiguously (tightly packed without any gaps), you can set bytePixelStride and/or byteRowStride to 0 to let the library compute the actual strides automatically, as a convenience.

Some special data used by filters are opaque/untyped (e.g. trained model weights blobs), which can be specified with the oidnSetSharedFilterData function:

void oidnSetSharedFilterData(OIDNFilter filter, const char* name,
                             void* ptr, size_t byteSize);

Filters may have parameters other than buffers as well, which you can set and get using the following functions:

void  oidnSetFilter1b(OIDNFilter filter, const char* name, bool  value);
void  oidnSetFilter1i(OIDNFilter filter, const char* name, int   value);
void  oidnSetFilter1f(OIDNFilter filter, const char* name, float value);
bool  oidnGetFilter1b(OIDNFilter filter, const char* name);
int   oidnGetFilter1i(OIDNFilter filter, const char* name);
float oidnGetFilter1f(OIDNFilter filter, const char* name);

Filters support a progress monitor callback mechanism that can be used to report progress of filter operations and to cancel them as well. Calling oidnSetFilterProgressMonitorFunction registers a progress monitor callback function (func argument) with payload (userPtr argument) for the specified filter (filter argument):

typedef bool (*OIDNProgressMonitorFunction)(void* userPtr, double n);

void oidnSetFilterProgressMonitorFunction(OIDNFilter filter,
                                          OIDNProgressMonitorFunction func,
                                          void* userPtr);

Only a single callback function can be registered per filter, and further invocations overwrite the previously set callback function. Passing NULL as function pointer disables the registered callback function. Once registered, Intel Open Image Denoise will invoke the callback function multiple times during filter operations, by passing the payload as set at registration time (userPtr argument), and a double in the range [0, 1] which estimates the progress of the operation (n argument). When returning true from the callback function, Intel Open Image Denoise will continue the filter operation normally. When returning false, the library will cancel the filter operation with the OIDN_ERROR_CANCELLED error code.

After setting all necessary parameters for the filter, the changes must be commmitted by calling

void oidnCommitFilter(OIDNFilter filter);

The parameters can be updated after committing the filter, but it must be re-committed for the changes to take effect.

Finally, an image can be filtered by executing the filter with

void oidnExecuteFilter(OIDNFilter filter);

which will read the input image data from the specified buffers and produce the denoised output image.

In the following we describe the different filters that are currently implemented in Intel Open Image Denoise.


The RT (ray tracing) filter is a generic ray tracing denoising filter which is suitable for denoising images rendered with Monte Carlo ray tracing methods like unidirectional and bidirectional path tracing. It supports depth of field and motion blur as well, but it is not temporally stable. The filter is based on a convolutional neural network (CNN), and it aims to provide a good balance between denoising performance and quality. The filter comes with a set of pre-trained CNN models that work well with a wide range of ray tracing based renderers and noise levels.

It accepts either a low dynamic range (LDR) or high dynamic range (HDR) color image as input. Optionally, it also accepts auxiliary feature images, e.g. albedo and normal, which improve the denoising quality, preserving more details in the image.

The RT filter has certain limitations regarding the supported input images. Most notably, it cannot denoise images that were not rendered with ray tracing. Another important limitation is related to anti-aliasing filters. Most renderers use a high-quality pixel reconstruction filter instead of a trivial box filter to minimize aliasing artifacts (e.g. Gaussian, Blackman-Harris). The RT filter does support such pixel filters but only if implemented with importance sampling. Weighted pixel sampling (sometimes called splatting) introduces correlation between neighboring pixels, which causes the denoising to fail (the noise will not be filtered), thus it is not supported.

The filter can be created by passing "RT" to the oidnNewFilter function as the filter type. The filter supports the parameters listed in the table below. All specified images must have the same dimensions. The output image can be one of the input images (i.e. in-place denoising is supported).

Parameters supported by the RT filter.
Type Format Name Default Description
Image float3 color input color image (LDR values in [0, 1] or HDR values in [0, +∞), 3 channels)
Image float3 albedo input feature image containing the albedo (values in [0, 1], 3 channels) of the first hit per pixel; optional
Image float3 normal input feature image containing the shading normal (world-space or view-space, arbitrary length, values in (−∞, +∞), 3 channels) of the first hit per pixel; optional, requires setting the albedo image too
Image float3 output output color image (3 channels); it can be one of the input images
Data weights trained model weights blob; optional
bool hdr false whether the color is HDR
float hdrScale NaN HDR color values are interpreted such that, multiplied by this scale, a value of 1 corresponds to a luminance level of 100 cd/m² (this affects the quality of the output but the output color values will not be scaled); if set to NaN, the scale is computed automatically (default)
bool srgb false whether the color is encoded with the sRGB (or 2.2 gamma) curve (LDR only) or is linear; the output will be encoded with the same curve
int maxMemoryMB 6000 approximate maximum amount of scratch memory to use in megabytes (actual memory usage may be higher); limiting memory usage may cause slower denoising due to internally splitting the image into overlapping tiles, but cannot cause the denoising to fail
const int alignment when manually denoising the image in tiles, the tile size and offsets should be multiples of this amount of pixels to avoid artifacts; note that manual tiled denoising of HDR images is supported only when hdrScale is set by the user
const int overlap when manually denoising the image in tiles, the tiles should overlap by this amount of pixels
Example noisy color image rendered using unidirectional path tracing (64 spp). Scene by Evermotion.
Example output image denoised using color and auxiliary feature images (albedo and normal).

Using auxiliary feature images like albedo and normal helps preserving fine details and textures in the image thus can significantly improve denoising quality. These images should typically contain feature values for the first hit (i.e. the surface which is directly visible) per pixel. This works well for most surfaces but does not provide any benefits for reflections and objects visible through transparent surfaces (compared to just using the color as input). However, in certain cases this issue can be fixed by storing feature values for a subsequent hit (i.e. the reflection and/or refraction) instead of the first hit. For example, it usually works well to follow perfect specular (delta) paths and store features for the first diffuse or glossy surface hit instead (e.g. for perfect specular dielectrics and mirrors). This can greatly improve the quality of reflections and transmission. We will describe this approach in more detail in the following subsections.

The auxiliary feature images should be as noise-free as possible. It is not a strict requirement but too much noise in the feature images may cause residual noise in the output. Also, all feature images should use the same pixel reconstruction filter as the color image. Using a properly anti-aliased color image but aliased albedo or normal images will likely introduce artifacts around edges.


The albedo image is the feature image that usually provides the biggest quality improvement. It should contain the approximate color of the surfaces independent of illumination and viewing angle.

For simple matte surfaces this means using the diffuse color/texture as the albedo. For other, more complex surfaces it is not always obvious what is the best way to compute the albedo, but the denoising filter is flexibile to a certain extent and works well with differently computed albedos. Thus it is not necessary to compute the strict, exact albedo values but must be always between 0 and 1.

Example albedo image obtained using the first hit. Note that the albedos of all transparent surfaces are 1.
Example albedo image obtained using the first diffuse or glossy (non-delta) hit. Note that the albedos of perfect specular (delta) transparent surfaces are computed as the Fresnel blend of the reflected and transmitted albedos.

For metallic surfaces the albedo should be either the reflectivity at normal incidence (e.g. from the artist friendly metallic Fresnel model) or the average reflectivity; or if these are constant (not textured) or unknown, the albedo can be simply 1 as well.

The albedo for dielectric surfaces (e.g. glass) should be either 1 or, if the surface is perfect specular (i.e. has a delta BSDF), the Fresnel blend of the reflected and transmitted albedos (as previously discussed). The latter usually works better but only if it does not introduce too much additional noise due to random sampling. Thus we recommend to split the path into a reflected and a transmitted path at the first hit, and perhaps fall back to an albedo of 1 for subsequent dielectric hits, to avoid noise. The reflected albedo in itself can be used for mirror-like surfaces as well.

The albedo for layered surfaces can be computed as the weighted sum of the albedos of the individual layers. Non-absorbing clear coat layers can be simply ignored (or the albedo of the perfect specular reflection can be used as well) but absorption should be taken into account.


The normal image should contain the shading normals of the surfaces either in world-space or view-space. It is recommended to include normal maps to preserve as much detail as possible.

Just like any other input image, the normal image should be anti-aliased (i.e. by accumulating the normalized normals per pixel). The final accumulated normals do not have to be normalized but must be in a range symmetric about 0 (i.e. normals mapped to [0, 1] are not acceptable and must be remapped to e.g. [−1, 1]).

Similar to the albedo, the normal can be stored for either the first or a subsequent hit (if the first hit has a perfect specular/delta BSDF).

Example normal image obtained using the first hit (the values are actually in [−1, 1] but were mapped to [0, 1] for illustration purposes).
Example normal image obtained using the first diffuse or glossy (non-delta) hit. Note that the normals of perfect specular (delta) transparent surfaces are computed as the Fresnel blend of the reflected and transmitted normals.


Instead of using the built-in trained models for filtering, it is also possible to specify user-trained models at runtime. This can be achieved by passing the model weights blob produced by the training tool.


The RTLightmap filter is a variant of the RT filter optimized for denoising HDR lightmaps. It does not support LDR images.

The filter can be created by passing "RTLightmap" to the oidnNewFilter function as the filter type. The filter supports the following parameters:

Parameters supported by the RTLightmap filter.
Type Format Name Default Description
Image float3 color input color image (HDR values in [0, +∞), 3 channels)
Image float3 output output color image (3 channels); it can be one of the input images
Data weights trained model weights blob; optional
float hdrScale NaN HDR color values are interpreted such that, multiplied by this scale, a value of 1 corresponds to a luminance level of 100 cd/m²; if set to NaN, the scale is computed automatically (default)
int maxMemoryMB 6000 approximate maximum amount of scratch memory to use in megabytes (actual memory usage may be higher)
const int alignment when manually denoising the image in tiles, the tile size and offsets should be multiples of this amount of pixels
const int overlap when manually denoising the image in tiles, the tiles should overlap by this amount of pixels



A minimal working example demonstrating how to use Intel Open Image Denoise can be found at apps/oidnDenoise/oidnDenoise.cpp, which uses the C++11 convenience wrappers of the C99 API.

This example is a simple command-line application that denoises the provided image, which can optionally have auxiliary feature images as well (e.g. albedo and normal). By default the images must be stored in the Portable FloatMap (PFM) format, and the color values must be encoded in little-endian format. To enable other image formats (e.g. OpenEXR, PNG) as well, the project has to be rebuilt with OpenImageIO support enabled.

Running oidnDenoise without any arguments will bring up a list of command line options.


The Intel Open Image Denoise source distribution includes a Python-based neural network training toolkit (located in the training directory), which can be used to train the denoising filter models with image datasets provided by the user. The toolkit consists of the following command-line scripts:


Before you can run the training toolkit you need the following prerequisites:


A dataset should consist of a collection of noisy and corresponding noise-free reference images. It is possible to have more than one noisy version of the same image in the dataset, e.g. rendered at different samples per pixel and/or using different seeds.

The training toolkit expects to have all datasets (e.g. training, validation) in the same parent directory (e.g. data). Each dataset is stored in its own subdirectory (e.g. train, valid), which can have an arbitrary name.

The images must be stored in OpenEXR format (.exr files), and the filenames must have a specific format but the files can be stored in an arbitrary directory structure inside the dataset directory. The only restriction is that all versions of an image (noisy images and the reference image) must be located in the same subdirectory. Each feature of an image (e.g. color, albedo) must be stored in a separate image file, i.e. multi-channel EXR image files are not supported. If you have multi-channel EXRs, you can split them into separate images per feature using the included tool.

An image filename must consist of a name, the number of samples per pixel or whether it is the reference (e.g. 0128spp, ref), the identifier (ID) of the feature (e.g. hdr, alb), and the file extension (.exr). The exact format as a regular expression is the following:


The number of samples per pixel should be padded with leading zeros to have a fixed number of digits. If the reference image is not explicitly named as such (i.e. has the number of samples instead), the image with the most samples per pixel will be considered the reference.

The following image features are supported:

Supported image features, their IDs, and their number of channels.
Feature ID Channels
color (HDR) hdr 3
color (LDR) ldr 3
albedo alb 3
normal nrm 3

The following directory tree demonstrates an example root dataset directory (data) containing one dataset (rt_train) with HDR color and albedo feature images:

`-- rt_train
    |-- scene1
    |   |-- view1_0001.alb.exr
    |   |-- view1_0001.hdr.exr
    |   |-- view1_0004.alb.exr
    |   |-- view1_0004.hdr.exr
    |   |-- view1_8192.alb.exr
    |   |-- view1_8192.hdr.exr
    |   |-- view2_0001.alb.exr
    |   |-- view2_0001.hdr.exr
    |   |-- view2_8192.alb.exr
    |   `-- view2_8192.hdr.exr
    |-- scene2_000008spp.alb.exr
    |-- scene2_000008spp.hdr.exr
    |-- scene2_000064spp.alb.exr
    |-- scene2_000064spp.hdr.exr
    |-- scene2_reference.alb.exr
    `-- scene2_reference.hdr.exr

Preprocessing (

Training and validation datasets can be used only after preprocessing them using the script. This will convert the specified training (-t or --train_data option) and validation datasets (-v or --valid_data option) located in the root dataset directory (-D or --data_dir option) to a format that can be loaded more efficiently during training. All preprocessed datasets will be stored in a root preprocessed dataset directory (-P or --preproc_dir option).

The preprocessing script requires the set of image features to include in the preprocessed dataset as command-line arguments. Only these specified features will be available for training. Preprocessing also depends on the filter that will be trained (e.g. determines which HDR/LDR transfer function has to be used), which should be also specified (-f or --filter option). The alternative is to manually specify the transfer function (-x or --transfer option) and other filter-specific parameters, which could be useful for training custom filters.

For example, to preprocess the training and validation datasets (rt_train and rt_valid) with HDR color, albedo, and normal image features, for training the RT filter, the following command can be used:

./ hdr alb nrm --filter RT --train_data rt_train --valid_data rt_valid

For more details about using the preprocessing script, including other options, please have a look at the help message:

./ -h

Training (

After preprocessing the datasets, it is possible to start training a model using the script. Similar to the preprocessing script, the input features must be specified (could be a subset of the preprocessed features), and the dataset names, directory paths, and the filter can be also passed.

The tool will produce a training result, the name of which can be either specified (-r or --result option) or automatically generated (by default). Each result is stored in its own subdirectory, and these are located in a common parent directory (-R or --results_dir option). If a training result already exists, the tool will resume training that result from the latest checkpoint, or from an earlier checkpoint at the specified epoch (-c or --checkpoint option).

The default training hyperparameters should work reasonably well in general, but some adjustments might be necessary for certain datasets to attain optimal performance, most importantly: the number of epochs (-e or --epochs option), the mini-batch size (--bs or --batch_size option), and the learning rate. The training tool uses a cyclical learning rate (CLR) with the triangular2 scaling policy and an optional linear ramp-down at the end. The learning rate schedule can be configured by setting the base learning rate (--lr or --learning_rate option), the maximum learning rate (--max_lr or --max_learning_rate option), and the total cycle size in number of epochs (--lr_cycle_epochs option). If there is an incomplete cycle at the end, the learning rate will be linearly ramped down to almost zero.

Example usage:

./ hdr alb --filter RT --train_data rt_train --valid_data rt_valid --result rt_hdr_alb

For finding the optimal learning rate range we recommend using the included script, which trains one epoch using an increasing learning rate and logs the resulting losses in a comma-separated values (CSV) file. Plotting the loss curve can show when the model starts to learn (the base learning rate) and when it starts to diverge (the maximum learning rate).

The model is evaluated with the validation dataset at regular intervals (--valid_epochs option), and checkpoints are also regularly created (--save_epochs option) to save training progress. Also, some statistics are logged (e.g. training and validation losses, learning rate) at a specified frequency (--log_steps option), which can be later visualized with TensorBoard by running the script, e.g.:

./ --result rt_hdr_alb

Inference (

A training result can be tested by performing inference on an image dataset (-i or --input_data option) using the script. The dataset does not have to be preprocessed. In addition to the result to use, it is possible to specify which checkpoint to load as well. By default the latest checkpoint is loaded.

The tool saves the output images in a separate directory (-O or --output_dir) in the requested formats (-F or --format option). It also evaluates a set of image quality metrics (-M or --metric option), e.g. PSNR, SSIM, for images that have reference images available. All metrics are computed in tonemapped non-linear sRGB space. Thus, HDR images are first tonemapped (with Naughty Dog’s Filmic Tonemapper from John Hable’s Uncharted 2: HDR Lighting presentation) and converted to sRGB before evaluating the metrics.

Example usage:

./ --result rt_hdr_alb --input_data rt_test --format exr png --metric ssim

Exporting Results (

The training result produced by the script cannot be immediately used by the main library. It has to be first exported to the runtime model weights format, a Tensor Archive (TZA) file. Running the script for a training result (and optionally a checkpoint) will create a binary .tza file in the directory of the result, which can be either used at runtime through the API or it can be included in the library build by replacing one of the built-in weights files.

Example usage:

./ --result rt_hdr_alb

Image Conversion and Comparison

In addition to the already mentioned script, the toolkit contains a few other image utilities as well. converts a feature image to a different image format (and/or a different feature, e.g. HDR color to LDR), performing tonemapping and other transforms as well if needed. For HDR images the exposure can be adjusted by passing a linear exposure scale (-E or --exposure option). Example usage:

./ view1_0004.hdr.exr view1_0004.png --exposure 2.5

The script compares two feature images (preferably having the dataset filename format to correctly detect the feature) using the specified image quality metrics, similar to the tool. Example usage:

./ view1_0004.hdr.exr view1_8192.hdr.exr --exposure 2.5 --metric mse ssim