SeeSharp is a flexibility-first framework for rapid prototyping of rendering algorithms. Currently, it only works with triangle meshes and only handles surface interactions. The framework implements some basic integrators, namely unidirectional and bidirectional path tracing, photon mapping, and vertex connection and merging. These offer a range of virtual functions to easily and cleanly inject additional logic, like better importance sampling, path guiding, and so on.
Images being rendered can be streamed interactively to the tev image viewer, via sockets.
Additionally, the framework offers utility classes to render comparisons of different integrators and/or settings. That is, we use C# to implement the integrators and logic, and as the scripting language to run experiments.
An example of how to successfully use this framework to conduct experiments for a research paper can be found in the implementation of our EG 2021 paper on correlation-aware MIS.
We use TinyEmbree, a simple C# wrapper around Embree, for ray tracing, and SimpleImageIO, a C# wrapper around TinyEXR and stb_image. On 64 Bit Windows, Linux, or macOS, you should be able to use the pre-built binaries included in the nuget packages. These are linked by default, so you do not need to do anything and can skip ahead to the next section.
On other platforms you will need to build these from source. Instructions how to do that can be found in the respective README.md files. After building your platform specific version of these, the easiest way is to pack them as a local nuget package. For example:
mkdir ~/LocalNuget
dotnet nuget add source ~/LocalNuget
cd [TinyEmbree ROOT]
dotnet pack -c Release
cp TinyEmbree/bin/Release/TinyEmbree.[VERSION].nupkg ~/LocalNuget
An alternative is to replace the <PackageReference .../>
in SeeSharp.csproj
by a <ProjectReference .../>
to your local
TinyEmbree.csproj, which also allows you to more easily modify both.
Note: The .fbx loader is currently relying on Assimp.NET. On Linux (and maybe MacOS) you might see a "library not found" exception for
libassimp.so
. In that case, you need to manually installlibminizip
. (Linkinglibassimp.so
fails if that dependency is not found. It is not bundled in the NuGet package).
The unit tests can be run via a simple
dotnet test
The validation test, which ensure that all integrators agree on the same results for a number of trivial test scenes, can be run in Release mode via:
dotnet run --project SeeSharp.Validation -c Release
Note that the validation tests assume that they are run from within the root directory of this project, as they
rely on some scene files stored in the Data
directory.
SeeSharp comes with a very basic Blender add-on. The see_blender.zip
file is automatically built for each release. You can install it in Blender via Edit -> Preferences -> Add-ons -> Install...
. Check the docs for more details.
When building locally, you can run build_blender.bat
or build_blender.sh
to generate the Blender add-on.
The add-on currently handles two things: exporting a scene to our .json
format and rendering a preview when hitting F12 (final render only, viewport rendering not yet supported). These are very rudimentary and might not work for all possible Blender features and scene configurations.
Also, the add-on offers some basic GUI support to configure material parameters and add an HDR background.
Some simple Cycles materials can be converted to a SeeSharp material. Currently, there is a coarse mapping from:
- DiffuseBSDF: either with a constant color or a texture, roughness is ignored
- PrincipledBSDF: roughly mapped to our GenericMaterial, which currently ignores SSS, sheen, and clearcoat. Also, only BaseColor and roughness can be textured at the moment
- Emission: mapped to a diffuse black body emitter
- The viewport preview color and roughness, if all else fails
For most existing scenes, you will need to manually simplify the usually complex shader graphs to one of the above, else only the viewport settings will be used.
All materials in the entire scene can be converted via File -> Import -> Convert all materials to SeeSharp
.
Converting an individual material is also possible, via a button in that material's parameter settings.
SeeSharp is designed to be used as a library, to write rendering experiments with. To get started, you should first create a new console application that will contain you experiment set-up, as well as any additional algorithms or other changes you will introduce.
dotnet new console -o MyFirstExperiment
cd MyFirstExperiment
dotnet add package SeeSharp
Now, you can implement new or derived integrators and write your own experiment setup,
for instance by deriving a class from ExperimentFactory
.
Another option is to use .NET interactive in a Jupyter notebook. Or, you could write a C# script or F# script.
The following example of a .csx script conducts an experiment that compares a path tracer to the vertex connection and merging algorithm, at equal sample count:
#r "nuget: SeeSharp, 1.4.0"
using System.Collections.Generic;
using System.Diagnostics;
using SeeSharp.Experiments;
using SeeSharp.Image;
using SeeSharp.Integrators;
using SeeSharp.Integrators.Bidir;
// Configure an experiment that compares VCM and path tracing.
class PathVsVcm : Experiment {
public override List<Method> MakeMethods() => new() {
new("PathTracer", new PathTracer() { MaxDepth = 5, TotalSpp = 4 }),
new("Vcm", new VertexConnectionAndMerging() { MaxDepth = 5, NumIterations = 2 })
};
}
// Register the directory as a scene file provider.
// Asides from the geometry, it is also used as a reference image cache.
SceneRegistry.AddSource("Data/Scenes");
// Configure a benchmark to compare path tracing and VCM on the CornellBox
// at 512x512 resolution. Display images in tev during rendering (localhost, default port)
Benchmark benchmark = new(new PathVsVcm(), new() {
SceneRegistry.LoadScene("CornellBox", maxDepth: 5),
SceneRegistry.LoadScene("CornellBox", maxDepth: 2).WithName("CornellBoxDirectIllum")
}, "Results/PathVsVcm", 512, 512, FrameBuffer.Flags.SendToTev);
// Render the images
benchmark.Run(format: ".exr");
// Optional, but usually a good idea: assemble the rendering results in an overview
// figure using a Python script.
Process.Start("python", "./SeeSharp.Examples/MakeFigure.py Results/PathVsVcm PathTracer Vcm")
.WaitForExit();
// For our README file, we further convert the pdf to png with ImageMagick
Process.Start("magick", "-density 300 ./Results/PathVsVcm/Overview.pdf ExampleFigure.png")
.WaitForExit();
The first line automatically downloads and installs the SeeSharp package using nuget. Hence, you can simply run the experiment via:
dotnet script MyExperiment.csx
Alternatively, you could paste the same code in the cells of a .ipynb and run it with Jupyter. Then, you can even display the generated figure right underneath.
In this example, we automatically invoke Python, which is assumed to be in the PATH, to assemble the rendered images in a comparison figure. We use figuregen for that, so make sure to run:
python -m pip install figuregen
And here's the generated overview figure:
The project-file based version of this example, as well as the MakeFigure.py script, can be found in the SeeSharp.Examples directory.
The framework loosely follows the following coding conventions:
- Class names, method names, properties, and public fields are PascalCase
- Private fields, parameters, and local variables are camelCase
- Opening brackets are on the same line
- Lines should be less than 110 characters long, for better readability and comparisons
The .editorconfig and omnisharp.json files configure some of these conventions for Visual Studio and VS Code, respectively.
Example:
class SomeClass {
public int PublicField = 1;
public int PublicProperty => 13;
public void ComputeSomething(byte complexParameterName, int anotherVeryLongParameterName,
SomeClass mysteriousParameterThatIsNotNamedWell) {
if (complexParameterName == PublicProperty) {
int localName = 5;
implementationDetail = $"SeeSharp {localName}";
}
}
string implementationDetail;
}