Starless is a CPU black hole raytracer in numpy suitable for both informative diagrams and decent wallpaper material.

It is still in development and will be presented on my site along with my real time black hole visualizer. Starless, instead, is not real time but offers greater functionality, performing actual raytracing (while the mentioned WebGL applet uses a precomputed lookup texture for deflection).

• Full geodesic raytracing in Schwarzschild geometry
• Predicts distortion of arbitrary objects defined implicitly
• Alpha-blended accretion disk
• Sky distortion
• Dust
• Bloom postprocessing
• Completely parallel - renders chunks of the image using numpy arrays arithmetic
• Easy debugging by saving masks and intermediate results as images

python, numpy, scipy

a default adisk.jpg is now included.

My sky texture is too heavy to put on Github. To use it in the program, you need to place a bgedit.png texture in the textures/ folder for the sky (a skysphere, x = lon, y = lat) and optionally replace adisk.jpg for the accretion disk (x = angle, y = radius).

Example adisk.jpg

Example bgedit.png

Texture sizes are irrelevant (these texture will never end up in the GPU). You can use your own texture names or formats by editing the source.

Texture files are not needed if rendering with modes that don't use them.

Clone with git:

$ git clone https://github.com/rantonels/starless.git

or download this zip file and extract.

Write a .scene file. Examine the provided .scenes and model your file on them.

Some options will lie under the 'lofi' and 'hifi' sections. This are respectively the rendering specs for tests and the final image.

To run a test:

$ python tracer.py -d yourscene.scene

and wait. The rendered image will be in the tests folder under the name out.png, along with some other useful images.

To run the full render, just omit the -d option. The results will still be saved in tests.

tracer.py accepts the following command line options:

-d: run test (render with [lofi] settings)

--no-display: do not open matplotlib preview window. This is a huge improvement in speed for large images.

--no-shuffle: do not shuffle pixels before chunking. This, in practice, means that instead of being rendered as a gradually densening cloud, the image is raytraced progressively starting from the top. The end result is identical, but with shuffling the preview window might give an idea of the render sooner. However, disabling shuffling provides a nice speed improvement for larger images, because:

1. copying rendered data to the large final image buffer is slightly faster if it's contiguous
2. per chunk, the raytracer performs full calculations relative to an object (disc, horizon, etc) if and only if at least one ray of it its the object. So, if chunks are actually contiguous, there is a certain probability that some of them will never hit certain objects and many computations will be skipped. Shuffled chunks almost surely hit every object in the scene.

The (single) scene filename can be placed anywhere, and is recognized as such if it doesn't start with the - character. If omitted, scenes/default.scene is rendered.

Please refer to the scenes/default.scene file for a commented overview of all options.

Many options have default values and can be omitted, but I make no guarantees on the existence and values of such defaults. To be sure, include all options from scenes/default.scene.

Some general rules:

• Precision and render time are obviously massively affected by the Iterations and Stepsize options. As a rule of thumb, in [lofi] fix Stepsize to a large value, such as 0.08, then render multiple times with -d progressively decreasing Iterations. You need to find the smallest value of Iterations for which no significant parts of the image are clipped. When you've fixed the Iterations number, decrease the step size to a desirable value (0.02 is enough for most purposes) and simultaneously increase Iterations by the same factor (e.g.: halve step size, double iterations) so as to keep clipping distance approximately fixed. When satisfied, copy these values to the [hifi] section. This procedure minimizes render time.