Skip to content

templated flame fractal renderer written in c++ supporting output as image or the internal buffer

License

Notifications You must be signed in to change notification settings

tkoz0/flame-fractal-renderer

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

flame fractal renderer

Image 1 Image 2 Image 3

Sample images. From left to right:

  1. The Sierpinski Triangle, a well known fractal.
  2. A flame fractal I created for a project in a parallel computing class.
  3. Another flame fractal I created while testing color implementation.

what are flame fractals?

Flame fractals are fractals created by certain iterated function systems. They were created by Scott Draves in 1992. See his paper here: https://flam3.com/flame_draves.pdf. He also has a software called flam3 which is the main inspiration for writing this renderer. You can find it here: https://github.com/scottdraves/flam3.

Flame fractals are based on iterated function systems. They are typically considered in 2 dimensions for an image, but can be generalized to higher dimensions. In $n$ dimensions, we have a set of functions $f_i:R^n\to R^n$ with weights $w_i$ that sum to 1. The details in specifying these functions will come later. To render a flame fractal, start with a random point $x$ in the biunit square/cube/hypercube centered at the origin. Then to perform an iteration, randomly choose a function $f_i$ according to the weights and update $x$ to be $f_i(x)$. This will generate a sequence of points. For each point (except some at the beginning to allow the sequence to settle), assign it to an area/volume in the space $R^n$ divided into tiny squares/cubes/hypercubes.

In 2D, we have functions $f_i:R^2\to R^2$. We divide the 2D plane into small squares corresponding to pixels. For each pixel, we have a counter that gets incremented each time the generated sequence lands inside its area. Finally, to plot the result, the value for each pixel is scaled and then output as a grayscale image.

Flame fractals can be rendered with color. This is done by assigning colors to each function and blending them with exponential weighting so that the pixels are colored according to which functions are most associated with that part of the image. There is a lot of complexity in methods used to render images with good appearance.

The functions $f_i$ are defined by a pre-affine and post-affine transformation. Each of these is a linear transformation that can be defined as $Ax+b$ for some $n\times n$ matrix $A$ and $n$-vector $b$. Between these, we apply a sum of variation functions. Then to apply $f_i$ to a point $x$, we first apply the pre-affine transformation. Next, use this point to compute the variation function sum. Finally, apply the post-affine transformation to this sum.

why another renderer when others exist?

I initially created this because I wanted to create the probability distribution buffer and then render images based on that with various post-processing and filtering algorithms. Since the probability distribution buffer is the most computationally expensive part to generate, it makes sense to save its results and spend more effort optimizing its computation.

It expanded a bit when I decided to write a whole renderer that I would try to optimize decently, although that is difficult to do when trying to support arbitrary functions to be specified in JSON files and when most of the runtime depends on random access to a large area of memory. I ended up switching from C to C++ and making the whole thing templated to be able to support different compile time parameters while maintaining good performance.

Another reason was interest in exploring 3D flame fractals. Higher dimensions are supported, but no tools have been created for displaying the results in a meaningful way yet.

json format

Flames fractals for this software are specified in JSON. See the example flames for a good idea of how to write them. Comments are allowed in JSON files used for this program.

The JSON format documentation needs to be improved, especially to show details about all the variation functions.

building

Make sure you have the following Ubuntu packages installed (or equivalent on other distros)

  1. build-essential (for make and g++)
  2. wget
  3. libboost-program-options-dev (or libboost-all-dev)
  4. libpng-dev

As of the time of writing this, there is a known issue with Boost 1.74 (which comes from the Ubuntu repositories) where it will not compile with C++20 (which is required for this project). In order to resolve this, install the latest version of boost. The following commands are based on this link: https://www.baeldung.com/linux/boost-install-on-ubuntu

  1. mkdir -p boost/
  2. cd boost
  3. wget https://boostorg.jfrog.io/artifactory/main/release/1.84.0/source/boost_1_84_0.tar.bz2
  4. tar -xvf boost_1_84_0.tar.bz2
  5. cd boost_1_84_0/
  6. ./bootstrap.sh
  7. ./b2
  8. sudo ./b2 install --prefix=/usr/local

Then if everything is setup correctly, running make all should build 2 executables, described below.

For development, static linking should be avoided because Valgrind will show memory errors when glibc is statically linked. The solution I have for now is to install Boost 1.84 locally so it can be used with dynamic linking.

This was only tested on Ubuntu 20.04 and 22.04 since I only use Ubuntu for development. Help to expand this section to support other systems would be appreciated.

command line options

Currently there are 2 executables: ffr-buf.out and ffr-img.out. The first renders a buffer, which can have a number of dimensions specified. The second is a basic implementation that renders an image from 2D buffers only.

The options for ffr-buf.out are:

-h,--help help message

-f,--flame flame JSON file (required), or - for stdin

-i,--input specify input buffers (can be used to add more rendering progress to an existing buffer), used once per input file, - for stdin

-o,--output output file (required), or - for stdout

-s,--samples number of samples to render, default is 0 (no rendering)

-t,--threads threads for multithreading, default is the number of logical CPU cores

-b,--batch_size number of samples per thread work unit, if 0 or unspecified then a reasonable size is calculated, larger means less overhead

-z,--bad_values number of bad values allowed before render terminates, default is 16, bad values are indicative of degenerate flame parameters

The options for ffr-img.out are:

-h,--help help message

-f,--flame flame JSON file (required), or - for stdin

-i,--input input buffers, need at least 1, usue - for stdin

-o,--output output file, PNG images only for now

-y,--gamma gamma parameter for adjusting overall brightness

-m,--monochrome binary image (white = nonzero histogram)

-g,--grayscale grayscale image from log scaling the histogram

-c,--color color image using 3D color vectors for RGB only for now

-b,--bits bit depth per channel for gray/color (8 or 16, default 8)

example commands

Render a flame with 1000000 samples and save the buffer.

ffr-buf.out -f flame.json -o flame.buf -s 1000000

Add 1000000 rendering samples to an existing buffer.

ffr-buf.out -f flame.json -i flame1.buf -o flame2.buf -s 1000000

Render a grayscale image based on a buffer file using gamma=2.

ffr-img.out -f flame.json -i flame.buf -o flame.png -g -y 2

Perform a single threaded render and directly feed the buffer with stdout/stdin to render a color image with gamma=2.5.

ffr-buf.out -f flame.json -o - -s 1000000 -t 1 | ffr-img.out -f flame.json -i - -o flame.png -c -y 2.5

About

templated flame fractal renderer written in c++ supporting output as image or the internal buffer

Topics

Resources

License

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published