Open-source project
rh12503/triangula avatar
rh12503/triangula

Triangula's newest release is from 2021, and its images are drawn in two commands

Generate high-quality triangulated and polygonal art from images.

3,876 stars120 forksGoMIT

At a glance

What is it?
Triangula fits a mesh of points to an image with a modified genetic algorithm, in a library, a desktop app and a command line tool that live in three separate repositories. Its newest tag is five years older than its last commit, its documented installer uses the older go get idiom, one default is an exponent rather than a count, and the guarantee it advertises against its competitors is that the triangles never overlap.
Who is it for?
Triangula suits someone who wants a low-poly look out of a photograph and is happy to wait a couple of minutes for it, and the two-step design is worth understanding before you start: the first command only searches and writes a JSON file, and the second reads that file and draws. Two things to weigh.
Can I use it commercially?
Yes. MIT is a permissive licence: you can use, modify and sell software built on it, as long as you keep its copyright and licence notices.
Is it still maintained?
Activity is slowing. The repository last received commits 8 months ago.
What is it written in?
Mainly Go, according to GitHub's language statistics.

Answers come from the project's GitHub data, last synced on November 15, 2026, and from our analysis. They are not legal advice.

From daily snapshots

Star history

  • rh12503/triangula3,876 ★ · +2
Star history
Repository2026-09-242026-11-15
rh12503/triangula3.9k3.9k

2 Stars in the last 52 days, measured from the daily GitHub snapshot. Compare with other projects

Editorial analysis

Five years of commits past the newest release

The release history stops in 2021. The newest tag is version one point two, titled for the feature it added, polygonal art, on 2021-05-21. Before it, one point one point two, a dark mode change, and one point one point one, a major bug fix, both within two days of each other in April 2021. The default branch has moved on: the last commit is dated 2026-03-21, so there are close to five years of changes that no tag covers and no release note describes. The readme also leads with a browser demo and then tells you the desktop app is typically twenty to fifty times faster, which is a wide claim with a range attached and no measurement behind it, and it is the difference between the two ways most people will meet this project.

The library, the desktop app and the tool are three repositories

The install section is really three projects. This repository is the library. The desktop application lives in a separate repository and is built with a framework that wraps a web frontend in a native binary, and its installation instructions point at a releases page, which happens to be this repository's. The command line tool is a third repository, installed with the older idiom for putting a binary in place:

bash
go get -u github.com/RH12503/Triangula-CLI/triangula

After that, three different paths are given for the directory your Go tools land in: the user directory under your home on macOS, a Go path variable on Linux, and the Go installation directory on Windows, which is a different scheme from the other two and points at where Go itself was installed rather than where a user binary would go.

Running is two commands and the first one only searches

The workflow is deliberately split. The first command takes an image and writes a JSON file, and nothing is drawn:

bash
triangula run -img <path to image> -out <path to output JSON>
triangula render -in <path to outputted JSON> -img <path to image> -out <path to output SVG>

The readme says to run the second one when you are happy with the result's fitness, which gives away that the search is a genetic algorithm scoring candidate meshes. Keeping the two steps apart is what makes a good result reusable: the JSON is the artifact, so you can iterate on rendering without paying for the search again. It is also the honest answer to a question the readme does not raise, which is that the same image will not produce the same mesh twice, and a result you like is worth keeping rather than rerolling.

Nine options, and the advice is to change one

The options table has nine rows, each with a long flag, a short flag, a default and a description, and the guidance above it is that for almost all cases you should change only the number of points and leave the rest. So the interesting rows are the defaults, and three of them are not what they look like. The cache size defaults to twenty two and is described as a power of two, so the real figure is millions, not twenty two entries. The thread count defaults to zero, and zero does not mean no threads; it means use every core. And the generations-before-saving option is the only one marked as specific to the command line tool, so the desktop application cannot set it. The alias counts are inconsistent too: points and threads have one short flag, mutations has two.

The guarantee against the competitors is that triangles never overlap

The readme compares itself to three other projects, and the comparisons are unusually candid. Against two similar tools, it notes that many people have commented on the resemblance and answers with one sentence: the others allow triangles to overlap, while this produces a triangulation. That is the whole technical difference claimed. Against the closest project, also written in Go, it concedes the advantage: that one generates an image almost instantaneously while this one needs many iterations, and the styles differ. It also documents where the pictures came from, which is the detail worth having: the rival's results were taken from the rival's own repository, and these were generated here over one to two minutes. Both comparison tables, and both sets of example images, are empty in the document as published.

The API hands you factories, and the example hardcodes its point count

The library interface is configured with code rather than an options struct. The example opens and decodes a file, converts the decoded image into the library's own data type, then supplies two factories: one returning a function that generates random points, and one returning an evaluator given a size. The point factory in the example generates two hundred points, while the command line tool's default is three hundred, so the documented example and the documented default do not agree. The evaluator factory's comment repeats the two numbers that matter most, twenty two for the cache and five for the block size, which are the same values as the command line defaults and are otherwise only described in prose. The example is cut off inside that comment.

A Go 1.16 module with three random number generators

The module file is a small archaeology of its own. It declares language version one point one six, which is the version current when the last release shipped. Its dependencies include a PCG random number generator and a fast generator and a xorshift generator, two of them indirect, so the project pulls in three different sources of randomness for a stochastic search and is explicit about which one it uses directly. The triangulation primitive is a commit from 2018 rather than a release, and the map projection library that appears at all, marked indirect, is a hint that points are being placed in a projected space rather than in raw pixels. The test library is also from 2019, and exactly one dependency carries a normal version number.

A file with no extension sits in the root

The repository root is a directory per pipeline stage, fourteen of them, including folders named for the genetic algorithm's own vocabulary: fitness, mutation and generator. Two root entries are odd. One is a file with the project's own name and no extension, which in a Go project is data rather than code, and the readme does not mention it. The other is a documentation file with an uppercase extension, sitting next to a conventionally cased readme. The contribute section is the most useful forward-looking text in the file: it asks for effects exported into the output vector format, for more image types on both front ends such as three formats it names, for drag and drop from a browser, and for optimisations. The first of those is narrower than it sounds, since the render step already writes that format; what is missing is effects inside it.

Editorial conclusion

Triangula suits someone who wants a low-poly look out of a photograph and is happy to wait a couple of minutes for it, and the two-step design is worth understanding before you start: the first command only searches and writes a JSON file, and the second reads that file and draws. Two things to weigh. The newest release is from 2021 while the branch has been worked since, so pin a commit rather than a version, and expect the desktop app to be dramatically faster than the browser demo the readme links first.

And one thing to do rather than weigh. Because the search is stochastic, treat the JSON as the artifact you keep: the run step is the expensive half and the render step is cheap, so once you have a mesh you like, save it and re-render from it instead of searching again. That habit, more than any setting, is what makes this tool pleasant to iterate with. And the search is stochastic, so the same image will not give you the same mesh twice; if a result is close, keep the JSON rather than rerunning, which is also why the two commands are separate.

Frequently asked questions

What does it mean to triangulate, or what is triangulation in math?

In geometry it is the decomposition of a shape into triangles. In this repository it means something narrower: fitting a mesh of points to a picture and optionally reducing it to polygons, with a modified genetic algorithm searching for a good result, which the readme expects to take a couple of minutes on an image under 3000 pixels.

What are gunshot location systems?

Not this project. Searches for the name mostly return a different field, systems that infer where a shot came from by comparing the sound's arrival time at several microphones. The repository here is an image-to-mesh tool written in Go, published under a permissive licence with a browser demo, a desktop app and a command line tool in separate repositories.

What is triangular in shape?

Not this repository either. It triangulates images: it fits a mesh of points to a picture and can emit a triangle mesh or polygons, producing a stylised low-poly result rather than a statement about geometric shape. It also guarantees that the triangles do not overlap, which is the difference it claims over the two projects it is most often confused with.

How do I install and run Triangula?

There is a browser demo, a desktop application installed from a releases page and built with a web-to-desktop framework, and a command line tool installed with a go get command after which your PATH must include your Go binary directory. The tool runs in two phases, writing JSON from an image and then rendering that JSON to an SVG.

How many points should I use with Triangula?

The default is 300, and the readme says to change only that number and leave every other option alone, describing that as producing an optimal result in almost all cases. The same guidance notes the algorithm works best with images smaller than 3000 pixels and fewer than 3000 points, and expects a couple of minutes to run.

Official sources

  1. Issues
  2. License: MIT
  3. README
  4. Releases
  5. rh12503/triangula on GitHub
Add this badge to your README

If you maintain this project, the badge below links readers to this analysis and shows its maintenance status from the daily GitHub snapshot. Paste the markdown into your README; add ?metric=license or ?metric=stars to the image URL for a different field.

Add this badge to your README

markdown
[![Hysen Labs](https://hysenlabs.com/badge/rh12503-triangula.svg)](https://hysenlabs.com/projects/rh12503-triangula)

Star history chart for your README

markdown
[![Star history](https://hysenlabs.com/chart/rh12503-triangula.svg)](https://hysenlabs.com/projects/rh12503-triangula)

Was this useful?