spleeter-web VS procedural-gl-js

or https://github.com/JeffreyCA/spleeter-web

There's a webservice to separate a song using Demucs here:

https://demucs.danielfrg.com/

And a great Dockerized webapp that lets you choose from several models and parameters here:

https://github.com/JeffreyCA/spleeter-web

Otherwise you can just install them locally and run them through the CLI, it's pretty easy (one command)

https://github.com/facebookresearch/demucs#for-musicians

https://github.com/deezer/spleeter#quick-start

  2. Now you can take the isolated vocals, and build the rest of the song yourself

For the more tech-savvy folks and programmers alike: I've been working self-hostable web app that does this and supports Spleeter and many other models (fully open source): https://github.com/JeffreyCA/spleeter-web

Spleeter Web, a self-hostable web app for music source separation. It lets you isolate the vocal, accompaniment, bass, and/or drums of any piece of music using deep learning-based source separation models. It's like moises.ai/ezstems.com but open-source.

I learned lots about building a full-stack web app ground-up as well as how to containerize the whole thing with Docker.

GitHub: https://github.com/JeffreyCA/spleeter-web

Nice writeup, I always like it when the shaders are highlighted like this. I got started in a similar way 7 years ago and have been making 3D terrains with THREE.js & WebGL since.

The real fun begins when you need to implement some sort of Level-of-Detail system and streaming in data to give the illusion of high detail everywhere without sacrificing performance.

Last year I released an open-source framework (https://github.com/felixpalmer/procedural-gl-js) for creating 3D terrains for web applications, you can see Uluru here: https://www.procedural.eu/map/?longitude=131.036&latitude=-2... (unfortunately the aerial imagery from our default provider isn't as high resolution as other places in Europe)

Great visualization and approach with compressing the tile data. Do you have a comparison of how much smaller the payload ends up being compared to simply sending PNG files?

I use PNGs to encode elevation data in my 3D mapping library (https://github.com/felixpalmer/procedural-gl-js/) and this does a pretty good job of compressing the data, for example in the ocean the PNG files are also very small as the image is mostly black. Different use case I now as your data is much more sparse, but I wonder how close the PNG compression would be compared to your approach.

Yeah, the React parts of this are very minimal. I'm not really sure what using it gets you, since it just manages a single div.

The _actual_ library that does all the work is here: https://github.com/felixpalmer/procedural-gl-js

- Tiny filesize means library is parsed fast. Package size is less than THREE.js thanks to code stripping

Check it out on Github: https://github.com/felixpalmer/procedural-gl-js/

The imagery comes from the Orthofoto dataset on https://www.basemap.at/ - the actual texturing is done by the Procedural GL JS library https://github.com/felixpalmer/procedural-gl-js