Check out our recent work on improving image vectorization using gradient descent: https://people.csail.mit.edu/tzumao/diffvg/

Those figures look amazing! Also, its crazy how many things can be improved by a clever use of gradient descent

Had to think of the paper about depixelizing pixel art, fun stuff if you're interested:

https://johanneskopf.de/publications/pixelart/paper/pixel.pd...

Beautifully written and illustrated. Makes the algorithm clear and seems to give everything you'd need to implement it (afraid to test this assumption because I'm afraid it would be too much fun, and I have deadlines ;)

It's indeed a very good read! I have done a similar thing, with a completely different approach, because my math is too bad. https://github.com/visioncortex/vtracer

It works quite well actually. Nice work.

I think this is the first time I read a paper that includes real code and even links to a github repo. Amazing.

You should check out https://paperswithcode.com/

Not just real code, live code

Not having looked at vectorization so far I wondered if some ideas from CT reconstruction can be applied to image vectorization. In both cases you start with a base image that is projected on one (or in the case of tomography several) pictures. The goal is to retrieve the original.

https://en.m.wikipedia.org/wiki/Tomographic_reconstruction

Inkscape includes a tool like this: https://inkscape.org/doc/tutorials/tracing/tutorial-tracing....

I have badly needed vectorization tools for client side geomatics work. But I haven't found anything that works in thr browser. Going to have to check out the examples here but if anyone wants to make a client side library around it, I think there's demand!

Here it is https://www.visioncortex.org/vtracer/ Have a nice day!

Similar, public domain: https://github.com/jankovicsandras/imagetracerjs

The site is full of great interactive explanations. Nice work!