Im2Vec: Synthesizing Vector Graphics Without Vector Supervision | Hacker News Reader

Im2Vec: Synthesizing Vector Graphics Without Vector Supervision

arxiv.org

91 pointstomduncalf16 comments

Wish I had the research budget to work on the intersection of SVG and ML. I feel there is a rich vein awaiting discovery apart from the semantic stuff. Compact, efficient machine representations that consume 10x less memory for large SVG images, for example.

Isn't this like a solved problem for ages?

This should be an autorespond link, but [cite]?

A few years back I had the idea to morph fonts using vector graphic interpolation. Unfortunately it was a bit more complicated than I had anticipated and as usual it was abandoned in favor of newer, shinier projects. But it's still fun to play around with:

https://heartofnoise.com/fontmix/

This is a standard feature of modern (variable) fonts [0]

However, they are limited to linear interpolation of the control point polygon, IIRC.

If you want arbitrary morphing (of the actual curve) you might want to look into implicit curves [1] represented in geometric algebra [2]. Here [3] is an example of it.

[0] https://v-fonts.com/

[1] https://en.wikipedia.org/wiki/Implicit_curve

[2] https://en.wikipedia.org/wiki/Geometric_algebra

[3] https://enkimute.github.io/ganja.js/examples/coffeeshop.html...

That's all great stuff, but my idea was rather to explore the latent space inbetween existing fonts, so to speak. V-fonts are nice but very limited. Implicit curves are cool but I don't see how they could represent existing font data?

>> the latent space in between existing fonts

Packing multiple different fonts into one parametrizable one is the inverse of using one parametrizable font to generate multiple different fonts, right?

>> V-fonts are nice but very limited

As I already mentioned, therefore I agree.

>> Implicit curves are cool but I don't see how they could represent existing font data?

It might not be obvious from the wikipedia page, but implicit curves are not a class of curves but a different way to formalize them (as opposed to parametric curves). So you can think of / deal with the bezier curve segments either as parametric curves or as implicit curves.

Related work from 2008: Mona Lisa in 50 polygons, using a genetic algorithm.[0][1] Sadly I think they refused to release their source, but I believe others have re-implemented it.

[0] https://news.ycombinator.com/item?id=4912964

[1] https://rogerjohansson.blog/2008/12/07/genetic-programming-e...

For example. Dave Hoskins (shadertoy) To get the triangle look comment out #define ADD_DITHER and press the Compile (play) button. https://www.shadertoy.com/view/MsX3WH

differentiable rasterization pipeline

That's really damn cool.

This is a dumb question, I'm sure:

Does the loss function have to be a function of the derivative somehow? Or does being differentiable just allow gradient descent learning to work?

> Source code, datasets and more results are available at http://geometry.cs.ucl.ac.uk/projects/2020/Im2Vec/.

When I visit that URL all I get currently is the PDF of the paper?

Seems they got the wrong URL, http://geometry.cs.ucl.ac.uk/projects/2021/im2vec/ has more info, also the link to a github repo: https://github.com/preddy5/Im2Vec

So the code has never been published? Edit: seems to be here https://github.com/preddy5/Im2Vec

I was really hopping this was some chain of CV, NLP, and GANs to implement word2vec but for images. I think I just found one of my new side-projects for the year.

https://github.com/preddy5/Im2Vec

Reply on news.ycombinator.com