5 ms·

Was this converted by hand or by using some vectorizer? Really impressive! Anyway, are there any ML-powered vectorizers out there?

by Jerry2 3y ago

murphyslab 3y ago
Parts of this definitely appear to have been traced, although I'm not sure if it was done manually or by a vectorizer. Around the `OF` in the `FIELD LIST` flow, the green edge has the same perspective change as in the original, cf: https://www.attentionspan.nl/posts/technology/apple-pascal-syntax-poster/ https://www.attentionspan.nl/posts/technology/apple-pascal-s... Lots of other little peculiarities of the original seem to be preserved. So it definitely wasn't re-made solely from reference.
nanoraptor 3y ago
Essentially all by hand. I'd tried to do it a few times before with different quality scans, and while I could see the original intent on all of them, auto-tracing never held up across the whole poster. I ended up taking a few days zoning out and re-drawing it as a white/black raster image over a highly scaled up scan, then auto-traced that in Illustrator and coloured it there. Text went in as some variant of Century Gothic with a few changes to match the original.
- klabetron 3y ago
  Great work!
- skywal_l 3y ago
  Do you provide the vectorized version? It could be used to restore the original colors [0]. [0] https://news.ycombinator.com/item?id=38876295 https://news.ycombinator.com/item?id=38876295
  - JKCalhoun 3y ago
    I opened the file (PDF) in Affinity Designer (a commercial vector app) and the paths are all there. I can at that point easily export as SVG.
kekub 3y ago
I had some good results with https://vectorizer.ai/ https://vectorizer.ai/
justinl33 3y ago
I think vector magic is the current state of the art: https://vectormagic.com/?=20 https://vectormagic.com/?=20 No one seems to have tried to leverage deep learning yet; either because they haven't thought of doing so, or it just wouldn't be worthwhile. image to SVG's are an inherently deterministic task, with not much room for the noisy error of most deep learning models like stable diffusion and such. I think algorithmic approaches are just better in this case - deep learning isn't always superior.
- gwern 3y ago
  Raster2vector is a researched area, going well back; most notable recent work would be StarVector https://browse.arxiv.org/html/2312.11556v1 https://browse.arxiv.org/html/2312.11556v1
justinl33 3y ago
5 years ago... someone tried abusing a text generating RNN to produce broken SVG, and then tidy it up. Genuinely curious as to how this might work with today's transformers. https://www.reddit.com/r/MachineLearning/comments/cnip4m/p_abusing_text_synthesis_into_svg_image/ https://www.reddit.com/r/MachineLearning/comments/cnip4m/p_a... https://github.com/artBoffin/GAN-XML-Fixer https://github.com/artBoffin/GAN-XML-Fixer
tempodox 3y ago
There's the Railroad Diagram Generator: https://github.com/GuntherRademacher/rr https://github.com/GuntherRademacher/rr

Was this converted by hand or by using some vectorizer? Really impressive! Anyway, are there any ML-powered vectorizers out there?

by Jerry2 3y ago

Was this converted by hand or by using some vectorizer? Really impressive! Anyway, are there any ML-powered vectorizers out there?

murphyslab 3y ago

Parts of this definitely appear to have been traced, although I'm not sure if it was done manually or by a vectorizer. Around the `OF` in the `FIELD LIST` flow, the green edge has the same perspective change as in the original, cf: https://www.attentionspan.nl/posts/technology/apple-pascal-syntax-poster/ https://www.attentionspan.nl/posts/technology/apple-pascal-s... Lots of other little peculiarities of the original seem to be preserved. So it definitely wasn't re-made solely from reference.

nanoraptor 3y ago

Essentially all by hand. I'd tried to do it a few times before with different quality scans, and while I could see the original intent on all of them, auto-tracing never held up across the whole poster. I ended up taking a few days zoning out and re-drawing it as a white/black raster image over a highly scaled up scan, then auto-traced that in Illustrator and coloured it there. Text went in as some variant of Century Gothic with a few changes to match the original.

klabetron 3y ago
Great work!
skywal_l 3y ago
Do you provide the vectorized version? It could be used to restore the original colors [0]. [0] https://news.ycombinator.com/item?id=38876295 https://news.ycombinator.com/item?id=38876295
- JKCalhoun 3y ago
  I opened the file (PDF) in Affinity Designer (a commercial vector app) and the paths are all there. I can at that point easily export as SVG.

kekub 3y ago

I had some good results with https://vectorizer.ai/ https://vectorizer.ai/

justinl33 3y ago

I think vector magic is the current state of the art: https://vectormagic.com/?=20 https://vectormagic.com/?=20 No one seems to have tried to leverage deep learning yet; either because they haven't thought of doing so, or it just wouldn't be worthwhile. image to SVG's are an inherently deterministic task, with not much room for the noisy error of most deep learning models like stable diffusion and such. I think algorithmic approaches are just better in this case - deep learning isn't always superior.

gwern 3y ago
Raster2vector is a researched area, going well back; most notable recent work would be StarVector https://browse.arxiv.org/html/2312.11556v1 https://browse.arxiv.org/html/2312.11556v1

justinl33 3y ago

5 years ago... someone tried abusing a text generating RNN to produce broken SVG, and then tidy it up. Genuinely curious as to how this might work with today's transformers. https://www.reddit.com/r/MachineLearning/comments/cnip4m/p_abusing_text_synthesis_into_svg_image/ https://www.reddit.com/r/MachineLearning/comments/cnip4m/p_a... https://github.com/artBoffin/GAN-XML-Fixer https://github.com/artBoffin/GAN-XML-Fixer

tempodox 3y ago

There's the Railroad Diagram Generator: https://github.com/GuntherRademacher/rr https://github.com/GuntherRademacher/rr