StarVector: Generating Scalable Vector Graphics Code from Images and Text

72 pointsposted 12 days ago
by lnyan
(starvector.github.io)

7 Comments

rahimnathwani

8 days ago

  Note: StarVector models will not work for natural images or illustrations, as they have not been trained on those images. They excel in vectorizing icons, logotypes, technical diagrams, graphs, and charts.

peter_d_sherman

12 days ago

This would be absolutely GREAT for generating icons for applications!

(Also would make a great SaaS... for $X/month ($9.95, $19.95, ??.??) generate unlimited icons...)

Congrats to the team for their pioneering hard work in this nascent area of LLM/Transformer research!

Well done!

smahs

7 days ago

No it won't (most likely). VTracer (which the authors compare with) is fast, runs in browser via wasm, consumes way less resources and can even convert natural images very decently. But the model seems cool for the usecase of prompt to logo or icon (over my current workflow of getting a jpg from flux and passing it through VTracer). I hope someone over at llama.cpp notices this (at least for the text-to-svg usecase, if not multimodal).

jasonjmcghee

7 days ago

Did anyone else notice - the molecule it generated did not match the source image.

Seems like this could be incredibly valuable, but I'd argue there needs to be validation steps in place to confirm it's actually generating the right thing, for the case of image -> vector generation.

aikinai

8 days ago

I've been waiting for someone to make something like this! Perfect.

iFire

12 days ago

Horray. Code is released