SVG-LLM

Inspiration

SVG images are very special (and unlike PNG or JPEG images) in that scaling them up or down does not result in a loss of quality. This is because SVG images are vector graphics, which means that they are defined by mathematical equations rather than pixels. Unfortunately, this also makes them difficult to render using traditional LLMs. We want to build an LLM which is fine-tuned on a dataset of SVG images. This allows it to generate SVG images from text descriptions, making it easy to create vector graphics without needing to know how to use a graphic design tool.

What it does

Generate accurate SVGs from user text

How we built it

We constructed a dataset of PNG, SVG image pairs. SVGs where gathered from a datasewt, and these were fed into a tool which converts SVG into PNG. We fed these pairs into a gemini model (with SVG being ground truth), and finetuned the model.

Challenges we ran into

NLX is not intuitive or easy to use to host LLM workflows

Accomplishments that we're proud of

What we learned

What's next for SVG-LLM

Actually getting access back to out Gemini model (it used up WAY too much credit without warning, and the payment was declined, so the model is kind of in limbo). Also, different file formats, so maybe taking 3d-looking images and creating a .dwg file which can be easily edited in AutoCAD or Solidworks