Previously: 241009-1513 Symbols project links

Some more new/fresher thoughts on this. Mostly focused on the SVG side of this.

• Unsymbols will be the name I use in my notes for this.
  • Interestingly it’s not even taken.

ML

SVG VAE (2024)

• NN / GAN that works on SVGs, incl. for English alphabet, really cool
  • Blog post: SVG VAE: Generating Scalable Vector Graphics Typography
  • [1904.02632] A Learned Representation for Scalable Vector Graphics
  • Code and stuff:
    • Main: magenta/magenta/models/svg_vae at main · magenta/magenta
    • Colab: magenta-demos/colab-notebooks/vae_svg_decoding.ipynb at main · magenta/magenta-demos
  • Dataset: A-Za-z0-9 based on 66k fonts w/ download links given in .txt

DeepSVG (2020)

• alexandre01/deepsvg: [NeurIPS 2020] Official code for the paper “DeepSVG: A Hierarchical Generative Network for Vector Graphics Animation”. Includes a PyTorch library for deep learning with SVG data.
• allegedly better than SVG-VAE for font generation (p.8)
• Animations between icons but not just that
• Interpolations a la “squarify” (minus $\Delta_{square}$) and between icons (p.7 of paper)
• They do font generation!
• really cool appendix with details of SVG preprocessing etc., normalization

IconShop: Text-Guided Vector Icon Synthesis with Autoregressive Transformers

• [2304.14400] IconShop: Text-Guided Vector Icon Synthesis with Autoregressive Transformers

  •  The key to success of our approach is to sequentialize and tokenize SVG paths (and textual descriptions as guidance) into a uniquely decodable token sequence. With that, we are able to fully exploit the sequence learning power of autoregressive transformers, while enabling both unconditional and text-conditioned icon synthesis.

  • #good

  •  which is validated by the objective Uniqueness and Novelty measures.

  • https://icon-shop.github.io/
• Tokenization of SVGs!
• I think we can work with this
  • instead of words do letters, or groups of letters
• kingnobro/IconShop: (Siggraph Asia 2023) Code of “IconShop: Text-Guided Vector Icon Synthesis with Autoregressive Transformers”
• code is readable
• transformer-based

Other

• SVGFormer
  • Newer than deepSVG
  • reconstruction: I guess pic -> SVG
  • Compare w/ or use DeepSVG
• LLMs to work on SVGs
  • An Investigation on LLMs’ Visual Understanding Ability Using SVG for Image-Text Bridging

Tangentially related

• Unicode
  • List of Latin letters by shape - Wikipedia
  • A | Unicode discussion | Fandom
  • List of Unicode Characters with Decomposition Mapping “Font variant” #good
• DeepVecFont | deepvecfont_homepage
  • yizhiwang96/deepvecfont: [SIGGRAPH Asia 2021] DeepVecFont: Synthesizing High-quality Vector Fonts via Dual-modality Learning
  • generate/predict fonts based on letters
  • 
• clovaai/dmfont: Official PyTorch implementation of DM-Font (ECCV 2020)

  • In this paper, we focus on compositional scripts, a widely used letter system in the world, where each glyph can be decomposed by several components. By utilizing the compositionality of compositional scripts, we propose a novel font generation framework,

Datasets

SVG

• googlecreativelab/quickdraw-dataset: Documentation on how to access and use the Quick, Draw! Dataset. drawings by categories

Misc

• ML Model for SVG - Sriram Alagappa - Medium

  • Exploring if there is any good AI model that will natively generate SVG

  • TODO this is excellent
• Magenta
  • “An open source research project exploring the role of machine learning as a tool in the creative process.”
  • SVG VAE above comes from this place
• Calligrapher.ai: Realistic computer-generated handwriting
• The Engineering behind Figma’s Vector Networks
  • really in-depth about SVG/vectors as well as the title — #good

Notes / ideas

• Inkscape can radically simplify SVG shapes, maybe I can use that approach
• How do I work with lines, w/o width?