Elamur Research⌕
☰
home / papers / 2026 / Vector Calligrapher: Generating Scalable Vector Graphics via Structured Linguistic Supervision

Vector Calligrapher: Generating Scalable Vector Graphics via Structured Linguistic Supervision

Paper · 2026 · Bo Zhou, Xikang Chen, Yan Gong et al.

Summary

Multimodal LLMs struggle to turn a textual style request into an SVG font glyph for two reasons. Free-form captions describe typographic style too loosely to control exact geometry. Generic tokenizers also split coordinate-heavy SVG markup into long token streams that carry little information each. Vector Calligrapher tackles both problems by treating glyph synthesis as conditional language modeling with a Qwen-based MLLM. Its Font Description Framework breaks style into separate, interpretable axes, such as historical period, serif/script structure, stroke contrast, mood and metaphor. These axes are encoded as discrete tagged tokens placed in front of the SVG sequence. For geometry, paths are normalized to a fixed view-box, and the x and y coordinates are quantized and emitted as independent tokens. As a result, vocabulary size grows linearly with resolution rather than quadratically. The model is trained on VectorFont, a new set of more than 10M openly licensed glyphs with automatically generated hierarchical descriptions. Against unstructured baselines, the system raises CLIP score, roughly halves geometric error and fits far more drawing commands into each token. This suggests that structured supervision plus compact encoding makes controllable vector generation more dependable.

Key points

  • VectorFont dataset: contains over 10M glyphs from open-source font collections such as Google Fonts, mostly under SIL OFL or Apache licenses, and is skewed toward Latin and Chinese scripts.
    • Qwen2.5-VL-70B produced all style annotations with no human annotation.
    • A post-hoc human check of 500 samples found 96.4% accuracy.
  • Reported gains: CLIP score improves by 23%, and Chamfer Distance drops by about 48% compared with unstructured baselines.
    • Tasks covered include text-to-SVG, raster-to-SVG and stylization/style editing.
  • Separated-coordinate tokenization: coordinates are normalized to [−1000, 1000] and x/y are tokenized independently, giving O(N) vocabulary instead of flattened y·W+x tokens.
    • It reaches an 18% Commands-per-Token (C/T) ratio, about 6× the information density of standard baselines.
  • Stated limitations:
    • Evaluation covers single glyphs only, with no metric for family-level consistency.
    • Generation is limited by a 150-token budget.
    • Baselines were not retrained on VectorFont.
    • The causal effect of the style axes has not been demonstrated.
  • Release: the authors promise to publish the dataset, checkpoints, code and evaluation scripts at github.com/baudzhou/vector-calligrapher.

Our summary of the paper, in our own words — not its text. The work belongs to its authors; read it at the links above.

 Institutions

🎓Zhejiang UniversityHangzhou · China224 works · 13,587 citations

 Fields, topics and concepts

📚 Shelves 5The field of science this work belongs to (OpenAlex fields).Computer Science · 18,614 worksArtificial Intelligence · 17,965 worksArts and Humanities · 17,450 worksLanguage and Linguistics · 17,296 worksComputer Vision and Pattern Recognition · 3,486 works

Cite

Cite as Bo Zhou, Xikang Chen, Yan Gong and Yin Zhang. 2026. Vector Calligrapher: Generating Scalable Vector Graphics via Structured Linguistic Supervision. Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers).

The paper itself, not this page. BibTeX from the library's record of the work. DOI 10.18653/v1/2026.acl-long.511

@inproceedings{zhou2026vector,
  title = {{Vector Calligrapher: Generating Scalable Vector Graphics via Structured Linguistic Supervision}},
  author = {Bo Zhou, Xikang Chen, Yan Gong and Yin Zhang},
  booktitle = {Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers)},
  year = {2026},
  doi = {10.18653/v1/2026.acl-long.511},
  url = {https://doi.org/10.18653/v1/2026.acl-long.511}
}

References

21 references, 7 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

  1. Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, and Shyamal Anadkat. 2023. GPT-4 technical report. arXiv preprint arXiv:2303.08774.
  2. Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katherine Millican, and Malcolm Reynolds. 2022. Flamingo: a visual language model for few-shot learning. Advances in neural information processing systems, 35:23716–23736.
  3. Samaneh Azadi, Matthew Fisher, Vladimir G Kim, Zhaowen Wang, Eli Shechtman, and Trevor Darrell. 2018. Multi-content gan for few-shot font style transfer. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 7564–7573. find
  4. Shuai Bai, Keqin Chen, Xuejing Liu, Jialin Wang, Wenbin Ge, Sibo Song, Kai Dang, Peng Wang, Shijie Wang, Junyang Lin, Chang Zhou, and Jingren Zhou. 2025. Qwen2.5-VL technical report. arXiv preprint arXiv:2502.13923.
  5. Ayan Banerjee, Nityanand Mathur, Josep LladÃg̀s, Umapada Pal, and Anjan Dutta. 2024. SVGCraft: Beyond single object text-to-svg synthesis with comprehensive canvas layout. arXiv preprint arXiv:2404.00412. graphics code from images and text. In Proceedings of the Computer Vision and Pattern Recogniti… — [arXiv:2404.00412](https://arxiv.org/abs/2404.00412) find
  6. Alexandre Carlier, Martin Danelljan, Alexandre Alahi, and Radu Timofte. 2020. DeepSVG: A hierarchical generative network for vector graphics animation. Advances in Neural Information Processing Systems, 33:16351–16361. find
  7. Alibaba Cloud / Qwen Team. 2025. Qwen2.5-VL-3B-Instruct: A 3b-parameter multimodal model for image, video, and text processing. Online Model Report. find
  8. Mark Chen, Jerry Tworek, Heewoo Jun, Qiming Yuan, Henrique Ponde De Oliveira Pinto, Jared Kaplan, Harri Edwards, Yuri Burda, Nicholas Joseph, and Greg Brockman. 2021. Evaluating large language models trained on code. arXiv preprint arXiv:2107.03374.
  9. Ximing Xing, Haitao Zhou, Chuang Wang, Jing Zhang, Dong Xu, and Qian Yu. 2024. SVGDreamer: Text guided svg generation with diffusion model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4546–4555. find
  10. Charles J Fillmore. 2006. Frame semantics. Cognitive linguistics: Basic readings, 34:373–400. find
  11. Yiying Yang, Wei Cheng, Sijin Chen, Xianfang Zeng, Fukun Yin, Jiaxu Zhang, Liao Wang, Gang Yu, Xingjun Ma, and Yu-Gang Jiang. Omnisvg: A unified scalable vector graphics generation model. In The Thirty-ninth Annual Conference on Neural Information Processing Systems. find
  12. Adele E Goldberg. 2003. Constructions: A new theoretical approach to language. Trends in cognitive sciences, 7(5):219–224. find
  13. Haibin He, Xinyuan Chen, Chaoyue Wang, Juhua Liu, Bo Du, Dacheng Tao, and Qiao Yu. 2024. Diff-font: Diffusion model for robust one-shot font generation. International Journal of Computer Vision, 132(11):5372–5386. find
  14. Peter Karow. 1994. Font technology. In Font Technology: Methods and Tools, pages 85–104. Springer. find
  15. Donald Ervin Knuth. 1986. The metafontbook. Addison-Wesley Longman Publishing Co., Inc. find
  16. Yujia Li, David Choi, Junyoung Chung, Nate Kushman, Julian Schrittwieser, Rémi Leblond, Tom Eccles, James Keeling, Felix Gimeno, and Agustin Dal Lago. 2022. Competition-level code generation with alphacode. Science, 378(6624):1092–1097.
  17. Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. 2023. Visual instruction tuning. Advances in neural information processing systems, 36:34892– 34916.
  18. Raphael Gontijo Lopes, David Ha, Douglas Eck, and Jonathon Shlens. 2019. A learned representation for scalable vector graphics. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7930–7939. find
  19. Stanislas Polu and Ilya Sutskever. 2020. Generative language modeling for automated theorem proving. arXiv preprint arXiv:2009.03393.
  20. Pradyumna Reddy, Michael Gharbi, Michal Lukac, and Niloy J Mitra. 2021. Im2vec: Synthesizing vector graphics without vector supervision. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7342–7351. find
  21. Juan A Rodriguez, Abhay Puri, Shubham Agarwal, Issam H Laradji, Pau Rodriguez, Sai Rajeswar, David Vazquez, Christopher Pal, and Marco Pedersoli. 2025. StarVector: Generating scalable vector find

Updated

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/en/papers/zhou-2026-vector-calligrapher-generating-scalable-vector/