Towards Readable Scalable Vector Graphic Generation
Primary Area: societal considerations including fairness, safety, privacy
Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.
Keywords: Scalable Vector Graphic, Readability
Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2024/AuthorGuide.
Abstract: The surge in the use of Scalable Vector Graphics (SVGs) for digital graphics, particularly with advancements in generative models, has seen a proliferation in the automatic creation of SVGs. Yet, as these models prioritize visual accuracy, they often neglect the readability of the underlying SVG code. However, the readability of the SVG code is equivalently, if not more, important in comparison to visual accuracy, for the convenience of editing and logical inference for downstream tasks. Therefore, this paper delves into the overlooked realm of SVG code readability, emphasizing its importance in ensuring efficient comprehension and modification of the generated graphics. Readability, encompassing aspects like logical struc- turing and minimized complexity, plays a pivotal role in ensuring SVGs are not just visually accurate but also human-friendly at the code level. We first propose a clear set of desiderata for SVG code readability, serving as a foundation for our subsequent analyses. Leveraging this, we introduce a set of dedicated metrics to evaluate SVG readability and design differentiable objectives to guide SVG gener- ation models towards producing more readable code. Our evaluation reveals that while most SVG generators can produce visually accurate graphics, the underlying code often lacks structure and simplicity. However, with our proposed metrics and objectives, SVG generators exhibit significant improvements in code readability without compromising visual accuracy.
Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors' identity.
No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.
Submission Number: 4809
Loading