Pith. sign in

REVIEW 4 major objections 5 minor 74 references

SVGen: Interpretable Vector Graphics Generation with Large Language Models

T0 review · 4 major / 5 minor · reviewed 2026-08-05 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read Small fine-tuned LLMs beat much larger general models at generating SVG code from text.

desk verdict Useful dataset, questionable evaluation: SVGen's own numbers undercut its headline claim, but SVG-1M is a real resource. read the letter →

arxiv 2508.09168 v1 pith:5HBPQQP3 submitted 2025-08-06 cs.LG cs.CV

Feiyu Wang , Zhiyuan Zhao , Yuandong Liu , Da Zhang , Junyu Gao , Hao Sun , Xuelong Li This is my paper
classification cs.LGcs.CV
keywords LargeLanguageModelsScalableVectorGraphicsText-to-SVGCurriculumLearningChain-of-ThoughtReinforcementGRPOGeneration
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

The paper claims that a compact, fine-tuned language model can turn a natural-language description into ready-to-use SVG code more reliably than much larger general-purpose LLMs and faster than optimization-based renderers. To make that work, it introduces SVG-1M, a dataset of one million text-SVG pairs built from standardized icons and auto-generated captions, with a subset annotated with chain-of-thought design steps. The training recipe combines curriculum learning (easy monochrome icons first, then multicolor complexity), chain-of-thought supervision for design reasoning, and GRPO reinforcement learning with rewards for structurally complete, path-count-matched output. A sympathetic reader would care because this offers a concrete path to real-time, editable vector graphics from text, a capability that previously required hand drawing or slow iterative optimization.

What carries the argument

The central object is the SVG-1M dataset plus the three-stage training protocol: (1) curriculum learning that orders training data by color category and command count (monochrome easy to monochrome difficult to multicolor easy to multicolor difficult); (2) chain-of-thought annotations that teach the model to articulate a step-by-step design plan before emitting SVG code; (3) GRPO reinforcement learning with a total reward combining an integrity reward (1 if the SVG parses and closes correctly, 0 otherwise) and a path-number matching reward that rewards generating at least as many paths as the reference and decays exponentially when the count falls short.

What would settle it

Take a random sample of 200 SVG-1M captions and have human annotators check whether each icon can be reconstructed from its caption alone; then run SVGen against GPT-4o on 100 captions written by designers (not by GPT-4o) and measure human-judged semantic match. If SVGen does not beat GPT-4o on those human-written prompts, the claim of superior semantic understanding would be contradicted.

Watch

Extended reading notes

Core claim

On its own terms, the paper's discovery is that a training pipeline - progressive curricula over difficulty-graded SVG data, chain-of-thought reasoning labels, and GRPO with integrity and path-match rewards - lets small open-weight LLMs (3B-7B parameters) generate SVG icons that are more semantically faithful, structurally complete, and aesthetically scored than the outputs of frontier models like GPT-4o, Gemini-2.5 Pro, and DeepSeek-R1, while also outpacing diffusion-optimization pipelines in speed. The supporting evidence is a set of automatic metrics (FID, CLIPScore, HPS, aesthetic score) and a blind human panel of 30 designers and engineers scoring SVGen highest on semantic match, visual

Load-bearing premise

The captions that GPT-4o wrote from rasterized icons are treated as accurate ground truth for both training and evaluation; if those captions are ambiguous or wrong, the model learns them and is judged against them, so the reported semantic accuracy may not reflect real user descriptions.

Editorial extensions

If this is right

  • If the central claim holds, design tools can integrate a small, fast model to produce editable vector icons in seconds rather than minutes or hours of optimization.
  • The SVG-1M dataset, with its normalized path commands and explicit difficulty tiers, provides a standard benchmark for future text-to-SVG research.
  • The combination of curriculum learning and chain-of-thought supervision could transfer to other structured code-generation domains, such as CAD drawings or UI layouts.
  • The two reward functions - integrity and path-count matching - offer a template for steering LLMs to produce structurally valid outputs of appropriate complexity in code-generation tasks.
  • A lightweight model beating frontier LLMs on a specialized generation task suggests that domain-specific fine-tuning can be more cost-effective than scaling up model size.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • The captions GPT-4o wrote from rasterized icons may be self-consistent with what the model learned, but they are not the same as a designer's description; a test set with human-written prompts would better isolate semantic understanding.
  • The path-number reward encourages at least as many paths as the reference, but when the reference icon is simpler than the user actually wants, this reward could push the model to add unnecessary complexity.
  • The efficiency comparison measures generation time only, not the total cost of building and fine-tuning on the 1M-pair dataset; an end-to-end cost comparison with zero-shot counterpart models would sharpen the practical advantage.
  • If the curriculum order is truly load-bearing, similar difficulty-based curricula should improve text-to-vector and text-to-diagram generation tasks, which are currently trained on uniformly shuffled data.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

4 major / 5 minor

Summary. SVGen addresses text-to-SVG generation by introducing SVG-1M, a dataset of one million Iconfont-derived SVG icons paired with GPT-4o-generated text descriptions, including a 65,745-sample Chain-of-Thought subset. The method fine-tunes 3B-7B LLMs (StarCoder2, Llama3.1, Qwen2.5 variants) with a curriculum over monochrome and multicolor SVGs, followed by CoT fine-tuning and GRPO reinforcement learning with an integrity reward and a path-number matching reward. The paper reports that SVGen outperforms general LLMs and optimization-based methods in FID, HPS, aesthetic score, token length, and generation time, and presents an ablation study and a human evaluation. The test set consists of 1,143 multicolored SVGs sampled from the same pool used for training, with prompts generated by the same GPT-4o captioning pipeline.

Significance. If the claims hold, SVGen would be a useful contribution: it demonstrates that lightweight open LLMs can produce structurally complex SVG code when given a large, standardized, text-SVG corpus, and it introduces a curriculum plus GRPO recipe that is applicable to code-to-structured-graphics tasks. The public release of code, model, and dataset is a concrete strength, as is the standardization of Iconfont data into a reduced command set. However, the evaluation currently does not support the central 'outperforms general large models' claim. The test distribution overlaps the training distribution, the prompts are auto-annotated by the same teacher model used to create training captions, the primary semantic metric (CLIPScore-T2I) actually favors the general LLMs, and the path-number reward explicitly rewards generating as many or more paths than the reference. These issues are load-bearing for the paper's main assertion.

major comments (4)
  1. [§5, §3] The test set is not independent of the training distribution. Section 5 states that 1,143 samples are randomly selected from the multicolored SVG data, and Section 3 describes that data as the same Iconfont pool used to build the training set. In addition, test prompts are produced by the same GPT-4o image-description pipeline used to annotate training pairs. The observed advantage over zero-shot general LLMs therefore partly measures in-distribution fit rather than text-to-SVG generalization. The central claim in the abstract and §5.1 would require a held-out test set with human-verified prompts or, at minimum, an out-of-distribution benchmark such as LLM4SVG or SVG-Stack to demonstrate transfer.
  2. [§5.1, Table 2] The paper's own semantic metric contradicts the claimed semantic superiority. In Table 2, SVGen's CLIPScore-T2I values are 0.229-0.241, while every general LLM baseline scores 0.259-0.274. The text reports FID, HPS, aesthetic score, token length, and generation time as evidence of superiority, but does not discuss the reversed CLIPScore-T2I trend. Since semantic alignment is one of the two core goals stated in the introduction, this metric must be addressed directly rather than omitted from the narrative. Additionally, FID is reported without confidence intervals or significance tests on a 1,143-image test set; given FID's known variance, the margins shown for several baselines may not be statistically meaningful.
  3. [§4.3, Eq. (4)] The path-number matching reward does not match the reference path count; it explicitly rewards generated path counts that are equal to or greater than the reference count. The equation r_match = max(γ, γ exp(-β(N(ĉ)-N(c_gt)))) yields the maximum reward whenever N(ĉ) >= N(c_gt) and only penalizes under-generation. This is a complexity-envelope reward, not a matching reward. Consequently, the RL phase, evaluated in Figure 4 only through completion rate and average path count, may improve those two proxies by inflating path counts rather than by improving fidelity to the reference structure. The claim that this reward 'enhances the structural realism and accuracy' is unsupported without measuring visual or semantic quality on a held-out set during RL.
  4. [§3, §2, §5.3] The dataset annotation process introduces a circularity risk. CoT annotations are generated by GPT-4o and the base captions are also generated by GPT-4o; the CoT subset is stated to be manually verified, but the base captions and the test prompts are not described as human-verified. Since SVGen is fine-tuned and evaluated on this same auto-annotated distribution, part of its advantage over GPT-4o may come from imitating the teacher's caption style and annotation distribution. The human evaluation (§5.3, Table 4) uses 50 prompts drawn from this same distribution, so it does not mitigate the concern. The authors should report human verification statistics for the base captions, or at least quantify caption noise, and add a human evaluation on novel, out-of-distribution prompts.
minor comments (5)
  1. [§2.2, references] Reference [33] is cited for SVG-VAE, but the listed reference is 'Language models are unsupervised multitask learners' by Radford et al. (2019), which appears unrelated. This appears to be a citation error that should be corrected.
  2. [§4.3, Eq. (3)] The integrity reward notation is corrupted: 'metricą (ď)' and 'Ďint(ď) = Ăą (ď)' contain non-ASCII characters and the meaning of Ă and its relation to the later statement that all reward coefficients are set to 1 are unclear. Please rewrite the equations with standard notation and define all variables.
  3. [§5.1, Table 2] The column 'Avg. Token' is ambiguous: it is not clear whether this is average generated token count, average tokenizer encoding length, and why higher values are interpreted as 'reflecting enhanced detail representation and richer content generation' without a human judgment on over-generation. Also, no error bars are given for any metric; please add standard errors or confidence intervals.
  4. [§5.3, Table 4] The human evaluation description in the supplementary says evaluators scored 200 sets from 50 prompts, while the main text says the evaluation covered 50 representative text prompts. This is consistent if each model produces one SVG per prompt, but the scoring method (10/8/6/4 for four models) means scores are forced rankings, not absolute quality ratings. Please clarify whether the reported scores are rankings and note the limitation that only three of the baselines plus SVGen were included.
  5. [§4.2] The phrase 'manually verify and correct all samples' for 65,745 CoT pairs is a strong claim. Please report the verification protocol, number of verifiers, and inter-annotator agreement. If only a subset was verified, state the actual coverage.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the derivation chain is self-contained and externally grounded.

full rationale

The paper's core claim is that SVGen, fine-tuned on the SVG-1M dataset, produces better text-to-SVG generation than zero-shot general LLMs and optimization-based methods. The target SVG code comes from Iconfont, an external vector-icon repository, and is not generated by any of the compared baseline models; therefore the central generation task has independent grounding. The text descriptions and CoT annotations are produced by GPT-4o, and the test set is sampled from the same multicolored SVG pool with prompts generated by the same captioning pipeline, which raises legitimate concerns about distribution fit and evaluation bias, but this is a data-construction and evaluation-design issue rather than a circular derivation. No parameter is fitted to the test labels and then reported as a prediction; the reward functions in Section 4.3 are training objectives, not test metrics. The paper does not rely on a load-bearing self-citation chain, does not import a uniqueness theorem from the authors' prior work, and does not smuggle in an ansatz via citation. The self-citations present are ordinary related-work references and are not load-bearing for the method's validity. The measured deficits, such as lower CLIPScore-T2I relative to general LLMs, are correctness/interpretation concerns, not circularity. Accordingly, no circular step meets the evidentiary standard required by the review rules.

Assumptions & free parameters 3 free parameters · 4 assumptions · 0 invented entities

The method rests on hand-set hyperparameters (reward weights, difficulty thresholds) and several domain assumptions about the quality of the auto-generated text descriptions, the validity of image-space metrics, and the path-count proxy for complexity. No new physical or conceptual entities are introduced.

free parameters (3)
  • reward coefficients lambda, alpha, beta = 1, 1, 1
    Section 4.3: 'The coefficients of all rewards in the reinforcement learning phase are set to 1.' These weights are chosen by hand and directly shape the RL objective.
  • difficulty thresholds = 50, 200, 100, 200 commands
    Section 3, Table 1: classification of monochrome/multicolor easy/difficult. These cutoffs are arbitrary and determine the curriculum ordering.
  • CoT step count limit = 2 to 6 steps
    Section 4.2: the prompt says 'limit to 2-6' and 'Fewer points are better.' This is a hand-set constraint on the supervision format.
assumptions (4)
  • domain assumption Iconfont icons are a representative, high-quality source of vector graphics
    Section 3 Data Collection states Iconfont 'adheres to unified technical specifications' and is 'an ideal benchmark dataset.' The generality of the method rests on this assumption.
  • domain assumption GPT-4o descriptions are accurate enough to reconstruct the icon from text alone
    Section 3 Data Inference: the goal is that 'the appearance of icons can be accurately reproduced solely based on the descriptions.' No human validation is reported for the bulk of the 1M pairs.
  • domain assumption Automatic metrics (FID, CLIPScore, HPS, aesthetic) capture the desired SVG quality
    Section 5.1 uses these as the primary evaluation. These are image-space metrics applied to rasterized SVGs, and they may not reflect editability or structural correctness.
  • domain assumption Path count is a valid proxy for icon complexity and quality
    Section 4.3 defines the path number matching reward to encourage generated path counts to match or exceed the reference. The paper provides no evidence that path count correlates with semantic accuracy or aesthetic value.

how reviews work

0 comments
Cite this review

Pith. "Pith review of SVGen: Interpretable Vector Graphics Generation with Large Language Models." pith.science (2026). https://pith.science/paper/5HBPQQP3

@misc{pith2026250809168,
  author       = {Pith},
  title        = {Pith review of: SVGen: Interpretable Vector Graphics Generation with Large Language Models},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/5HBPQQP3}},
  note         = {Machine review of arXiv:2508.09168}
}
read the original abstract

Scalable Vector Graphics (SVG) is widely used in front-end development and UI/UX design due to its scalability, editability, and rendering efficiency. However, turning creative ideas into precise vector graphics remains a time-consuming challenge. To address this, we introduce SVG-1M, a large-scale dataset of high-quality SVGs paired with natural language descriptions. Through advanced data augmentation and annotation, we create well-aligned Text to SVG training pairs, including a subset with Chain of Thought annotations for enhanced semantic guidance. Based on this dataset, we propose SVGen, an end-to-end model that generates SVG code from natural language inputs. Our approach ensures semantic accuracy and structural completeness, supported by curriculum learning and reinforcement learning optimization. Experiments show that SVGen outperforms general large models and traditional rendering methods in both effectiveness and efficiency. Code, model, and dataset are available on GitHub.

Figures

Figures reproduced from arXiv: 2508.09168 by the authors.

Figure 1
Figure 1. SVGen analyzes the creative requirements of users, gradually deriving the design logic to generate high-quality SVG [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. The distribution of data complexity and samples [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. 4.1 Curriculum learning Curriculum learning[6] is a training strategy that mimics the hu￾man learning process. The main idea is to begin with simple and easy-to-understand content, gradually transitioning to more com￾plex and challenging tasks[44]. Based on this concept, our research implements the curriculum learning strategy through a phased training process. Initially, we train with structurally simple mono￾chrom… view at source ↗
Figures from the paper (13 more)
Figure 3
Figure 3. Figure 3: The work ow of SVGen. We start with curriculum learning, where the model gradually trains on SVGs increasing in [PITH_FULL_IMAGE:figures/full_fig_p005_3.png]
Figure 4
Figure 4. Figure 4: The structural integrity and path quantity changes [PITH_FULL_IMAGE:figures/full_fig_p007_4.png]
Figure 5
Figure 5. Figure 5: Qualitative comparison between the proposed SVGen and both optimization-based and LLM-based methods [PITH_FULL_IMAGE:figures/full_fig_p008_5.png]
Figure 1
Figure 1. Figure 1: Data samples under di erent di culty categories from SVG-1M [PITH_FULL_IMAGE:figures/full_fig_p011_1.png]
Figure 3
Figure 3. Figure 3: Prompt: GPT-4o image description You are an image analysis assistant. Please help me brie y describe the provided icon. Make sure that everyone can clearly imagine the content and style of this icon through your description. The description text is as short and clear a…
Figure 2
Figure 2. Figure 2: A comparison of the code and visual quality before [PITH_FULL_IMAGE:figures/full_fig_p012_2.png]
Figure 4
Figure 4. Figure 4: Prompts for generating basic descriptions of the [PITH_FULL_IMAGE:figures/full_fig_p012_4.png]
Figure 5
Figure 5. Figure 5: Prompts for constructing detailed CoT steps for [PITH_FULL_IMAGE:figures/full_fig_p012_5.png]
Figure 6
Figure 6. Figure 6: Samples of chain-of-thought data pairs from SVG-1M. [PITH_FULL_IMAGE:figures/full_fig_p013_6.png]
Figure 8
Figure 8. Figure 8: Testing the diversity of generated results. [PITH_FULL_IMAGE:figures/full_fig_p014_8.png]
Figure 7
Figure 7. Figure 7: Human evaluation interface 4 Results of SVGen [PITH_FULL_IMAGE:figures/full_fig_p014_7.png]
Figure 9
Figure 9. Figure 9: Monochrome results generated by SVGen [PITH_FULL_IMAGE:figures/full_fig_p015_9.png]
Figure 10
Figure 10. Figure 10: Multicolored results generated by SVGen [PITH_FULL_IMAGE:figures/full_fig_p015_10.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

74 extracted references · 52 canonical work pages

  1. [1]

    Iconfont

    2025. Iconfont. https://www.iconfont.cn/. [2] Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Floren- cia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al . 2023. Gpt-4 technical report. arXiv preprint arXiv:2303.08774 (2023)

    arXiv 2025

  2. [2]

    Draw two small, curved maroon lines for the eyes and position them symmetrically near the top of the circle

    work page

  3. [3]

    Anthropic. 2025. Claude 3.7 Sonnet. https://www.anthropic.com/claude/sonnet

    work page 2025

  4. [4]

    Shuai Bai, Keqin Chen, Xuejing Liu, Jialin Wang, Wenbin Ge, Sibo Song, Kai Dang, Peng Wang, Shijie Wang, Jun Tang, et al . 2025. Qwen2. 5-vl technical report. arXiv preprint arXiv:2502.13923 (2025)

    arXiv 2025

  5. [5]

    Jonas Belouadi, Anne Lauscher, and Steffen Eger. 2023. Automatikz: Text-guided synthesis of scienti/f_ic vector graphics with tikz.arXiv preprint arXiv:2310.00367 (2023)

    arXiv 2023

  6. [6]

    Yoshua Bengio, Jérôme Louradour, Ronan Collobert, and Jason Weston. 2009. Curriculum learning. In Proceedings of the 26th annual international conference on machine learning . 41–48

    work page 2009

  7. [7]

    Defu Cao, Zhaowen Wang, Jose Echevarria, and Yan Liu. 2023. Svgformer: Representation learning for continuous vector graphics using transformers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 10093–10102

    work page 2023

  8. [8]

    Alexandre Carlier, Martin Danelljan, Alexandre Alahi, and Radu Timofte. 2020. Deepsvg: A hierarchical generative network for vector graphics animation. Ad- vances in Neural Information Processing Systems 33 (2020), 16351–16361

    work page 2020

Show all 74 references
  1. [9]

    Zehao Chen and Rong Pan. 2024. SVGBuilder: Component-Based Colored SVG Generation with Text-Guided Autoregressive Transformers. arXiv preprint arXiv:2412.10488 (2024)

    2024

  2. [10]

    Louis Clouâtre and Marc Demers. 2019. Figr: Few-shot image generation with reptile. arXiv preprint arXiv:1901.02199 (2019)

    2019 arXiv

  3. [11]

    The Vision Cortex. 2023. VTracer. https://www.visioncortex.org/vtracer-docs. [12] Google DeepMind. 2025. Gemini 2.0. https://deepmind.google/technologies/ gemini/

    2023

  4. [13]

    Jon Ferraiolo, Fujisawa Jun, and Dean Jackson. 2000. Scalable vector graphics (SVG) 1.0 speci/f_ication. iuniverse Bloomington

    2000

  5. [14]

    Kevin Frans, Lisa Soros, and Olaf Witkowski. 2022. Clipdraw: Exploring text- to-drawing synthesis through language-image encoders. Advances in Neural Information Processing Systems 35 (2022), 5207–5218

    2022

  6. [15]

    Junyu Gao, Hao Yang, Da Zhang, Yuan Yuan, and Xuelong Li. 2024. Imbalanced aircraft data anomaly detection. IEEE Trans. Aerospace Electron. Systems (2024)

    2024

  7. [16]

    Junyu Gao, Yuan Yuan, and Qi Wang. 2020. Feature-aware adaptation and density alignment for crowd counting in video surveillance. IEEE transactions on cybernetics 51, 10 (2020), 4822–4833

    2020

  8. [17]

    Junyu Gao, Da Zhang, Feiyu Wang, Lichen Ning, Zhiyuan Zhao, and Xuelong Li. 2025. Combining SAM With Limited Data for Change Detection in Remote Sensing. IEEE Transactions on Geoscience and Remote Sensing (2025)

    2025

  9. [18]

    Junyu Gao, Liangliang Zhao, and Xuelong Li. 2024. NWPU-MOC: A benchmark for /f_ine-grained multicategory object counting in aerial images.IEEE Transactions on Geoscience and Remote Sensing 62 (2024), 1–14

    2024

  10. [19]

    Aaron Gratta/f_iori, Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al-Dahle, Aiesha Letman, Akhil Mathur, Alan Schelten, Alex Vaughan, et al. 2024. The llama 3 herd of models. arXiv preprint arXiv:2407.21783 (2024)

    2024 arXiv

  11. [20]

    Daya Guo, Dejian Yang, Haowei Zhang, Junxiao Song, Ruoyu Zhang, Runxin Xu, Qihao Zhu, Shirong Ma, Peiyi Wang, Xiao Bi, et al . 2025. Deepseek-r1: Incentivizing reasoning capability in llms via reinforcement learning. arXiv preprint arXiv:2501.12948 (2025)

    2025 arXiv

  12. [21]

    Junwei Han, Dingwen Zhang, Shifeng Wen, Lei Guo, Tianming Liu, and Xuelong Li. 2015. Two-stage learning to predict human eye /f_ixations via SDAEs. IEEE transactions on cybernetics 46, 2 (2015), 487–498

    2015

  13. [22]

    Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. 2017. Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems 30 (2017)

    2017

  14. [23]

    Ajay Jain, Amber Xie, and Pieter Abbeel. 2023. Vectorfusion: Text-to-svg by ab- stracting pixel-based diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition . 1911–1920

    2023

  15. [24]

    Xiaoheng Jiang, Yanwei Pang, Xuelong Li, Jing Pan, and Yinghong Xie. 2018. Deep neural networks with elastic recti/f_ied linear units for object recognition. Neurocomputing 275 (2018), 1132–1139

    2018

  16. [25]

    Diederik P Kingma, Max Welling, et al . 2013. Auto-encoding variational bayes

    2013

  17. [26]

    Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. 2022. Blip: Bootstrapping language-image pre-training for uni/f_ied vision-language understanding and generation. In International conference on machine learning . PMLR, 12888–12900

    2022

  18. [27]

    Tzu-Mao Li, Michal Lukáč, Michaël Gharbi, and Jonathan Ragan-Kelley. 2020. Differentiable vector graphics rasterization for editing and learning. ACM Trans- actions on Graphics (TOG) 39, 6 (2020), 1–15

    2020

  19. [28]

    Aixin Liu, Bei Feng, Bing Xue, Bingxuan Wang, Bochao Wu, Chengda Lu, Cheng- gang Zhao, Chengqi Deng, Chenyu Zhang, Chong Ruan, et al . 2024. Deepseek-v3 technical report. arXiv preprint arXiv:2412.19437 (2024)

    2024 arXiv

  20. [29]

    Anton Lozhkov, Raymond Li, Loubna Ben Allal, Federico Cassano, Joel Lamy- Poirier, Nouamane Tazi, Ao Tang, Dmytro Pykhtar, Jiawei Liu, Yuxiang Wei, et al . 2024. Starcoder 2 and the stack v2: The next generation. arXiv preprint arXiv:2402.19173 (2024)

    2024 arXiv

  21. [30]

    Chengyuan Peng. 2000. Scalable vector graphics (svg). In Research Seminar on Interactive Digital Media. Citeseer

    2000

  22. [31]

    Antoine Quint. 2003. Scalable vector graphics. IEEE MultiMedia 10, 3 (2003), 99–102

    2003

  23. [32]

    Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. 2021. Learning transferable visual models from natural language supervision. In International conference on machine learni...

    2021

  24. [33]

    Alec Radford, Jeffrey Wu, Rewon Child, David Luan, Dario Amodei, Ilya Sutskever, et al. 2019. Language models are unsupervised multitask learners. OpenAI blog 1, 8 (2019), 9

    2019

  25. [34]

    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 . 7342–7351

    2021

  26. [35]

    Juan A Rodriguez, Shubham Agarwal, Issam H Laradji, Pau Rodriguez, David Vazquez, Christopher Pal, and Marco Pedersoli. 2023. Starvector: Generating scalable vector graphics code from images. arXiv preprint arXiv:2312.11556 (2023)

    2023 arXiv

  27. [36]

    Juan A Rodriguez, David Vazquez, Issam Laradji, Marco Pedersoli, and Pau Rodriguez. 2023. Figgen: Text to scienti/f_ic /f_igure generation.arXiv preprint arXiv:2306.00800 (2023)

    2023 arXiv

  28. [37]

    Juan A Rodriguez, David Vazquez, Issam Laradji, Marco Pedersoli, and Pau Rodriguez. 2023. Ocr-vqgan: Taming text-within-image generation. InProceedings of the IEEE/CVF winter conference on applications of computer vision . 3689–3698

    2023

  29. [38]

    Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. 2022. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition . 10684–10695

    2022

  30. [39]

    Peter Selinger. 2024. Potrace. https://github.com/tatarize/potrace. [40] Zhihong Shao, Peiyi Wang, Qihao Zhu, Runxin Xu, Junxiao Song, Xiao Bi, Haowei Zhang, Mingchuan Zhang, YK Li, Y Wu, et al . 2024. Deepseekmath: Pushing the limits of mathematical reasoning in open language...

    2024 arXiv

  31. [41]

    Vikas Thamizharasan, Difan Liu, Matthew Fisher, Nanxuan Zhao, Evangelos Kalogerakis, and Michal Lukac. 2024. Nivel: Neural implicit vector layers for text-to-vector generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition . 4589–4597

    2024

  32. [42]

    Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothée Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, et al. 2023. Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971 (2023)

    2023 arXiv

  33. [43]

    Yael Vinker, Ehsan Pajouheshgar, Jessica Y Bo, Roman Christian Bachmann, Amit Haim Bermano, Daniel Cohen-Or, Amir Zamir, and Ariel Shamir. 2022. Clipasso: Semantically-aware object sketching. ACM Transactions on Graphics (TOG) 41, 4 (2022), 1–11

    2022

  34. [44]

    Xin Wang, Yudong Chen, and Wenwu Zhu. 2021. A survey on curriculum learning. IEEE transactions on pattern analysis and machine intelligence 44, 9 (2021), 4555–4576

    2021

  35. [45]

    Yizhi Wang and Zhouhui Lian. 2021. Deepvecfont: synthesizing high-quality vector fonts via dual-modality learning. ACM Transactions on Graphics (TOG) 40, 6 (2021), 1–15

    2021

  36. [46]

    Martin Weber. 2024. Autotrace. https://github.com/autotrace/autotrace. [47] Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Fei Xia, Ed Chi, Quoc V Le, Denny Zhou, et al. 2022. Chain-of-thought prompting elicits reasoning in large language models. Advances in neural in...

    2024

  37. [48]

    Chaoyi Wu, Weixiong Lin, Xiaoman Zhang, Ya Zhang, Weidi Xie, and Yanfeng Wang. 2024. PMC-LLaMA: toward building open-source language models for medicine. Journal of the American Medical Informatics Association 31, 9 (2024), 1833–1843

    2024

  38. [49]

    Ronghuan Wu, Wanchao Su, Kede Ma, and Jing Liao. 2023. Iconshop: Text-guided vector icon synthesis with autoregressive transformers. ACM Transactions on Graphics (TOG) 42, 6 (2023), 1–14

    2023

  39. [50]

    Shijie Wu, Ozan Irsoy, Steven Lu, Vadim Dabravolski, Mark Dredze, Sebas- tian Gehrmann, Prabhanjan Kambadur, David Rosenberg, and Gideon Mann

  40. [51]

    Xiaoshi Wu, Keqiang Sun, Feng Zhu, Rui Zhao, and Hongsheng Li. 2023. Human preference score: Better aligning text-to-image models with human preference. In Proceedings of the IEEE/CVF International Conference on Computer Vision . 2096– 2105

    2023

  41. [52]

    xAI. 2025. Grok 3 Beta — The Age of Reasoning Agents. https://x.ai/news/grok-3

    2025

  42. [53]

    Ximing Xing, Juncheng Hu, Guotao Liang, Jing Zhang, Dong Xu, and Qian Yu

  43. [54]

    Ximing Xing, Chuang Wang, Haitao Zhou, Jing Zhang, Qian Yu, and Dong Xu

  44. [55]

    Advances in Neural Information Processing Systems 36 (2023), 15869– 15889

    Diffsketcher: Text guided vector sketch synthesis through latent diffusion models. Advances in Neural Information Processing Systems 36 (2023), 15869– 15889

    2023

  45. [56]

    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. 4546–4555

    2024

  46. [57]

    An Yang, Baosong Yang, Beichen Zhang, Binyuan Hui, Bo Zheng, Bowen Yu, Chengyuan Li, Dayiheng Liu, Fei Huang, Haoran Wei, et al . 2024. Qwen2. 5 technical report. arXiv preprint arXiv:2412.15115 (2024)

    2024 arXiv

  47. [58]

    Shengbin Yue, Wei Chen, Siyuan Wang, Bingxuan Li, Chenchen Shen, Shujun Liu, Yuxuan Zhou, Yao Xiao, Song Yun, Xuanjing Huang, et al . 2023. Disc-lawllm: Fine-tuning large language models for intelligent legal services. arXiv preprint arXiv:2309.11325 (2023)

    2023 arXiv

  48. [59]

    Peiying Zhang, Nanxuan Zhao, and Jing Liao. 2024. Text-to-vector generation with neural path representation. ACM Transactions on Graphics (TOG) 43, 4 (2024), 1–13

    2024

  49. [60]

    Tong Zhang, Haoyang Liu, Peiyan Zhang, Yuxuan Cheng, and Haohan Wang

  50. [61]

    http://www.w3.org/2000/svg

    Beyond pixels: Exploring human-readable svg generation for simple images with vision language models. arXiv preprint arXiv:2311.15543 (2023). SVGen: Interpretable Vector Graphics Generation with Large Language Models Monocolor_easy Monocolor_difficult Muticolor_difficult Motic...

    2023 arXiv

  51. [62]

    Create a large yellow circle for the face and add a slightly darker yellow crescent on the left side for shading

  52. [63]

    Add two blue teardrop shapes, one on each side of the face, angled downwards to represent tears of joy

  53. [64]

    Design a wide, open mouth using a maroon semicircle and place it in the lower center of the face

  54. [65]

    Prompt: Please generate an SVG icon that meets the following description: Red bird on a brown branch with a yellow body and a green eye

    Insert a smaller, red semicircle inside the mouth to depict the tongue, and add a white rectangle above it to represent teeth. Prompt: Please generate an SVG icon that meets the following description: Red bird on a brown branch with a yellow body and a green eye. CoT:

  55. [66]

    Create a red circle for the bird’s head and a smaller red circle for the bird’s eye, /f_illing the eye with a brown color

  56. [67]

    Draw a red oval for the bird’s body, ensuring it overlaps with the head

  57. [68]

    Add a yellow circle behind the bird’s body to represent the sun

  58. [69]

    Construct two yellow rectangles for the bird’s legs and place them below the body

  59. [70]

    Prompt: Please generate an SVG icon that meets the following description: A pink sunburst with a yellow circle, featuring a brown-faced smiley wearing black sunglasses

    Design a brown horizontal line for the branch and add two small brown rectangles at the ends to represent the branch’s supports. Prompt: Please generate an SVG icon that meets the following description: A pink sunburst with a yellow circle, featuring a brown-faced smiley weari...

  60. [71]

    Create a large circle with a pink /f_ill to form the outer gear-like shape, adding wavy edges to mimic the gear’s teeth

  61. [72]

    Draw a smaller circle with a yellow /f_ill and place it centrally within the pink gear shape to represent the sun

  62. [73]

    Add two rounded rectangles with a brown /f_ill for the sunglasses, positioning them symmetrically on the yellow circle

  63. [74]

    Connect the sunglasses with a small horizontal rectangle in brown to form the bridge of the glasses

  64. [75]

    Figure 6: Samples of chain-of-thought data pairs from SVG-1M

    Draw a small curved line with a brown /f_ill below the sunglasses to create a smiling mouth. Figure 6: Samples of chain-of-thought data pairs from SVG-1M. 3 Human Evaluation This study designed a systematic human evaluation scheme, im- plementing a professional online scoring ...

    2025

  65. [2023]

    arXiv preprint MM ’25, October 27–31, 2025, Dublin, Ireland Wang et al

    Bloomberggpt: A large language model for /f_inance. arXiv preprint MM ’25, October 27–31, 2025, Dublin, Ireland Wang et al. arXiv:2303.17564 (2023)

    2025 arXiv

  66. [2024]

    arXiv preprint arXiv:2412.11102 (2024)

    Empowering LLMs to Understand and Generate Complex Vector Graphics. arXiv preprint arXiv:2412.11102 (2024)

    2024 arXiv

Pith tools

Reviewed August 5, 2026 · model on record in the stance chip above.