Pith. sign in

REVIEW 3 major objections 7 minor 68 references

VecFontLLM claims to be the first system to synthesize complex Chinese vector glyphs directly in the vector domain, without raster generation or vectorization, using an anchor-guided generation process.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · deepseek-v4-flash

2026-08-01 18:34 UTC pith:LB45KYLM

load-bearing objection Real anchor-to-curve decomposition with honest failure cases, but the 'first time/SOTA' claim needs missing baselines, error bars, and a softer mechanistic story. the 3 major comments →

arxiv 2607.17251 v1 pith:LB45KYLM submitted 2026-07-19 cs.CV

VecFontLLM: Anchor-Guided Direct Synthesis of Chinese Vector Fonts

Hao Yuan , Yuxuan Luo , Xing Chen , Zhouhui Lian This is my paper
classification cs.CV
keywords Chinese vector fontsfew-shot font synthesisanchor-guided generationmultimodal large language modelSVG generationBézier curve completiontest-time scalingstyle transfer
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

VecFontLLM claims to be the first system to synthesize complex Chinese vector glyphs directly in the vector domain, without raster generation or vectorization. The paper's central claim is that previous direct vector generators fail because they write path commands, anchors, and Bézier control points in one long token sequence, tightly coupling layout and curvature. The proposed fix is to generate glyphs through anchors: the model first builds and refines a scaffold of path-segment endpoints that fixes component layout, then infills Bézier control points for local style. The paper argues this decoupling, plus a confidence-guided generation chain that selects reliable components at test time, yields editable vector fonts competitive with raster-domain generators. A sympathetic reader would care because editable vector output is what typographers actually need, and few-shot style transfer directly in the vector domain removes the raster-to-vector postprocessing step.

Core claim

The paper claims that for the first time, high-quality few-shot synthesis of complex Chinese vector glyphs can be performed directly in the vector domain, without raster generation or vectorization. The evidence is a three-stage pipeline in which a multimodal large language model first predicts an anchor scaffold—the endpoints of every path segment—so that the coarse layout of components and contours is fixed before any Bézier control points are generated, then refines that scaffold, and finally completes Bézier control points to recover local curvature and style. On several Chinese font datasets, the paper shows that this anchor-to-curve staging outperforms previous vector-domain generators

What carries the argument

The central object is the anchor: the endpoint of each M, L, and C path segment. By generating only anchors first, the model fixes the coarse layout of components and contours, then fills in Bézier control points to recover curvature and style. This decomposition is carried by a vector style encoder that embeds SVG instructions and aggregates them with a 1D-CNN and Transformer, trained with a supervised contrastive objective, alongside a multimodal LLM backbone that receives content image features and style features; the same architecture is fine-tuned separately for the three stages. At test time, a confidence-guided generation chain samples multiple candidate components, scores them by tok

Load-bearing premise

The load-bearing premise is that a glyph's coarse layout is fully captured by the endpoint positions of its path segments (anchors), with Bézier control points contributing only local curvature and style; if an ambiguous endpoint polygon cannot determine the true component arrangement, Stages 1–2 cannot fix layout errors and the whole pipeline inherits them.

What would settle it

Construct a set of Chinese glyphs whose path-segment endpoints are identical, or nearly identical, across two different valid component layouts, and test whether Stage-1/2 resolves them correctly; if the model consistently fails on such ambiguous anchor scaffolds, the anchor decomposition is insufficient. A simpler check: measure how often Stage-2 refinement rewrites the topology of the Stage-1 scaffold (e.g., changes component boundaries or contour nesting); frequent topology changes would indicate that anchors carry too little layout information.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • If the central claim holds, font designers get editable, resolution-independent vector glyphs from one or a few style exemplars, with no raster-to-vector conversion step.
  • The anchor-to-curve staging suggests that long structured vector sequences are better generated by first fixing global structure and then local geometry, a principle that could extend beyond fonts to icons, diagrams, and other SVG graphic classes.
  • Style interpolation of vector glyphs becomes possible through linearly interpolated style features, opening a path to continuous font morphing and design-space exploration.
  • Lightweight fine-tuning with roughly 75 glyphs per font adapts the model to a new style, implying that the method can be adopted for font libraries at modest annotation cost.
  • Confidence-guided test-time scaling improves structural reliability for complex characters, making autoregressive vector generation more robust and suggesting a general strategy for long-sequence structured output.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • Editorial inference: The anchor-sufficiency assumption—that a glyph's coarse layout is determined by path-segment endpoints alone—is not formally proven in the paper; a stress test on glyphs with long curved strokes whose endpoint polygons are ambiguous would reveal whether the decomposition has a hard ceiling.
  • Editorial inference: The paper's finding that generated command distributions are close to real fonts suggests a distribution-level fidelity measure for vector fonts that goes beyond rasterized-image metrics like FID or LPIPS; this could become a standard evaluation axis for vector-domain generative models.
  • Editorial inference: The three-stage decomposition (scaffold, refine, complete curves) is a natural fit for other structured vector formats with global layout and local style, such as circuit schematics or architectural drawings, where the same decoupling could improve autoregressive generation.
  • Editorial inference: The failure cases shown in the appendix—collapse into meaningless token sequences for dense glyphs with many short strokes—point to a ceiling of discrete-token autoregressive generation; a latent-continuous anchor model, which the paper itself suggests as future work, may be needed to push further.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 7 minor

Summary. The paper proposes VecFontLLM, an MLLM-based system for few-shot Chinese vector font synthesis that generates vector glyphs directly without intermediate rasterization. The key idea is to decompose generation into three stages: (Stage-1) construction of an initial 'anchor scaffold' consisting of path-segment endpoints, (Stage-2) refinement of that scaffold, and (Stage-3) completion of Bézier control points conditioned on the refined scaffold. The model uses a vector style encoder with contrastive training and a confidence-guided test-time generation chain that selects high-confidence component-level candidates. Experiments compare against older vector-domain baselines (DeepVecFont, DeepVecFont-V2, DualVector) and image-domain methods (MX-Font, CG-GAN, NTF, FontDiffuser), reporting large gains in FID/SSIM and competitive LPIPS/L1/RMSE. The paper claims, 'for the first time, high-quality few-shot synthesis of complex Chinese vector glyphs directly in the vector domain.'

Significance. If the results hold, this is a meaningful step toward editable vector-output font generation. The anchor-guided decomposition is a clean and intuitive idea, and the reported quantitative gains over older vector baselines are substantial. The confidence-guided component-level generation chain is also an interesting adaptation of test-time scaling to structured vector output. The paper is honest about some failure cases in its appendices. However, the 'first-time / state-of-the-art' claim is currently under-supported because recent vector-domain methods cited in the paper are not evaluated, and no repeated-run statistics are reported. The theoretical claim about anchor sufficiency is asserted rather than demonstrated. Despite these issues, the core empirical contribution is defensible and addressable in revision.

major comments (3)
  1. [§4.2, Table 1; §2 Related Work] The claim of 'state-of-the-art' and 'for the first time' is not fully supported by the chosen comparison set. VecFusion (Thamizharasan et al. 2024a), DiffVecFont (Liu et al. 2025), and VecGlypher (Huang et al. 2026) are cited in §2 but never evaluated as baselines. The only vector-domain competitors are DeepVecFont, DeepVecFont-V2, and DualVector, all of which predate these works. Without results for recent direct-vector synthesis methods, the superiority claim is incomplete. In addition, all tables report single numbers without error bars or repeated-seed statistics; given the sampling sensitivity shown in Table 6, the reported gains cannot be assessed for statistical reliability.
  2. [§3.1 and Appendix B.1] The central decomposition assumes that anchor (segment-endpoint) placement fixes the coarse layout, while Bézier control points only add local curvature/style. This is not guaranteed for cubic Bézier segments: with endpoints and path order fixed, control points can move a curve to either side of another component, create loops, or change containment. Stage-3 is conditioned on the refined scaffold and cannot revise component grouping, contour order, or nesting already fixed in Stage-1/2. The 'anchor-to-curve' ablation in Table 5 compares the full multi-stage pipeline with direct generation, which also differs in stage-specific data and conditioning; it does not isolate whether the anchor scaffold itself carries the layout. As a concrete test, I suggest an oracle experiment: condition Stage-3 on ground-truth anchors (and on deliberately corrupted anchors) to measure whether Stage-3 can cor
  3. [§4.1.1 (HBTS) and Table 3] The Hard Base Test Set is defined as '6,582 failure cases produced by the Stage-1 model on BTS.' Evaluating the confidence-guided test-time scaling on this self-selected hard set is circular to a degree: the model's own failure distribution determines the evaluation set. The improvements in Table 3 (e.g., FID 20.6 -> 17.6 for the generation chain) may not generalize to the full BTS or to an independent hard set defined by character complexity. The claim that component-level selection 'filters uncertain structures earlier' should be validated on the full BTS or on a hard set selected by criteria independent of the model. As reported, the comparison measures relative improvement on the model's own failure subset.
minor comments (7)
  1. [Abstract and §5] The abstract claims 'high-quality' synthesis and 'for the first time' without the caveats that appear in Appendix G and §5, where the authors acknowledge generation failures on highly complex glyphs and unstable interpolation between distant styles. I recommend softening the wording to reflect those limitations.
  2. [Appendix D.1, Eq. (11)] The 'token confidence' measure C_TC_i = -1/V * sum_j log p_i(j) is unusual: it is the negative average log-probability over the whole vocabulary, which is essentially a transformed perplexity rather than a token-level confidence. It also diverges when any p_i(j)=0. Please clarify why this is a confidence measure, define its behavior for zero probabilities, or compute it on the top-k distribution as hinted in the text.
  3. [Table 3 and Fig. 19] The aggregation fractions (t%, b%) for the top/bottom rows are not specified in Table 3; Fig. 19 mentions top-45% only for the generation chain. Please report the exact fractions used for all rows so the comparison is reproducible.
  4. [Fig. 7] The claim that command distributions are 'close to real fonts' is only qualitative. A numerical distance (e.g., KL divergence or total variation) between the generated and ground-truth distributions would make the claim testable.
  5. [§3.2, Eq. (1)] The notation 'W_A ∈ R^{d_E×8d_E}' should be 'R^{8d_E×d_E}' or the text should explicitly say the concatenated vector has dimension 8d_E before projection; the current notation is ambiguous.
  6. [§4.4, Table 4] The OCR accuracy of the proposed method is lower than the image-domain baselines in most settings. The discussion should acknowledge this directly, not only state that vector outputs remain recognizable, so readers can calibrate the 'recognition quality' claim.
  7. [Table 8] Typo: 'steming' should be 'stemming'; 'K-V cache' should be 'KV cache.'

Circularity Check

0 steps flagged

No circular derivation found; anchor-guided decomposition is a testable modeling choice, not an input–output identity.

full rationale

The paper's central claim is empirical: an anchor-guided MLLM directly synthesizes Chinese vector glyphs. The three-stage pipeline is trained with autoregressive cross-entropy on ground-truth vector sequences (Eqs. 3–4) and evaluated on held-out sets (BTS, UFTS, UCTS, STS) against external raster and vector baselines using rendering metrics, OCR, and human studies. The anchor-scaffold decomposition is a modeling assumption about the roles of segment endpoints versus Bézier control points, but it is not derived from, nor does it reduce to, the outputs it predicts: Stage-1 anchors are ground-truth endpoints, Stage-2 refines them, and Stage-3 predicts control points conditioned on the refined scaffold; each stage's target is independent of the conditioning by construction, and the ablation in Table 5 directly compares the full pipeline against direct generation. Self-citations (DeepVecFont-V2, HFH-Font, Cai et al., CalliReader) appear in related work, representation choices, and baselines, but the central synthesis claim does not rest on any unverified self-citation chain or imported uniqueness theorem. The HBTS test set is self-referential—it is built from Stage-1 failures (Section 4.1.1)—and the confidence-guided scaling study is therefore potentially optimistic, but the improvement is measured against ground-truth rendering metrics, and the confidence scores are heuristics rather than fitted predictors of those metrics, so this is a selection-bias limitation, not circularity by construction. Appendix G explicitly acknowledges residual failures and inaccuracies, further showing the method is not presented as a forced consequence of its assumptions. No step reduces to its own inputs by definition or fit; the paper contains at most minor, non-load-bearing self-citations.

Axiom & Free-Parameter Ledger

7 free parameters · 5 axioms · 0 invented entities

The central method introduces no new physical entities. The free parameters are sampling and confidence-selection settings plus learned network weights. The axioms are modeling assumptions about the vector-font domain that, if false, would weaken the anchor-guided decomposition and the component-level confidence chain.

free parameters (7)
  • Sampling temperature = 0.8
    Chosen via Table 6 parameter analysis on BTS; not derived.
  • Top-k sampling = 50
    Chosen via Table 6 on BTS.
  • Beam size = 4
    Chosen via Table 6 on BTS.
  • Confidence aggregation top/bottom fractions (t%, b%) = top-45% used for best generation chain; b% not specified
    Appendix E.4 states top-45% aggregation gives the best visual results; these fractions are tuned on HBTS.
  • Number of component candidates in generation chain = 10
    §4.3 evaluates the chain with 10 sampled candidates per component.
  • Number of glyph candidates in parallel thinking = 20
    §4.3 evaluates parallel thinking with 20 complete scaffold candidates.
  • Learned neural-network weights = trained on 80% of Basis Fonts dataset
    Includes fine-tuned StarCoder-1B, adapters, vector style encoder, and image encoder; values are trained, not enumerated, but are ultimately free parameters of the empirical method.
axioms (5)
  • domain assumption Path-segment endpoints (anchors) fully capture coarse glyph layout; Bézier control points only carry local curvature and style.
    Introduced in §3.1 and Appendix B.1; load-bearing for the three-stage decomposition.
  • domain assumption Contour orientation (clockwise outer, counterclockwise holes) yields a valid <component> decomposition of Chinese glyphs.
    Appendix D.2; the component-level confidence generation chain depends on this grouping being correct.
  • domain assumption Token-level confidence correlates with structural correctness of generated SVG components.
    §3.3; central to test-time scaling; validated only empirically on HBTS, not by a formal relation.
  • domain assumption The 6,763-character set and 345 training fonts are sufficient to represent all components of unseen characters for compositional generalization.
    §4.1.1; the UCTS generalization results depend on this coverage assumption.
  • domain assumption A contrastive vector style encoder trained with same-font positive pairs learns font-level style independent of character content.
    §3.2.1; the style-conditioning mechanism relies on this for few-shot style transfer.

pith-pipeline@v1.3.0-alltime-deepseek · 22401 in / 12804 out tokens · 122482 ms · 2026-08-01T18:34:13.936347+00:00 · methodology

0 comments
read the original abstract

Direct generation of Chinese vector fonts is a challenging and ongoing problem. A Chinese vector glyph contains complex component structure, anchor layout, and B\'ezier curve details, which work at different scales, but a standard vector sequence writes them together in one long sequence, making the task of vector font synthesis challenging. Existing direct vector generators often fail on complex characters, while raster-domain methods must vectorize the synthesized glyph images afterward. To address the above-mentioned problem, this paper proposes VecFontLLM, an anchor-guided multimodal large language model for direct few-shot synthesis of Chinese vector fonts. Our key idea is to generate vector glyphs through anchors rather than a standard vector sequence. Specifically, the proposed VecFontLLM first predicts and refines an anchor scaffold that fixes the coarse layout of components and contours, and then completes B\'ezier control points to recover local curvature and style. At test time, a confidence-guided generation chain samples multiple component candidates and continues synthesis from the highest-confidence one, improving stability for complex glyphs. This work demonstrates, for the first time, high-quality few-shot synthesis of complex Chinese vector glyphs directly in the vector domain, without raster generation or vectorization. Experiments on several Chinese font datasets show substantial improvements over existing vector font synthesis methods, competitive glyph rendering quality against raster-domain baselines, and vector command distributions close to real fonts.

Figures

Figures reproduced from arXiv: 2607.17251 by Hao Yuan, Xing Chen, Yuxuan Luo, Zhouhui Lian.

Figure 1
Figure 1. Figure 1: VecFontLLM is an MLLM-based framework optimized for high-fidelity Chinese vector font synthesis (a). Compared with existing vector font synthesis [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Overview of VecFontLLM. (a) We factorize Chinese vector font synthesis into anchor scaffold construction, anchor scaffold refinement, and Bézier [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 4
Figure 4. Figure 4: Unseen-font and cross-language visualization results. (a) English glyph [PITH_FULL_IMAGE:figures/full_fig_p011_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Qualitative evaluation of image-modality methods. We ensure high generation quality while preserving scalability. For UFTS, our model adapts to [PITH_FULL_IMAGE:figures/full_fig_p011_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: A fine-grained comparison with ground-truth glyphs. Red boxes highlight locally consistent strokes between the synthesized and ground truth. [PITH_FULL_IMAGE:figures/full_fig_p011_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Token count distributions of <M>, <L>, <C>, and <component> ((a)–(d)) across vector glyphs. The generated fonts closely follow human line and curve design patterns. Font a Linear interpolation Font b [PITH_FULL_IMAGE:figures/full_fig_p012_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Interpolation results between two fonts of distinct styles. Our generative method facilitates smooth transitions between two distinct font styles, [PITH_FULL_IMAGE:figures/full_fig_p012_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: Illustration of vector glyph instructions. We normalize path data into four command types, [PITH_FULL_IMAGE:figures/full_fig_p014_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: Challenges of vector generation. (a) Vector glyphs mix command tokens and coordinate arguments, while raster glyphs use fixed-grid pixels. (b) [PITH_FULL_IMAGE:figures/full_fig_p014_10.png] view at source ↗
Figure 11
Figure 11. Figure 11: Complexity statistics of vector data. We compare Chinese glyphs, English glyphs from the 345 training fonts, and general vector graphics from [PITH_FULL_IMAGE:figures/full_fig_p015_11.png] view at source ↗
Figure 12
Figure 12. Figure 12: Definition of <component>. (a) A <component> contains one outer closed contour and its enclosed interior contours. (b) A vector glyph is composed of one or more <component>s. VecFontLLM (a) Parallel thinking. (b) Generation chain. Select VecFontLLM Select [PITH_FULL_IMAGE:figures/full_fig_p016_12.png] view at source ↗
Figure 13
Figure 13. Figure 13: Comparison of parallel thinking and the generation chain. (a) Par￾allel thinking samples multiple complete glyph candidates and selects the one with the highest confidence. (b) The generation chain samples multiple candidates for each <component>, appends the highest-confidence candi￾date to the partial scaffold, and repeats the process until the full glyph is generated. E.2 Comparison with Other Image-do… view at source ↗
Figure 14
Figure 14. Figure 14: Additional visual comparison with vector-domain font synthesis methods on STS. [PITH_FULL_IMAGE:figures/full_fig_p017_14.png] view at source ↗
Figure 15
Figure 15. Figure 15: Additional visual comparison with image-domain font synthesis methods on BTS. [PITH_FULL_IMAGE:figures/full_fig_p018_15.png] view at source ↗
Figure 16
Figure 16. Figure 16: Additional visual comparison with image-domain font synthesis methods on [PITH_FULL_IMAGE:figures/full_fig_p019_16.png] view at source ↗
Figure 17
Figure 17. Figure 17: Additional visual comparison with image-domain font synthesis methods on [PITH_FULL_IMAGE:figures/full_fig_p020_17.png] view at source ↗
Figure 18
Figure 18. Figure 18: Additional comparisons between our generated vector glyphs and the ground truth. The results show smooth local curves and consistent outline [PITH_FULL_IMAGE:figures/full_fig_p021_18.png] view at source ↗
Figure 19
Figure 19. Figure 19: Additional comparison of test-time scaling strategies on HBTS. The generation chain with token confidence and top-45% aggregation gives the best [PITH_FULL_IMAGE:figures/full_fig_p022_19.png] view at source ↗
Figure 20
Figure 20. Figure 20: Qualitative ablation of Stage-2 anchor refinement. Stage-2 corrects missing parts and structural errors in the initial anchor scaffold. [PITH_FULL_IMAGE:figures/full_fig_p023_20.png] view at source ↗
Figure 21
Figure 21. Figure 21: Qualitative ablation of the anchor-to-curve pipeline. The multi-stage pipeline reduces missing parts, distorted structures, and generation failures. [PITH_FULL_IMAGE:figures/full_fig_p024_21.png] view at source ↗
Figure 22
Figure 22. Figure 22: Qualitative comparison under different sampling parameters. The red box marks the result closest to the ground truth. [PITH_FULL_IMAGE:figures/full_fig_p025_22.png] view at source ↗
Figure 23
Figure 23. Figure 23: Interpolation between vector fonts. The red boxes highlight the gradual change of local geometric features. [PITH_FULL_IMAGE:figures/full_fig_p026_23.png] view at source ↗
Figure 25
Figure 25. Figure 25: Qualitative results of vector glyph completion. Given a partial vector [PITH_FULL_IMAGE:figures/full_fig_p027_25.png] view at source ↗
Figure 26
Figure 26. Figure 26: Examples of generation failure. We generate only a partial charac [PITH_FULL_IMAGE:figures/full_fig_p028_26.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

68 extracted references · 15 linked inside Pith

  1. [1]

    Communications of the ACM , volume=

    Generative adversarial networks , author=. Communications of the ACM , volume=. 2020 , publisher=

    2020

  2. [2]

    Advances in neural information processing systems , volume=

    Denoising diffusion probabilistic models , author=. Advances in neural information processing systems , volume=

  3. [3]

    Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

    High-resolution image synthesis with latent diffusion models , author=. Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

  4. [4]

    Advances in Neural Information Processing Systems , volume=

    Deepsvg: A hierarchical generative network for vector graphics animation , author=. Advances in Neural Information Processing Systems , volume=

  5. [5]

    arXiv preprint arXiv:1704.03477 , year=

    A neural representation of sketch drawings , author=. arXiv preprint arXiv:1704.03477 , year=

    Pith/arXiv arXiv

  6. [6]

    Proceedings of the IEEE/CVF International Conference on Computer Vision , pages=

    A learned representation for scalable vector graphics , author=. Proceedings of the IEEE/CVF International Conference on Computer Vision , pages=

  7. [7]

    ACM Transactions on Graphics (TOG) , volume=

    Deepvecfont: synthesizing high-quality vector fonts via dual-modality learning , author=. ACM Transactions on Graphics (TOG) , volume=. 2021 , publisher=

    2021

  8. [8]

    Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

    Deepvecfont-v2: Exploiting transformers to synthesize vector fonts with higher quality , author=. Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

  9. [9]

    IEEE MultiMedia , volume=

    Scalable vector graphics , author=. IEEE MultiMedia , volume=. 2003 , publisher=

    2003

  10. [10]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

    Vecfusion: Vector font generation with diffusion , author=. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

  11. [11]

    International Conference on Computational Visual Media , pages=

    DiffVecFont: Fusing Dual-Mode Reconstruction Vector Fonts via Masked Diffusion Transformers , author=. International Conference on Computational Visual Media , pages=. 2025 , organization=

    2025

  12. [12]

    Proceedings of the Computer Vision and Pattern Recognition Conference , pages=

    Starvector: Generating scalable vector graphics code from images and text , author=. Proceedings of the Computer Vision and Pattern Recognition Conference , pages=

  13. [13]

    Proceedings of the Computer Vision and Pattern Recognition Conference , pages=

    Chat2SVG: Vector Graphics Generation with Large Language Models and Image Diffusion Models , author=. Proceedings of the Computer Vision and Pattern Recognition Conference , pages=

  14. [14]

    ACM Transactions on Graphics (TOG) , volume=

    Iconshop: Text-guided vector icon synthesis with autoregressive transformers , author=. ACM Transactions on Graphics (TOG) , volume=. 2023 , publisher=

    2023

  15. [15]

    Proceedings of the Computer Vision and Pattern Recognition Conference , pages=

    Empowering llms to understand and generate complex vector graphics , author=. Proceedings of the Computer Vision and Pattern Recognition Conference , pages=

  16. [16]

    arXiv preprint arXiv:2504.06263 , year=

    Omnisvg: A unified scalable vector graphics generation model , author=. arXiv preprint arXiv:2504.06263 , year=

    arXiv

  17. [17]

    arXiv preprint arXiv:2203.11171 , year=

    Self-consistency improves chain of thought reasoning in language models , author=. arXiv preprint arXiv:2203.11171 , year=

    Pith/arXiv arXiv

  18. [18]

    arXiv preprint arXiv:2305.11860 , year=

    Let's Sample Step by Step: Adaptive-Consistency for Efficient Reasoning and Coding with LLMs , author=. arXiv preprint arXiv:2305.11860 , year=

    Pith/arXiv arXiv

  19. [19]

    arXiv preprint arXiv:2402.13212 , year=

    Soft self-consistency improves language model agents , author=. arXiv preprint arXiv:2402.13212 , year=

    Pith/arXiv arXiv

  20. [20]

    arXiv preprint arXiv:2311.08298 , year=

    A survey of confidence estimation and calibration in large language models , author=. arXiv preprint arXiv:2311.08298 , year=

    Pith/arXiv arXiv

  21. [21]

    arXiv preprint arXiv:2403.04696 , year=

    Fact-checking the output of large language models via token-level uncertainty quantification , author=. arXiv preprint arXiv:2403.04696 , year=

    Pith/arXiv arXiv

  22. [22]

    arXiv preprint arXiv:2502.18581 , year=

    Scalable best-of-n selection for large language models via self-certainty , author=. arXiv preprint arXiv:2502.18581 , year=

    arXiv

  23. [23]

    arXiv preprint arXiv:2508.15260 , year=

    Deep think with confidence , author=. arXiv preprint arXiv:2508.15260 , year=

    Pith/arXiv arXiv

  24. [24]

    arXiv preprint arXiv:1312.6114 , year=

    Auto-encoding variational bayes , author=. arXiv preprint arXiv:1312.6114 , year=

    Pith/arXiv arXiv

  25. [25]

    Proceedings of the AAAI Conference on Artificial Intelligence , volume=

    Cg-gan: An interactive evolutionary gan-based approach for facial composite generation , author=. Proceedings of the AAAI Conference on Artificial Intelligence , volume=

  26. [26]

    Proceedings of the IEEE/CVF international conference on computer vision , pages=

    Multiple heads are better than one: Few-shot font generation with multiple localized experts , author=. Proceedings of the IEEE/CVF international conference on computer vision , pages=

  27. [27]

    Proceedings of the AAAI conference on artificial intelligence , volume=

    Fontdiffuser: One-shot font generation via denoising diffusion with multi-scale content aggregation and style contrastive learning , author=. Proceedings of the AAAI conference on artificial intelligence , volume=

  28. [28]

    Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

    Neural transformation fields for arbitrary-styled font generation , author=. Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

  29. [29]

    Proceedings of the 29th ACM international conference on multimedia , pages=

    Zigan: Fine-grained chinese calligraphy font generation via a few-shot style transfer approach , author=. Proceedings of the 29th ACM international conference on multimedia , pages=

  30. [30]

    European conference on computer vision , pages=

    Few-shot compositional font generation with dual memory , author=. European conference on computer vision , pages=. 2020 , organization=

    2020

  31. [31]

    Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

    Dg-font: Deformable generative networks for unsupervised font generation , author=. Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

  32. [32]

    Proceedings of the AAAI conference on artificial intelligence , volume=

    Strokegan: Reducing mode collapse in chinese font generation via stroke encoding , author=. Proceedings of the AAAI conference on artificial intelligence , volume=

  33. [33]

    International Journal of Computer Vision , volume=

    Diff-font: Diffusion model for robust one-shot font generation , author=. International Journal of Computer Vision , volume=. 2024 , publisher=

    2024

  34. [34]

    ACM Transactions on Graphics (TOG) , volume=

    HFH-font: Few-shot Chinese font synthesis with higher quality, faster speed, and higher resolution , author=. ACM Transactions on Graphics (TOG) , volume=. 2024 , publisher=

    2024

  35. [35]

    Neural computation , volume=

    Long short-term memory , author=. Neural computation , volume=. 1997 , publisher=

    1997

  36. [36]

    Advances in neural information processing systems , volume=

    Attention is all you need , author=. Advances in neural information processing systems , volume=

  37. [37]

    Proceedings of the IEEE , volume=

    Gradient-based learning applied to document recognition , author=. Proceedings of the IEEE , volume=. 2002 , publisher=

    2002

  38. [38]

    arXiv preprint arXiv:2305.06161 , year=

    Starcoder: may the source be with you! , author=. arXiv preprint arXiv:2305.06161 , year=

    Pith/arXiv arXiv

  39. [39]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

    Vectorfusion: Text-to-svg by abstracting pixel-based diffusion models , author=. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

  40. [40]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

    Dualvector: Unsupervised vector font synthesis with dual-part representation , author=. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

  41. [41]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

    Svgdreamer: Text guided svg generation with diffusion model , author=. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

  42. [42]

    ACM Transactions on Graphics (TOG) , volume=

    Text-to-vector generation with neural path representation , author=. ACM Transactions on Graphics (TOG) , volume=. 2024 , publisher=

    2024

  43. [43]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

    Nivel: Neural implicit vector layers for text-to-vector generation , author=. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

  44. [44]

    arXiv preprint arXiv:2209.14988 , year=

    Dreamfusion: Text-to-3d using 2d diffusion , author=. arXiv preprint arXiv:2209.14988 , year=

    Pith/arXiv arXiv

  45. [45]

    arXiv preprint arXiv:2501.12948 , year=

    Deepseek-r1: Incentivizing reasoning capability in llms via reinforcement learning , author=. arXiv preprint arXiv:2501.12948 , year=

    Pith/arXiv arXiv

  46. [46]

    arXiv preprint arXiv:2407.10972 , year=

    Vgbench: Evaluating large language models on vector graphics understanding and generation , author=. arXiv preprint arXiv:2407.10972 , year=

    Pith/arXiv arXiv

  47. [47]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

    SVGEditBench: A Benchmark Dataset for Quantitative Assessment of LLM's SVG Editing Capabilities , author=. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

  48. [48]

    arXiv preprint arXiv:2505.24499 , year=

    Reason-SVG: Hybrid Reward RL for Aha-Moments in Vector Graphics Generation , author=. arXiv preprint arXiv:2505.24499 , year=

    Pith/arXiv arXiv

  49. [49]

    Proceedings of the 33rd ACM International Conference on Multimedia , pages=

    SVGen: Interpretable Vector Graphics Generation with Large Language Models , author=. Proceedings of the 33rd ACM International Conference on Multimedia , pages=

  50. [50]

    Proceedings of the 33rd ACM International Conference on Multimedia , pages=

    SVGThinker: Instruction-Aligned and Reasoning-Driven Text-to-SVG Generation , author=. Proceedings of the 33rd ACM International Conference on Multimedia , pages=

  51. [51]

    arXiv preprint arXiv:2510.22684 , year=

    RoboSVG: A Unified Framework for Interactive SVG Generation with Multi-modal Guidance , author=. arXiv preprint arXiv:2510.22684 , year=

    arXiv

  52. [52]

    arXiv preprint arXiv:2510.11341 , year=

    InternSVG: Towards Unified SVG Tasks with Multimodal Large Language Models , author=. arXiv preprint arXiv:2510.11341 , year=

    arXiv

  53. [53]

    arXiv preprint arXiv:2503.06472 , year=

    CalliReader: Contextualizing Chinese Calligraphy via an Embedding-Aligned Vision-Language Model , author=. arXiv preprint arXiv:2503.06472 , year=

    arXiv

  54. [54]

    arXiv preprint arXiv:2510.18234 , year=

    DeepSeek-OCR: Contexts Optical Compression , author=. arXiv preprint arXiv:2510.18234 , year=

    Pith/arXiv arXiv

  55. [55]

    Advances in neural information processing systems , volume=

    Supervised contrastive learning , author=. Advances in neural information processing systems , volume=

  56. [56]

    Advances in neural information processing systems , volume=

    Gans trained by a two time-scale update rule converge to a local nash equilibrium , author=. Advances in neural information processing systems , volume=

  57. [57]

    Proceedings of the IEEE conference on computer vision and pattern recognition , pages=

    The unreasonable effectiveness of deep features as a perceptual metric , author=. Proceedings of the IEEE conference on computer vision and pattern recognition , pages=

  58. [58]

    IEEE transactions on image processing , volume=

    Image quality assessment: from error visibility to structural similarity , author=. IEEE transactions on image processing , volume=. 2004 , publisher=

    2004

  59. [59]

    Proceedings of the IEEE conference on computer vision and pattern recognition , pages=

    Rethinking the inception architecture for computer vision , author=. Proceedings of the IEEE conference on computer vision and pattern recognition , pages=

  60. [60]

    Advances in neural information processing systems , volume=

    Imagenet classification with deep convolutional neural networks , author=. Advances in neural information processing systems , volume=

  61. [61]

    arXiv preprint arXiv:2505.20793 , year=

    Rendering-Aware Reinforcement Learning for Vector Graphics Generation , author=. arXiv preprint arXiv:2505.20793 , year=

    arXiv

  62. [62]

    arXiv preprint arXiv:1409.1556 , year=

    Very deep convolutional networks for large-scale image recognition , author=. arXiv preprint arXiv:1409.1556 , year=

    Pith/arXiv arXiv

  63. [63]

    2017 , publisher=

    Classification and regression trees , author=. 2017 , publisher=

    2017

  64. [64]

    The Bell system technical journal , volume=

    A mathematical theory of communication , author=. The Bell system technical journal , volume=. 1948 , publisher=

    1948

  65. [65]

    2026 , eprint=

    Beyond Patches: Global-aware Autoregressive Model for Multimodal Few-Shot Font Generation , author=. 2026 , eprint=

    2026

  66. [66]

    2026 , eprint=

    VecGlypher: Unified Vector Glyph Generation with Language Models , author=. 2026 , eprint=

    2026

  67. [67]

    arXiv preprint arXiv:2510.14528 , year=

    Paddleocr-vl: Boosting multilingual document parsing via a 0.9 b ultra-compact vision-language model , author=. arXiv preprint arXiv:2510.14528 , year=

    arXiv

  68. [68]

    2020 , eprint=

    Visual Transformers: Token-based Image Representation and Processing for Computer Vision , author=. 2020 , eprint=

    2020