acceptodds
Under review as a conference paper at ICLR 2027

Continuous Representations Improve On Discrete Ones for SVG Generation

Abstract

Current autoregressive SVG generation models represent coordinates as a sequence of discrete tokens, even though those values describe continuous, geometric positions. We propose an alternative, *continuous* approach that encodes coordinates through a value encoder and predicts them via regression through a value decoder, while retaining ordinary token prediction for drawing commands and markup. In evaluations, our approach achieves higher reconstruction quality than all evaluated discrete baselines. Moreover, data-scaling experiments with 5%, 25%, and 100% of the training data show a persistent advantage over discrete approaches at every training-set size, showing that our approach learns faster and peaks higher than discrete baselines. We investigate this advantage by analyzing *numerical representations*, *coordinate generalization*, and *coordinate precision*. For numerical representations, we find that our embeddings vary more smoothly and are more linearly decodable than discrete representations. For coordinate generalization, continuous predictions remain stable on held-out coordinates, while discrete predictions degrade. For coordinate precision, continuous predictions yield lower teacher-forced coordinate error than their discrete counterparts on fractional targets. Together, these results favor a more natural division of labor: continuous values for geometry and discrete tokens for drawing structure.

Primary area:
generative models
Keywords:
SVG generation, continuous numerical representations, numerical encoding, coordinate regression, vector graphics, multimodal generation, vision-language models

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_k98xOpenReview ↗
Rating
2
Confidence
3
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

The authors tackle the problem of SVG generation with autoregressive LLMs. A prevalent issue in this setting is the generation of coordinate values, which are usually expressed as floating-point numbers and require high precision. The authors speculate that generating these floats as tokenized sequences is suboptimal and propose a continuous-value generation subnetwork as an alternative. This subnetwork fills in placeholder tokens generated by the autoregressive model at positions where coordinate values are expected. The authors show that this approach outperforms the standard approach across a wide range of metrics.

Critical strengths

The paper is well written and easy to follow. The idea makes sense intuitively and works well in practice. The quantitative evaluation shows a large gap compared to previous methods. The findings will be useful to the broader SVG generation community.

Critical weaknesses

My biggest gripe is with the presentation of the results. I believe the results require more discussion. As it stands, the quantitative and qualitative results are presented without much explanation. For example, in lines 312f, what qualitative improvements and differences can we observe in the figures? And in lines 316f, what do the scores tell us? Looking at the results myself, I am confused because the quantitative results rank the discrete models above the standard baseline, yet in the qualitative examples, they seem to perform worse. It also seems that the continuous approach generates better SVGs overall but suffers from greater variance in the predicted coordinates. How would the authors put these observations into perspective? The paper would benefit from a discussion of these points.

Other parts of the paper could benefit from further discussion as well. For example, I couldn't find details on how the discrete baseline works. Does it introduce new tokens for a predefined range of values? Overall, I don't think the paper is ready yet for publication and believe it would benefit from further revision.

Questions for authors

The approach reminds me of a method employed by Kulits et al. (2024) for Blender code generation. What do the authors see as the major similarities and differences between the two approaches?

Reviewer_i8jeOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper introduces new ways of representing coordinates in autoregressive, VLM-based SVG generation. Commands and markup remain ordinary tokens, while each coordinate (scalar variant) or coordinate pair (paired variant) becomes a marker token whose value is encoded with Fourier features and an MLP and regressed from the hidden state by an MLP head. Five coordinate interfaces (plain-text numbers, paired and scalar discrete tokens, and the two continuous variants) are compared on a shared Qwen2-VL backbone at 2B and 7B, trained for one epoch on SVG-Stack. The scalar continuous variant achieves the best raster-to-SVG scores at both scales and across training-data subsets, with additional results on MMSVGBench and on text-to-SVG at 2B.

Critical strengths
  • The problem is well motivated: how coordinates are represented is a basic design choice in LLM-based SVG generation, and the proposed continuous interface yields consistent improvements over matched discrete tokenizations, most clearly for the scalar variant.
  • The comparison is well controlled, with five interfaces sharing the same backbone, data and grammar. It is complemented by analyses of precision on medium-sized models (2B and 7B).
Critical weaknesses
  • Limited novelty: the continuous interface is taken directly from existing work. It combines the marker-token formulation of xVal and the Fourier number embeddings of FoNE with a regression head, and adapts them to a specific domain (SVG generation).

  • The gains are demonstrated mainly on reconstruction. Raster-to-SVG is arguably the most favorable setting for regression, since the input image largely determines the target geometry and a point estimate suffices. Even text-to-SVG is evaluated against a single paired reference drawing, hence generation is still scored as reconstruction, which favors averaged rather than diverse outputs. The paper would be more convincing if it showed a clear benefit in generation, measured with distribution-level metrics (e.g., FID, CLIP score).

  • The references appear small: just among SVG generative models, several relevant works are missing, e.g., SVGFusion: Scalable Text-to-SVG Generation via Vector Space Diffusion (Xing et al., arXiv 2024), Vector Grimoire: Codebook-based Shape Generation under Raster Image Supervision (Cipriano et al., ICML 2025), vHector and HeisenVec: Scalable Vector Graphics Generation Through Large Language Models (Zini et al., NeurIPS 2025) and A Scalable Vector Graphics Latent Space (Zini et al., arXiv 2026).

Questions for authors
  • It would help to show more cases where continuous modeling is actually needed or beneficial.
  • An extension to the generative setting would strengthen the paper, for instance a generative continuous head instead of a single-value regressor, such as a small flow-matching or diffusion head per marker.
  • Please report the text-to-SVG results at 7B; Table 1 reports them only for 2B.
Reviewer_5FmwOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper studies continuous coordinate representations for autoregressive SVG generation. It combines discrete drawing commands with Fourier-based numerical embeddings and coordinate regression, comparing paired and scalar interfaces against discrete alternatives under matched training conditions. Experiments cover reconstruction, text-to-SVG generation, data scaling, and coordinate behavior.

Critical strengths
  1. Matched comparisons across model sizes and data budgets provide useful evidence about coordinate representation choices.

  2. The numerical analyses and encoder/output-head ablations help distinguish coordinate accuracy from successful generation of complete SVGs.

Critical weaknesses
  1. Limited methodological novelty. The controlled study is useful, but the SVG-specific advance beyond adapting numerical representations from xVal and FoNE remains unclear.

  2. Evaluation depends on reference-guided processing. Although the reference raster is available as input, DINOv3-based candidate selection and SSIM-guided prefix repair make gains in raw generation quality difficult to isolate (Sec. 4.3; Appendix A.4).

  3. Complex-SVG generalization remains unclear. MMSVGBench provides some transfer evidence, but examples are icon-centric, and teacher-forced coordinate tests do not establish reliable generation of complex illustrations or long SVG sequences.

Questions for authors
  1. What methodological advances go beyond adapting xVal/FoNE to SVG generation?

  2. Do the gains persist before candidate selection and repair? Please report single-candidate quality and native validity.

  3. Could you report free-running results by path count or sequence length, including complex illustrations and representative failures?

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.