acceptodds
Withdrawn from ICLR 2027

Differentiable One-Line Drawings

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Image-based one-line drawing seeks to turn a reference into a single stroke that preserves the subject's structure and provides a drawing process that users can inspect and imitate. Existing image generators can suggest the appearance of one-line art without providing a continuous path, while vector synthesis can change the reference geometry or route the stroke without following the subject's form. We propose a contour-guided differentiable framework that couples structural fidelity with locally organized drawing order. It first constructs a path along sampled foreground-boundary points, inserting attention-selected interior points near their closest boundary points. It then refines a cubic spline through image-feature matching and geometric regularization; optional width optimization varies line thickness along the stroke. An attachment-length bound explains the locality of the initialization. Progressive replay exposes the drawing order, as illustrated in the supplemental HTML. Our aim is to turn the visual economy associated with skilled one-line drawing into a process that non-experts can inspect, imitate, and experience.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
One-line drawing, Continuous-line illustration, Euler trail, Route inspection, Differentiable vector graphics

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_DKkKOpenReview ↗
Rating
1
Confidence
2
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

The paper addresses image-conditioned one-line drawing. Given an input image, it outputs a single continuous stroke, represented as an ordered sequence of cubic B-spline control points, optionally with a width field along the stroke. For initialization, anchor points are sampled along the foreground boundary in tracing order and interior points are sampled from a CLIP saliency map. The curve, and optionally the width, is then optimized through a differentiable renderer with a CLIP feature-matching loss and regularizers on bending, control-point spacing, and proximity of non-adjacent segments.

Critical strengths

The method places points not only on the boundary but also in the interior. Attaching saliency-sampled interior points to their nearest boundary anchors lets a single stroke include interior details while the route still follows the outline.

Critical weaknesses
  1. The technical contribution is limited.

The pipeline selects boundary anchors and salient interior points, connects them in a fixed order as B-spline control points, and refines the curve by optimizing a CLIP-based loss. Each step is natural, but the paper offers neither a new representation for drawings nor an algorithmic advance. It states itself that it relies "on standard spline properties rather than introducing a new curve primitive". In particular, the method closely follows CLIPasso (Vinker et al., 2022), and the paper does not clarify this relationship. The loss, the augmentations, the sampling of initial points from a CLIP saliency map, and the use of U2-Net for foreground extraction all follow CLIPasso, which is cited only as related work and as a baseline.

The theoretical statements are elementary: Proposition 1 follows directly from the triangle inequality, and the bounds in Eq. (12) follow directly from the partition of unity. The evidence for the main contribution is also limited. Contour-local insertion is compared only with random assignment, which is a weak baseline. Although the introduction argues against the TSP initialization used by SLDgen and Calligraph, no TSP-initialized variant of the same pipeline is evaluated.

Moreover, although this is subjective, the results of the prior method SLDgen (e.g., Figures 3, 6, and 10) are of very high quality as one-line drawings. The main advantage claimed over SLDgen is closer adherence to the input geometry, which SLDgen, being guided by text and image-derived depth, does not enforce. I understand it is difficult to compare subjective quality, but I woud have liked human evaluations or quantitative measures to be included.

  1. Key elements of the method are not specified.

Renderer: The method relies on "a differentiable renderer", but the renderer is never specified, including how the stroke and its variable width are rasterized and how gradients are obtained.

Scales and normalization: The normalization of coordinates and widths is not stated, so constants such as the clamping of width coefficients to [0.1, 3.0] "in renderer units" are difficult to interpret.

These omissions make the method difficult to understand and to reproduce.

Questions for authors
  1. Which differentiable renderer is used?
  2. Eq. (3) assigns each interior point to an anchor once, and the traversal order is then fixed. How are cases handled in which, after the points move during optimization, this order is no longer appropriate, for example when an interior point ends up closer to a different part of the contour, producing detours or self-intersections?
  3. How are coordinates and widths normalized, and what do the "renderer units" represent?
  4. Beyond the contour-local initialization, which parts of the method differ from CLIPasso, and what do the authors regard as the main technical novelty?
  5. How does contour-local insertion compare with a TSP-based initialization, as used in SLDgen and Calligraph, within the same optimization pipeline?
Reviewer_zQcNOpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
3
Soundness
2
Read review
Paper summary

This paper presents a method for generating single-line drawings from images while preserving the original object structure. The key idea is to follow the outer contour and insert important interior points near their closest contour points to form a continuous drawing path. The path is then refined using differentiable rendering, CLIP-based image loss, and geometric regularization. The method also optimizes line width along the curve and supports drawing replay by progressively revealing the path. Experiments show that the method achieves the highest Image CLIP, Spatial DINO, and Edge F1 scores among the evaluated single-line baselines. It also produces shorter paths than random-anchor initialization and reports the lowest runtime among the compared methods under their native settings.

Critical strengths
  1. Simple and practical design. The contour-guided initialization provides a simple way to construct a continuous drawing path while preserving the object's structure. The method requires no additional training and is easy to implement.

  2. Good structural fidelity. The method achieves the highest Image CLIP, Spatial DINO, and Edge F1 scores among the evaluated single-line baselines. The ablation study also shows that local anchor insertion improves image similarity and reduces path length compared with random assignment. These results demonstrate the effectiveness of the proposed approach.

Critical weaknesses

W1.Insufficient validation of the proposed initialization. The main contribution is the contour-guided initialization strategy, but the ordering ablation only compares nearest-anchor assignment with random assignment. Stronger comparisons, such as TSP-based ordering [1, 2] and other local insertion strategies, using the same sampled points and optimization settings, are missing. In addition, the contribution of the CLIP-based interior sampling is not separately evaluated. These experiments are important to clarify which components are responsible for the reported improvements.

W2. Unclear contribution of the contour prior and subsequent refinement. The method directly extracts the outer contour from the foreground mask, which already provides a strong geometric prior for preserving the object structure. However, it remains unclear how much of the reported improvement comes from this prior rather than the proposed pipeline. A controlled comparison among (1) contour-only spline fitting, (2) contour-guided initialization with interior points but no optimization, and (3) the full method would help clarify the contributions of each stage.

References

[1] Bosch, R. A., and Herman, A. (2004). Continuous line drawings via the traveling salesman problem. Operations Research Letters, 32(4), 302–303.

[2] Magne, T., et al. (2026). Single line drawing generation via semantics-driven optimization. Computer Graphics Forum, e70502.

Questions for authors

Q1. Necessity of Differentiable Optimization. Since the initial drawing path is already constructed from contour and interior points, could a similar result be achieved using conventional spline fitting and geometric refinement without differentiable rendering? How much improvement does CLIP-guided optimization provide over these simpler alternatives, particularly in terms of structural fidelity and computational cost?

Q2. Loss Design and Robustness. The bending and proximity losses encourage smooth curves and discourage nearby segments, but these features may be necessary for fine details or narrow structures. Although the ablation results show overall improvements, how does the method handle such conflicting objectives? How sensitive are the results to the loss weights, especially for objects with complex internal structures?

1 comment

I have not seen a response addressing my main concerns about the technical novelty, comparison with TSP-based initialization, and the separate contributions of the contour prior and differentiable optimization.

I therefore keep my rating.

Reviewer_QUY9OpenReview ↗
Rating
3
Confidence
3
Informativeness
3
Presentation
3
Soundness
3
Read review
Paper summary

This paper studies the problem of how to turn a reference picture into an one-line drawing. To overcome the shortcomings of previous methods, a contour-guided differentiable framework is introduced, which has better structural fidelity under the one-line drawing paradigm. In this framework, the authors use a contour-guided method to initialize the B-spline and then refine it by using a differentiable renderer with clip image alignment and geometry regularization. Experimental results demonstrate a significant improvement of this method compared to previous ones.

Critical strengths

This paper is well written and very interesting. The issue of insufficient structural fidelity in previous methods is addressed through contour-guided initialization, CLIP image alignment, and differentiable rendering. Furthermore, the authors apply geometry regularization to encourage a compact and smooth trajectory.

Critical weaknesses

The main weakness of this paper lies in the experimental evaluation, where only around 50 images are used to compute the metrics, and some of the primary metrics are closely related to the optimization objective. Increasing the size of the test set and introducing additional evaluation methods (e.g., human judgments) may further strengthen this work.

Questions for authors
  1. Could the authors increase the size of the test set and include human judgements in the evaluation?

  2. How sensitive is the method to errors in foreground contour extraction? Can the subsequent image-guided optimization correct an inaccurate initialization?

  3. Is it feasible to replace CLIP with other models, such as DINO models or VAEs used in image generation models?

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

Related papers

Discussion (0)

Sign in to comment.