- 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
- 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.
- 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
- Which differentiable renderer is used?
- 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?
- How are coordinates and widths normalized, and what do the "renderer units" represent?
- Beyond the contour-local initialization, which parts of the method differ from CLIPasso, and what do the authors regard as the main technical novelty?
- How does contour-local insertion compare with a TSP-based initialization, as used in SLDgen and Calligraph, within the same optimization pipeline?
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.