acceptodds
Under review as a conference paper at ICLR 2027

TopoContour: Persistent-Homology-Gated Counterfactual Path Deletion for Fixed-Budget Filled SVG Repair

Abstract

Repairing a filled SVG requires mapping a raster-space defect to the vector path that caused it. We introduce TopoContour, a topology-aware framework that converts unmatched PH events into path proposals and tests them through counterfactual B´ezier path deletion under a fixed geometry budget. Across synthetic data and two real filled-SVG domains, TopoContour consistently improves topology repair and path responsibility relative to residual-driven alternatives. On Tabler Filled and Bootstrap Filled, the topology-repair rate improves over residual-centered gating by 24.7 and 27.5 percentage points, respectively. Compared with exhaustive ungated search, TopoContour evaluates 47.3% and 48.0% fewer candidates while retaining 95.8% and 96.5% of its topology-repair performance. Boundary F1 remains comparable. The same pattern holds for asymmetric blobs and rounded-square B´ezier contours, with topology-repair rates between 0.894 and 0.952. Increasing the path budget from 4 to 16 raises the reduction in candidate evaluations to 86.2% on Bootstrap and 93.1% on Tabler, with counterfactual repair-decision speedups of 12.5× and 13.7×. The results show that PH events can serve as effective, transferable action proposals for efficient path-level vector repair

Primary area:
learning on graphs and other geometries & topologies
Keywords:
Persistent Homology, Counterfactual Path Deletion, SVG Repair, Vector Repair

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_P3dNOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This work studies fixed-budget repair of filled SVGs, where a closed Bézier contour creates a hole absent from the target raster, and the responsible path must be identified and removed without disturbing unrelated geometry. The proposed TopoContour uses unmatched persistent homology hole events between the target raster and the current rendering to gate a small set of candidate paths, and tests each candidate by temporary deletion and re-rendering, abstaining when no deletion helps. Experiments on synthetic data and on Tabler and Bootstrap Filled icons report better topology repair and path responsibility than residual-driven alternatives, while evaluating fewer candidates than exhaustive ungated search at a slightly lower topology-repair rate.

Critical strengths

Matched controls attribute the gain to the PH event. Residual-centered and shuffled-event gating share TopoContour’s gate and counterfactual utility and differ only in the proposal signal, yet both give lower topology repair and responsibility accuracy on Tabler and Bootstrap (Table 5). This is the main evidence that the observed event, rather than counterfactual verification or candidate restriction alone, drives the improvement.

Path responsibility is measured against exact labels. In the vector-native benchmarks, inserting one closed contour into an unused slot gives an exact responsibility label that no method receives, and the gate multiplier selected on Tabler is reused on Bootstrap without retuning. This makes the path-responsibility claim testable in a controlled setting and adds a transfer check.

Critical weaknesses

The tested defect is exactly invertible by the method’s only action. The method can only delete one path or abstain (Sec. 3.1), and every evaluation domain uses a constructed corruption (Table 1). In the main causal evaluation, one closed contour is inserted into an unused slot of a four-path bank without changing the remaining paths, so deleting that path restores the pre-insertion bank exactly. No experiment uses holes produced by a vectorizer or a generative model. This is consequential because the results show that the method can locate and remove an injected contour, while the abstract and conclusion describe the icon benchmarks as “real-domain” evidence for path-level vector repair without testing any naturally occurring defect.

Questions for authors
  1. Could you please explain how the topology-repair rate is defined? It seems the main text lists it as a primary metric (Sec. 4.2) but does not define it. In particular, in Table 4 the global leave-one-out reference has a higher responsibility accuracy (0.7734) than topology-repair rate (0.6719), and Identity, which leaves the bank unchanged, has a topology-repair rate of 0.1094.

  2. Sec. 4.1 states that the target raster is the only reference information available at decision time. Which input does the global leave-one-out reference require beyond that, and why is it excluded from Table 5?

Reviewer_J2o4OpenReview ↗
Rating
1
Confidence
1
Informativeness
1
Presentation
1
Soundness
3
Read review
Paper summary

This paper addresses a specific problem: suppose you have a target icon bitmap along with a current SVG, but the latter renders with an extra white hole that isn't in the target. The paper aims to figure out which path needs to be deleted to fix the hole.

To do this, it first uses persistent homology (PH) to locate the extra hole. Then, it treats only the paths near the hole as candidates, testing the deletion of each one by one and re-rendering to see if the hole disappears.

Critical strengths

The paper poses a well-defined question: which vector path is responsible for a topological defect observed in raster space. The evaluation separates the proposal signal from the decision rule (residual-centered and shuffled-event gates, ungated search) and reports path responsibility alongside the topology metric, including references that outperform the proposed method on some metrics. Beyond this, I find the strengths of the submission limited, for the reasons given below.

Critical weaknesses
  1. The setting is too narrow. Everything uses four closed contours at 128×128, with exactly one hole created by inserting a contour into an empty slot, and the only action is to delete one path. The "real" Tabler and Bootstrap SVGs are crude four-contour fits of the icons (Figs. 2 and 5), not real SVGs. With only four paths, simply trying every deletion takes a few tens of milliseconds, so the problem is not hard in this form. A more convincing setting would be SVGs from vectorizers or LLM-based generators (e.g., StarVector, Rodriguez et al., CVPR 2025; https://openaccess.thecvf.com/content/CVPR2025/html/Rodriguez_StarVector_Generating_Scalable_Vector_Graphics_Code_from_Images_and_Text_CVPR_2025_paper.html), which have many more paths and real defects. I would like to see results there. The B = 8/16 experiment does not answer this, since the extra paths are cancelling duplicates.

  2. Limited novelty. The main problem is that the paper does not give any new knowledge or insight. That a spurious hole is caused by a contour near it, and that checking nearby paths first is cheaper than checking all of them, is what one would expect before running any experiment. The results confirm exactly this (fewer candidates, lower accuracy than trying every path), but the paper does not tell us anything beyond it, for example when and why the PH proposal fails. The only formal part, the Soft Even-Odd composition, is a standard identity and is not used in the experiments. Closely related work is also not cited: Optimize & Reduce (Hirschorn et al., AAAI 2024; https://ojs.aaai.org/index.php/AAAI/article/view/27987) already scores shapes by removing each one and comparing the rendering with the target, and Betti matching (Stucki et al., ICML 2023; https://proceedings.mlr.press/v202/stucki23a.html) addresses how to match persistence features between two images.

  3. Baselines are mostly the authors' own ablations, and there is no comparison with published methods. The one reference that is not a variant of the method, global leave-one-out, is better than TopoContour on responsibility in Table 4 (0.773 vs 0.633), but the paper says it is "not available" (L308-310) even though it only uses the target raster.

  4. Key details are missing: the filtration and how q, r and π are defined on binary images, how hole events are matched, the full utility function, the definition of the topology-repair rate, what the target raster is in the vector-native protocol, and how the icons were selected. There is no code statement.

  5. Presentation (minor). The abstract is hard to follow, with many undefined terms and numbers. PH is not expanded at first use. Figure 1 adds little because it never shows the paths. Tables 4 and 5 bold the whole TopoContour row even where other methods are better, and Table 6 has no baseline rows. A few references are off, e.g., 3D mesh renderers are cited as differentiable vector graphics work (L74-77).

Questions for authors
  1. Can you report results on SVGs with many paths and real defects, e.g., from a vectorizer or an LLM-based generator?
  2. Why is global leave-one-out considered unavailable, given that it only needs the target raster?
Reviewer_3YgqOpenReview ↗
Rating
1
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes TopoContour, a topology-aware framework for repairing filled SVGs with spurious holes. The method uses persistent homology to detect unmatched hole events between the current rendering and the target image, uses these events to gate a small set of candidate Bézier paths, and then verifies each candidate through counterfactual path deletion and re-rendering.The main contribution is the combination of topology-based event localization with path-level counterfactual verification under a fixed path budget. The authors evaluate the method on controlled synthetic cases, Tabler Filled SVGs, Bootstrap Filled SVGs, different corruption shapes, and larger path banks. Results show that TopoContour achieves repair quality close to exhaustive ungated search while substantially reducing the number of candidate evaluations and improving decision-time efficiency.

Critical strengths

-The paper cleanly separates topology-based candidate proposal from counterfactual path verification, giving persistent homology a concrete role in identifying structurally relevant SVG paths. -The evaluation includes synthetic and real SVG domains, transfer across datasets, non-elliptical corruptions, ablations, and scaling to larger path banks, showing consistent repair performance and substantial reductions in candidate evaluations.

Critical weaknesses

-The core method uses persistent homology to propose a smaller candidate set and then applies counterfactual deletion to verify the responsible path. While effective in the studied setting, I am not fully convinced that this combination provides sufficient methodological novelty or broader ML significance for ICLR.

  • The paper suffers from a lack of clear problem definition, use of many terms before they are introduced (e.g. responsible in line 114 used several times earlier) -The main intervention is deletion of a single closed path, and the vector-native benchmark is constructed by inserting one extra contour to create the topological defect. It is therefore unclear how well the approach extends to more general SVG errors that require modifying control points, changing fill structure, or coordinating edits across multiple paths.
  • The paper makes a number of fixed settings that are not evaluated (lambda, path bank...)
  • The empirical evaluation setup seems tailored to the proposed method, with limited scope, and lack of justification as to the fairness of comparison with existing work.
  • Figure 1 is not informative - the illustration does not actually show what is happening.
Questions for authors

The vector-native benchmark creates defects by inserting a single closed contour, while the repair action is also single-path deletion. How dependent is TopoContour on this alignment between the corruption mechanism and the intervention space? Would the same event-to-path reasoning remain valid for defects caused by control-point errors or interactions among multiple paths?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.