{"id":"b475bb91-90ec-45e5-a0f1-3bede46382c1","arxiv_id":"2501.13812","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":8,"one_line_summary":"A pipeline converts a single raster texture into a scalable vector texture by segmenting, clustering, and rearranging textons, enabling vector editing operations.","lead":"An image of a rough natural texture is converted into an editable vector texture by cutting it into pieces, grouping similar pieces, and rearranging them over a gradient background. The payoff is that the result is a set of polygons that can be moved, recolored, or deleted, which is impractical with normal pixel images.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The quantitative evidence for the 'competitive' claim rests on a metric table with degenerate SIFID/EMD zeros and no error bars; until those numbers are reproduced with a working protocol, the central demonstration is unverified.","rationale":"The reader correctly flagged the statistics as untrustworthy, but their stated weakest assumption was SAM segmentation reliability. In my reading, the more load-bearing weakness for the central claim is the quantitative comparison itself: the claim of competitiveness is explicitly supported by Table 1, and that table contains internally implausible exact zeros plus no variance information. This is not an ad hominem point; the pipeline could still work, and the qualitative figures and vector-editing demonstrations are genuinely suggestive. However, the objective demonstration that the method 'highly resemble[s] the input' and 'is competitive' depends on metrics that appear degenerate. The concrete test would settle whether the zeros are an artifact or real: if they persist on non-identical images, the comparison protocol is invalid and the paper's central claim is not established as written. I therefore keep the reader's CONDITIONAL verdict, with the condition that Table 1 be recomputed with a corrected, reproducible metric pipeline and per-exemplar statistics.","tokens_in":12589,"tokens_out":7062,"duration_ms":66952,"concrete_test":"Request or rerun the authors' metric pipeline on the actual outputs: compute SIFID and EMD between an exemplar and a known non-identical image, such as the Image Quilting result in Figure 6. If these values still round to 0.000, the metric harness is broken. Then recompute Table 1 using only validated metrics, reporting per-exemplar standard errors; if Ours remains best or second-best on no validated metric, the 'competitive' claim should be weakened to 'produces plausible vector textures with comparable PieAPP/SIFID on select exemplars.'","verdict_should_be":"UNCHANGED","load_bearing_attack":"The strongest claim explicitly asserts that experiments show the method 'is competitive when compared to methods that operate directly on the raster domain.' Table 1 is the only quantitative support for that assertion, and it is suspect: SIFID is reported as 0.000 for Image Quilting, Self-Tuning, PSGAN, and Ours, and EMD as 0.000 for three baselines, even though the compared images are visibly non-identical (Figure 6). A SIFID of exactly zero to three decimals for four different outputs indicates a degenerate computation (e.g., a broken feature-statistics path or unintended resizing), not genuine perceptual equivalence; EMD zero likewise implies identical intensity histograms, which is implausible across different synthesis methods. Once those degenerate entries are set aside, Ours is best on no metric and worst or near-worst on LPIPS, DISTS, BRISQUE, and CLIP-IQA, so 'comparable scores in general' is not demonstrated. The table also reports single means over eight exemplars with no per-exemplar variance, so even the non-degenerate differences are not statistically supported. Section 4.6's admission that 'segmentation is not always reliable' appropriately limits scope, but it does not repair the evaluation; the central capability may still be real, but the published evidence for it is currently unreliable.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a two-stage method for synthesizing arbitrarily sized vector textures from a single raster exemplar. In the analysis phase, Segment Anything is used to segment primary textons, which are clustered by color, area, and compactness; neighborhood descriptors capture inter-texton spatial relationships, secondary textons are extracted by flood fill, and the background is summarized by a Voronoi-based color field. In the synthesis phase, primary textons are placed by descriptor-guided scoring on a target map, secondary textons are placed by Poisson-disk sampling, and the background is generated by inverse-distance interpolation of sampled colors. The authors claim that their method produces vector textures that highly resemble the input exemplars and is competitive with raster-domain texture synthesis methods, and they demonstrate several vector-editing operations enabled by the representation.","tokens_in":1691,"tokens_out":2061,"duration_ms":58198,"significance":"If the central result is established, this is a genuinely novel contribution: it directly converts a natural raster texture into an editable vector representation and synthesizes arbitrarily large vector textures from a single exemplar, which previous vector-texture methods do not do. The paper has clear strengths: the pipeline is described in enough detail to be reimplemented, the qualitative results in Figures 5 and 6 are often convincing, an ablation study is included, and Section 4.6 honestly states the segmentation-dependent scope of the method. However, the quantitative support for the 'competitive' claim is currently unreliable because the headline metric table contains degenerate entries and no measures of variance; until those numbers are corrected and reproduced, the central claim rests mainly on the qualitative figures.","major_comments":[{"comment":"The SIFID value is reported as 0.000 for Image Quilting, Self-Tuning, PSGAN, and Ours, and the EMD value is 0.000 for three baselines, even though Figure 6 shows visibly non-identical outputs. Exact zeros to three decimals are implausible for a genuine feature-statistics or histogram-distance computation and indicate a degenerate evaluation path (for example, an unintended resize, a zero-covariance feature branch, or a failure to compare comparable image sizes). This directly undermines the sentence in Section 4.2 claiming that the method is 'competitive with state-of-the-art methods, with comparable scores in general and among the top SIFID scores.' The authors need to explain the exact metric setup, correct these entries, and report values that distinguish the four methods.","section":"Section 4.2, Table 1"},{"comment":"All quantitative comparisons report only means over eight exemplars, with no per-exemplar values, error bars, confidence intervals, or significance tests in the main text. Observed differences such as LPIPS 0.527 vs 0.467, or BRISQUE 20.665 vs 19.400, are plausibly within sampling variability, so the claim of being 'competitive' is not statistically supported. The ablation table is especially concerning: SIFID is 0.042 for all six conditions, which suggests the metric is insensitive to the ablations or is computed incorrectly. Please report per-exemplar distributions or error bars, and clarify why the ablation SIFID values are identical.","section":"Section 4.2-4.3, Tables 1-3"},{"comment":"The global density correction explicitly deletes textons from each category until the fractional area covered by the category drops below the exemplar's fractional area. This means that any statement about matching the exemplar's texton density, including the coverage comparison in Table 2, is partly enforced by construction rather than produced by the synthesis model. The paper should state this explicitly when interpreting coverage results and, ideally, report coverage with the correction disabled to show how much of the match is due to this post-processing step.","section":"Section 3.3, Global density correction"},{"comment":"Several central parameters are hand-tuned without sensitivity analysis: the 40th-percentile secondary-texton spacing, the scoring weights w = (0.5, -0.4, -0.2, -0.5, -0.5), the threshold for promoting secondary textons to primary, and the density-correction target. Since the paper's quality claim depends on these choices, a small sensitivity study (for example, varying the percentile across the 30th-60th range that the authors say is stable, and varying the scoring weights by a fixed percentage) would show that the results are not an artifact of a single tuned setting.","section":"Section 3.1.2 and Section 3.2.1"},{"comment":"The paper dismisses a raster-to-vector-then-synthesize pipeline by saying that vectorization tools 'often create output that is not separable, with many overlapping polygons,' but it does not provide an experiment demonstrating that failure. Because the novelty claim is that no existing method directly synthesizes vector textures from natural raster images, a concrete comparison against a vectorization-plus-existing-vector-synthesis baseline would make that claim much stronger and would also clarify the advantage of the proposed pipeline.","section":"Section 1 and Section 2"}],"minor_comments":[{"comment":"There is a typo in the phrase 'representative of the state of the are of GAN-based methods'; it should read 'state of the art.'","section":"Section 2"},{"comment":"The scoring formula is written as 'sum from i=5 to 5 w_i A_i'; it should be 'sum from i=1 to 5 w_i A_i' to match the five listed area terms.","section":"Section 3.2.1"},{"comment":"The sentence 'Table 4 reports the average amount of time required to synthesize 4 500 × 500 exemplars' is ambiguous; it should say four 500x500 exemplars. The table caption also uses 'average time to synthesize a 500 x 500 exemplar,' which should be reconciled.","section":"Section 4.4"},{"comment":"The Image Quilting timing is flagged as coming from an unoptimized third-party implementation, so the timing comparison is not apples-to-apples; consider reporting only the authors' own timings or moving the third-party timing to a footnote with a clearer caveat.","section":"Section 4.4, Table 4"},{"comment":"The ablation discussion states that condition E is preferred by the metrics but that its textures contain substantial repetition not captured by the metrics; consider adding a repetition-aware quantitative measure or at least reporting the visual-inspection criterion more explicitly, since the current text makes the metric-based selection and the final choice inconsistent.","section":"Section 4.3"}],"recommendation":"major_revision","confidential_remarks":"The paper is likely a good fit for a graphics or computational aesthetics venue, and the qualitative results suggest a real contribution. The key obstacle to acceptance is the quantitative evaluation: the degenerate SIFID/EMD zeros in Table 1 and the lack of error bars in Tables 1-3 mean that the 'competitive' claim is unsupported as written. These issues are fixable within the manuscript's scope, so I recommend major revision rather than rejection. I would also ask the editor to ensure that the supplementary material contains the per-exemplar metric tables promised in Section 4.2, since those are essential for judging the corrected results."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Read it. The core idea is real: this is the first pipeline I know of that goes straight from a raster exemplar to a synthesizable vector texture, and the qualitative figures show it can work for the narrow class of textures it targets—distinct, non-overlapping textons on a smooth background. The descriptor-based placement and the secondary-texton/background-gradient construction are sensible engineering. Credit where due: the paper states its assumptions clearly, and Section 4.6 openly admits that segmentation reliability is the weak link. That is honest.\n\nThe soft spot is exactly where the stress-test note lands. Table 1 reports SIFID = 0.000 for Image Quilting, Self-Tuning, PSGAN, and Ours, and EMD = 0.000 for three baselines. Those numbers cannot be right for visibly different outputs; a degenerate metric computation is the likely cause. Set them aside and 'Ours' is best on nothing and worst or near-worst on LPIPS, DISTS, BRISQUE, CLIP-IQA. So the claim of being 'competitive' with raster methods is not demonstrated by the evidence shown. The table also gives single means over eight exemplars without variance, so even the plausible numbers carry no statistical weight. The hand-tuned weights, the 40th-percentile spacing, and the global density correction that forces the output texton area fraction to match the exemplar all make the evaluation weaker than the qualitative results alone. The ablation is also a bit odd: variant E wins on the metrics but looks repetitive, which says the metrics are not the whole story.\n\nThe circularity burden is low in the classic sense—nothing is derived and then tested against itself—but the density correction is an explicit constraint that hides coverage errors, so the reported 'similarity' partly comes from a fitted knob rather than from the synthesis model.\n\nWho is this for? Graphics practitioners who want editable vector textures from natural images, and researchers working on element-based texture synthesis. A serious referee should engage with it, because the capability is new and the limitations are acknowledged. But the revision must fix the metric pipeline, report per-exemplar variance, and ideally release code or at least the synthesized vector outputs. Without that, the quantitative claims are unverifiable.\n\nRecommendation: send to peer review, but expect a major-revision outcome. The idea is sound; the evidence is not yet.","headline":"Genuinely new raster-to-vector texture synthesis pipeline with plausible qualitative results, undermined by a degenerate quantitative evaluation that does not support the 'competitive' claim.","tokens_in":13386,"tokens_out":2297,"would_cite":false,"duration_ms":18738,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"A single raster texture can be converted into an editable vector texture of any size.","keywords":["vector texture synthesis","by-example synthesis","texton","image segmentation","non-parametric texture synthesis","raster to vector","editable vector graphics","perceptual similarity metrics"],"falsifier":"Take an exemplar with overlapping textons or cast shadows, run the pipeline, and check whether the output vector texture contains detached shadow fragments and distorted texton shapes, as the paper's own failure cases suggest; alternatively, compare the pair-correlation function of texton centers in the synthesized output against the exemplar and look for statistically significant deviations.","tokens_in":12366,"feed_emoji":"🎨","tokens_out":6075,"duration_ms":49872,"temperature":0.7,"pith_summary":"This paper proposes a method to take one raster image of a natural, irregular texture—rust, bark, pebbles—and turn it into a vector texture made of individual solid-colored polygons on a gradient background. The authors aim to show that this conversion and re-synthesis preserves the look of the original, so the vector result is both faithful and arbitrarily scalable. The payoff is editability: since every texton is a separate polygon, operations such as swapping, deleting, recoloring, or orienting elements become simple, while the same edits are difficult on a raster image. The paper reports that its outputs are perceptually competitive with state-of-the-art raster texture synthesis methods when judged by several perceptual metrics.","feed_headline":"Raster texture becomes an editable vector texture of any size","feed_subtitle":"New method segments textons and re-synthesizes them with neighborhood descriptors, enabling per-element edits.","key_machinery":"The load-bearing object is the texton neighborhood descriptor: a two-dimensional grid centered on a primary texton whose cells store either the category label of a texton found at that location or a code for 'empty', extended to include textons that protrude past the initial boundary. During synthesis, candidates are placed by scoring them against an incrementally-built target map using a weighted sum of five terms: overlap between the candidate and the desired target region, uncovered target area, coverage of cells marked empty, overlap with a mismatched category, and overlap with another region of the same category. The placement with the highest score is written into the map, while secondary textons are scattered by Poisson-disk sampling and the background is reconstructed from a Voronoi-derived color pool using inverse-distance weighted interpolation.","core_discovery":"The central claim is that a raster exemplar composed of distinct, non-overlapping textons can be decomposed by segmentation into primary textons, secondary textons, and a background, then re-synthesized as a novel vector texture through a non-parametric placement process. The key discovery is that a local neighborhood descriptor—a grid map recording which texton categories surround a given texton—plus a weighted scoring of candidate placements based on overlap and coverage is sufficient to reproduce convincing inter-texton arrangements without modeling global statistics. The paper demonstrates the method on a variety of natural textures and shows that the resulting vector images support element-wise editing operations that raster outputs do not.","pith_inferences":["If the descriptor-based placement strategy generalizes, the same arrangement machinery could be applied to non-texture domains, such as procedural arrangement of discrete objects in illustrations or scene layouts, where the 'textons' are any category-labelled elements.","Because the pipeline leans on an automatic segmentation model, future improvements in open-world segmentation should directly widen the class of textures the method can handle, including partially occluded or shadowed ones that currently fail.","The ablation results suggest the perceptual metrics used do not penalize visible texton repetition, since the condition without a repetition penalty scores best on several metrics while producing visibly repetitive output; a natural testable extension is to use a metric that explicitly measures repetition or pair-correlation statistics.","A direct extension of the method could condition the target map on a user-supplied density or orientation field, turning the current stationary-texture synthesizer into a tool for controlled, non-stationary vector texture design."],"forward_implications":["Vector textures can be produced directly from ordinary raster photographs of materials, removing the need for manual vectorization or pre-separated vector elements.","The synthesized textures preserve enough inter-texton spatial relationships to be perceptually comparable to raster-only methods such as Image Quilting, PSGAN, GCD Loss, and Self-Tuning Optimization.","Because the output is a set of discrete polygons, users can perform element swaps, density-map-driven deletion, forced anisotropy, and texture simplification, all demonstrated in the paper.","The method scales to arbitrary output size: a 500×500 exemplar yields a 1000×1000 vector texture in about four minutes of synthesis time, and larger outputs are limited only by placement effort."],"supporting_citations":[{"why":"Supplies the segmentation step that extracts primary textons from the raster exemplar.","marker":"[KMR∗23]"},{"why":"Most closely related prior work in clustered vector textures; the paper compares against it and positions its own contribution relative to it.","marker":"[TWZ22]"},{"why":"Image Quilting is the raster patch-based baseline the paper compares against.","marker":"[EF01]"},{"why":"Fast Poisson disk sampling provides point distributions for secondary textons and background synthesis.","marker":"[Bri07]"},{"why":"Inverse-distance weighted interpolation is the mechanism that turns the background color samples into a gradient field.","marker":"[She68]"},{"why":"Cited as the source for Poisson disk/stochastic sampling used to estimate inter-element spacing.","marker":"[DW85]"},{"why":"PSGAN is the GAN-based comparator in the evaluation.","marker":"[BJV17]"},{"why":"Guided correspondence loss is the neural texture synthesis comparator in the evaluation.","marker":"[ZCXH23]"},{"why":"Self-tuning texture optimization is the optimization-based comparator in the evaluation.","marker":"[KNL∗15]"}],"fun_headline_variants":["Single raster exemplar becomes scalable editable vector texture","Vector texture synthesis from one raster sample","Editable vector textures from a single raster patch","By-example vector textures, editable and scalable"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The whole pipeline depends on the segmentation reliably isolating distinct, non-overlapping, shadow-free textons; if the segmentation misses or merges textons, no descriptor can be built and the synthesis fails.","fun_headline_variants_meta":{"raw":{"variants":["Single raster exemplar becomes scalable editable vector texture","Vector texture synthesis from one raster sample","Editable vector textures from a single raster patch","By-example vector textures, editable and scalable"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000147,"raw_usage":{"total_tokens":1106,"prompt_tokens":784,"completion_tokens":322,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":400,"completion_tokens_details":{"reasoning_tokens":265}},"tokens_in":400,"tokens_out":322,"duration_ms":3019,"temperature":1.0,"reasoning_tokens":265,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T15:34:27.523715+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take an exemplar with overlapping textons or cast shadows, run the pipeline, and check whether the output vector texture contains detached shadow fragments and distorted texton shapes, as the paper's own failure cases suggest; alternatively, compare the pair-correlation function of texton centers in the synthesized output against the exemplar and look for statistically significant deviations.","supporting_citations":[],"review_version":1}