{"id":"61f076ce-dd4c-456a-b07f-08128f470327","arxiv_id":"2506.01441","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"Local color strokes propagate to semantically similar image regions by optimizing an edited palette in a combined color and semantic feature space.","lead":"A new color editing method uses a neural network to read image content, builds a small palette tied to semantic regions, then solves an optimized palette that spreads a user's local color strokes across all pixels with the same object or material. It aims to make recoloring as simple as a few strokes while avoiding the global color shifts that earlier semantic editing methods produce.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Eq. 9's propagation term is internally inconsistent: α_j is high for pixels most similar to the user stroke, so the objective penalizes the exact color changes the method claims to propagate, and the solved palette cannot faithfully deliver the user's target color.","rationale":"The reader's evaluation-focused concerns about five examples, tuned parameters, and missing code are valid but secondary. The more decisive problem is internal to the method: the energy function that is supposed to implement content-aware propagation appears, as written, to penalize the changes it is meant to achieve. This is an internal inconsistency, not a disagreement with an external baseline, so it bears directly on the central claim that 'local user edits are extended in a content-aware manner.' If Eq. (9) is exactly what was optimized, the solved palette will be a compromise between the target color and the original color, so accurate propagation is not achieved even on the reported examples. If Eq. (9) is a typographical error, the correct weighting must be stated and the experiments rerun with it. I would keep the reader's conditional verdict but shift the binding condition: the first requirement is not a larger dataset but a corrected, self-consistent Eq. (9) plus a minimal numerical demonstration that the objective moves similar-region pixels toward the target color while leaving dissimilar pixels unchanged. No judgment about authorial intent is implied; the concern is about the technical content of the paper as written.","tokens_in":7992,"tokens_out":8185,"duration_ms":97857,"concrete_test":"Implement Eq. (7) verbatim on a synthetic image with one palette entry and two regions: one region semantically similar to the user stroke and one dissimilar. Set the target color of the stroke to a known shift δ, compute α_j by similarity to the stroke as stated, sample both regions, and solve for P′ using COBYLA. Record the resulting color of the similar region. If the printed formula is used, the stroke and similar-region colors will converge to a compromise strictly between the original and target colors; if the text's intent is followed, the loss should instead use a dissimilarity weight such as Σ(1−α_j)||I'_j−I_j||². Only the variant that drives the similar-region pixels to the target color supports the paper's propagation claim.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The central claim is that sparse strokes are propagated to all pixels sharing the edited semantics. That depends on the energy in Section II-B. In Eq. (9), the propagation term is E_prop = (1/Σα_j) Σ α_j ||I_j.C − I'_j.C||², with α_j defined as 'the similarity of I_j to the most similar pixel in user strokes.' So α_j is largest for pixels most similar to the stroke, and the term assigns the largest penalty to exactly those pixels when their color changes. The sentence following Eq. (9) says the term should 'penalize those pixels [that] have larger color variations but not similar with user strokes'; that intent requires weighting by dissimilarity, e.g., (1−α_j), not by α_j. As printed, for a stroke pixel that appears in both H and G, the fidelity term (Eq. 8) pulls I' toward the target while the propagation term pulls I' back toward the original color, so the optimum cannot satisfy the user's requested color. More generally, the objective suppresses color change in the very semantic region where propagation is supposed to occur. Since I'_j.C is linear in the palette edit through Eq. (6), E_prop as written is a quadratic that vanishes at zero edit and biases the solution toward no change for all α-weighted samples. This is not a tuning issue: it is a mismatch between the stated objective and the formula. Without code or a corrected derivation, the reported results do not establish that the implemented optimization follows Eq. (7) as written.","agreement_with_reader":"disagree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a semantic palette-guided color propagation method. In a first step, per-pixel semantic features from a neural network (Aksoy et al.) are reduced via PCA to three dimensions and concatenated with RGB color, yielding a 6D feature space. A modified k-means algorithm with SLIC-based sampling and an adaptive initial-center selection (Eqs. 1-2) extracts a semantic palette. In the second step, an edited palette is obtained by minimizing an energy function (Eq. 7) that combines a fidelity term (Eq. 8) and a propagation term (Eq. 9), and the edited palette is applied to all pixels via Eq. 6. Experiments on five images compare the method with five prior approaches, reporting lower MSE and higher PSNR/SSIM, along with qualitative comparisons and an ablation of the propagation term.","tokens_in":8327,"tokens_out":5164,"duration_ms":58674,"significance":"If the method works as described, the idea of jointly using color and semantic features in a palette-based editing framework is a reasonable and potentially useful contribution to interactive image editing. The paper clearly specifies the feature construction, the clustering procedure, and the energy function, which makes the approach reproducible in principle. The qualitative examples suggest that the method can achieve content-aware propagation when the energy is implemented consistently with the stated intent. However, the paper's experimental support is thin: only five images, no error bars or statistical tests, and the hyperparameters w_c, w_s, and t are tuned on the same examples used for evaluation. The most serious issue is that Eq. (9), as printed, appears to contradict the stated goal of propagation, which undermines confidence that the reported results follow from the described method.","major_comments":[{"comment":"The propagation term as written is internally inconsistent with its stated purpose. In Eq. (9), E_propagation is a weighted sum of squared color deviations ||I_j.C - I'_j.C||^2 with weights α_j, where α_j is defined as \"the similarity of I_j to the most similar pixel in user strokes.\" Thus α_j is largest for pixels most similar to the user strokes, and the term assigns the largest penalty to exactly those pixels whose colors should be propagated. The sentence following Eq. (9) says the term should \"penalize those pixels [that] have larger color variations but not similar with user strokes,\" which would require weighting by a dissimilarity measure such as (1 - α_j), not by α_j. As printed, the objective suppresses color change in the semantic region where propagation is supposed to occur, and the solution of Eq. (7) cannot faithfully deliver the user's target color to non-stroke pixels in the target semantic region. Please correct the formula or carefully redefine α_j; if the implementation actually uses a different weight, provide the correct derivation and, ideally, code to verify the energy that is minimized.","section":"Section II-B, Eq. (9)"},{"comment":"The hyperparameters w_c, w_s, and t are evaluated and selected on the same five images that are later used for the quantitative comparison in Table I. For example, Section III-A states that \"we find that setting w_c = 1, ws = 3 and t = 0.80 can generate palettes with an appropriate number of entries... So we use these values to extract semantic palettes for all examples,\" with the evaluation shown in Fig. 3 on the same examples. This is tuning on the evaluation set, which inflates the apparent performance and means the reported MSE/PSNR/SSIM values do not estimate performance on unseen images. Please provide a separate test set, cross-validation, or at least a sensitivity analysis that shows the results are stable over a range of parameters.","section":"Section III-A and III-C"},{"comment":"The ground truth column \"GT\" in Fig. 5 is not defined anywhere in the manuscript. It is unclear whether the ground-truth edited images were produced by a human expert, by another algorithm, or by a synthetic procedure (e.g., applying a known color change to a segmentation mask). Since Table I reports quantitative errors against GT, the meaning of those numbers depends entirely on how GT was created. Please specify the GT generation protocol, including the number of annotators (if any) and the exact editing instructions.","section":"Section III-C, Fig. 5 and Table I"},{"comment":"The quantitative evaluation is based on only five images (Man, Fox, Sky, Milk, Bear) with a single metric value per method per image and no error bars, confidence intervals, or significance tests. With n = 5 and hyperparameters tuned on the same images, the claim that the method is consistently better than the baselines is not statistically supported. Please expand the evaluation set and report variance or perform a paired test.","section":"Section III-C"}],"minor_comments":[{"comment":"The normalization in Eq. (9) has an index inconsistency: the denominator is written as 1/∑_{i=1}^{|G|} α_j, mixing indices i and j. It should be 1/∑_{j=1}^{|G|} α_j.","section":"Section II-B, Eq. (9)"},{"comment":"The description of σ_c and σ_s is ambiguous: \"both determined by averaging the color and semantic features across all palette entries\" could mean the average of per-palette-entry standard deviations, the standard deviation of palette entries, or something else. Please clarify the exact formula.","section":"Section II-A, Eq. (3)"},{"comment":"The coefficients λ_{i,j} are said to be obtained by solving a linear system from [12], but the system is not stated and the role of the RBF expansion in Eq. (4) is not explained. A brief derivation or citation to the specific equations in [12] would improve reproducibility.","section":"Section II-A, Eq. (4)"},{"comment":"The energy is described as a \"weighted sum\" of a fidelity term and a propagation term, but no explicit weights appear in Eq. (7). If the two terms are meant to be equally weighted, please state that explicitly; otherwise, the balancing weights should be part of the equation.","section":"Section II-B, Eq. (7)"},{"comment":"The statement \"we uniformly sample 256 pixels\" should clarify whether these 256 pixels are sampled from the whole image, from all non-stroke pixels, or from the set of superpixel centroids. This affects the definition of the propagation term.","section":"Section II-B"},{"comment":"The text says that combining a smaller w_c and a larger w_s usually yields more palette entries, but the exact relationship and the impact on the final edit are only shown qualitatively in Fig. 3. A quantitative sensitivity plot would be more informative.","section":"Section III-A"},{"comment":"The abstract and conclusion claim \"Extensive experiments\" and \"efficient and accurate pixel-level color editing,\" but the experimental section uses only five images. Please either expand the experiments or soften the wording to match the actual scope.","section":"Abstract and Conclusion"}],"recommendation":"major_revision","confidential_remarks":"The main technical concern in this report—the inconsistency in Eq. (9)—is load-bearing and should be resolved before the paper can be accepted. Even if the formula is corrected, the experimental section needs substantial strengthening (more images, defined ground truth, and a proper parameter-selection protocol). If the authors can provide code to confirm that the implemented energy matches a corrected Eq. (9), that would materially increase confidence. The paper is within the scope of the journal, but the current evidence is insufficient for acceptance."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Here's my take on the semantic palette paper. The central problem is Eq. (9). The propagation term is supposed to penalize color changes in pixels that are not similar to the user strokes, but α_j is defined as the similarity to the stroke. So the term heavily penalizes changes in exactly the pixels that should be recolored. The fidelity term pulls stroke pixels toward the target color, the propagation term pulls them back; the optimum can't satisfy the user's request. This is a load-bearing inconsistency: without it, the method doesn't propagate. Maybe the implementation used (1−α_j) or something, but as written, the math doesn't support the claim.\n\nWhat's genuinely new: the idea of extracting a palette in a 6D color+semantic PCA space, with adaptive k-means (threshold t) for the number of entries, and using that palette to carry local edits to distant similar regions. That's a plausible extension of the authors' earlier work and of LoCoPalettes, and the pipeline is clearly motivated. The visualization of similarity weights (Fig. 2) is helpful.\n\nThe experiments are thin: five images, no error bars, no code or data, and the hyperparameters (wc, ws, t) are tuned on the same images used in the comparisons. That alone would justify a conditional acceptance request. But the Eq. 9 problem is more fundamental: it's the mechanism that is supposed to make propagation content-aware. A referee would need to see a corrected derivation and ideally code to verify the sign.\n\nThe rest of the method section is coherent and the math is otherwise fine. The citation pattern looks honest, with appropriate references to prior palette methods. The paper is not a home run, but it's not a desk reject either; the idea is worth exploring.\n\nWho's it for: someone working on interactive color editing or edit propagation. They'd get a plausible idea and a cautionary example of how a small sign error can sink a core claim. I'd send it to peer review, but flag Eq. 9 as a major revision. If the authors can fix the formula and add real evaluation (more images, separate parameter selection, code), it could be acceptable. As it stands, I'd not cite it.","headline":"The propagation term in Eq. (9) is backwards as written, which undermines the core method, though the semantic palette idea itself has some merit.","tokens_in":8832,"tokens_out":4052,"would_cite":false,"duration_ms":41989,"reading_group":"maybe","serious_thinker":"no","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"The paper claims that local color edits made with sparse, non-fine-tuned strokes can be propagated to all semantically similar regions by extracting a semantic palette in a 6D color-semantic feature space and solving an edited palette via…","keywords":["color propagation","semantic palette","content-aware color editing","palette-based recoloring","edit propagation","semantic features","energy optimization","user strokes"],"falsifier":"Take a held-out image containing two visually distinct instances of the same semantic category (for example, two flowers under different lighting) next to a similarly colored but semantically different object, place a stroke on one flower, and run the method with the fixed parameters $w_c = 1$, $w_s = 3$, $t = 0.80$; the central claim is falsified if the stroke fails to reach the second flower, or reaches the similar-colored distractor.","tokens_in":7825,"feed_emoji":"🎨","tokens_out":8499,"duration_ms":81988,"temperature":0.7,"pith_summary":"The paper proposes a two-step color-propagation method built around a semantic palette: first it builds a small set of palette entries in a six-dimensional space that mixes RGB color with three PCA-compressed semantic-feature dimensions, then it solves for an edited palette whose color shifts match sparse user strokes while a propagation term suppresses changes to dissimilar regions. The central claim is that local edits made with only a few rough strokes extend faithfully to every region that shares the same object or material, without the global color drift that palette recoloring often causes and without the dense stroke requirements of classic edit propagation. If the method works, image recoloring becomes a matter of marking one petal, one patch of wall, or one object and having the edit reach all its counterparts while leaving neighboring but different content untouched.","feed_headline":"Sparse strokes recolor every matching object","feed_subtitle":"Sparse strokes spread to every region that shares the same content, without global color shifts.","key_machinery":"The central object is the semantic palette, a set of k cluster centers in a 6D feature space where each pixel is represented as $(r, g, b, \\kappa_1, \\kappa_2, \\kappa_3)$, with $\\kappa$s being PCA-reduced semantic features from the soft-segmentation network. A modified k-means builds the palette without a preset k: sampling points come from SLIC superpixel centroids, initial centers are chosen greedily by importance weights updated with a repulsion factor $(1 - \\exp(-d^2))$, and selection stops when the largest weight falls below threshold $t$. Pixel-to-palette similarity is computed by a radial basis function that factorizes color and semantic distances, and the coefficients in the similarity definition are recovered by solving the linear system used in palette-based recoloring. The edited palette is the minimizer of the energy $E = E_{\\text{fidelity}} + E_{\\text{propagation}}$, where $E_{\\text{fidelity}}$ averages the $L^2$ error between transferred and target colors on stroke pixels and $E_{\\text{propagation}}$ penalizes color shifts at 256 sampled pixels weighted by their similarity to the strokes. This palette-plus-energy machinery is what lets local strokes drive a global but content-aware recoloring.","core_discovery":"On the paper's own terms, the discovery is that a palette extracted from a fused color-semantic feature space can act as a carrier for content-aware color propagation. Each palette entry is not merely a dominant color but a cluster center in the 6D space, so its similarity weights (computed with a radial basis function over color and semantic components and normalized over the palette) concentrate on pixels belonging to the same semantic regions. Given user strokes, the paper solves an edited palette by minimizing an energy function with two terms: a fidelity term that forces stroked pixels to their target colors and a propagation term that penalizes large color changes at sampled pixels dissimilar to the strokes. The solved palette is then applied through a standard palette color-transfer equation, giving pixel-level edits that extend to semantically similar but spatially distant regions. The paper reports that this requires only sparse, non-fine-tuned strokes and that on five test examples it yields lower MSE and higher PSNR and SSIM than the compared methods.","pith_inferences":["The authors leave open whether the solved palette transfers across images: if the same semantic palette were applied to a second image of the same scene or object class, propagation might work with no new strokes, but the paper does not test this.","The 6D space separates color from semantics, so the same energy equation could be repurposed to propagate other attributes such as lightness, texture, or material by swapping the color component; this is an extension, not a claim of the paper.","The uniform 256-point sampling for the propagation term is a hidden capacity limit; on very large or detail-rich images, propagation quality may depend on whether those samples cover the semantically relevant regions, a sensitivity the paper does not report."],"forward_implications":["A user can recolor one instance of an object and have the edit apply to all other instances of that object, even ones with different lighting or texture.","Palette editing and pixel-level edit propagation merge into one interaction: the palette is an intermediate, not the final editing surface, so users can stroke the image directly.","The method's reliance on continuous semantic features rather than a hard segmentation result means propagation does not inherit segmentation-boundary errors.","Because the edited palette is solved per image from sparse strokes, the same interaction can edit many semantically similar regions in a single stroke without global color shift.","On the paper's five test examples, the method reports lower MSE and higher PSNR and SSIM than the five compared methods."],"supporting_citations":[{"why":"Supplies the per-pixel semantic features that are PCA-compressed and fused with RGB to form the 6D feature space; without this network the palette would have no semantic component.","marker":"[24]"},{"why":"Provides SLIC superpixels whose centroid pixels are used as sampling points for the modified k-means palette extraction.","marker":"[25]"},{"why":"Supplies the palette-based recoloring machinery, including the linear system for the RBF similarity coefficients and the color-transfer equation that applies palette changes to pixels.","marker":"[12]"},{"why":"Compared method; an RGB-space geometric palette approach that lacks semantic information and serves as a baseline the method must beat on content-aware propagation.","marker":"[18]"},{"why":"Compared method; an edit-propagation-via-color-palettes approach that provides a palette-based baseline without semantics.","marker":"[9]"},{"why":"Compared method; a content-aware palette recoloring approach that neither adaptively determines palette size nor supports pixel-level editing, the baseline closest in aim.","marker":"[23]"},{"why":"Compared method; a deep learning edit-propagation approach that represents the dense-stroke requirement and limited distant-pixel propagation the method aims to overcome.","marker":"[11]"},{"why":"Compared method; a hierarchical semantic-segmentation-based palette editing approach whose segmentation dependence the method contrasts with its own continuous semantic-feature propagation.","marker":"[22]"}],"fun_headline_variants":["Semantic palette spreads sparse strokes to matching objects","Recolor all similar objects with one semantic palette","Content-aware color propagation via semantic palette","Sparse strokes, semantic palette, accurate object recoloring"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The method assumes that the PCA-reduced semantic features reliably put a stroke on the intended semantic class and that the parameter values $w_c = 1$, $w_s = 3$, and $t = 0.80$, chosen on the evaluation examples, transfer to unseen images; if those features or settings fail, unrelated regions can join the same palette entry and the edit propagates incorrectly.","fun_headline_variants_meta":{"raw":{"variants":["Semantic palette spreads sparse strokes to matching objects","Recolor all similar objects with one semantic palette","Content-aware color propagation via semantic palette","Sparse strokes, semantic palette, accurate object recoloring"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000529,"raw_usage":{"total_tokens":2512,"prompt_tokens":869,"completion_tokens":1643,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":485,"completion_tokens_details":{"reasoning_tokens":1593}},"tokens_in":485,"tokens_out":1643,"duration_ms":14645,"temperature":1.0,"reasoning_tokens":1593,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T11:42:43.979078+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a held-out image containing two visually distinct instances of the same semantic category (for example, two flowers under different lighting) next to a similarly colored but semantically different object, place a stroke on one flower, and run the method with the fixed parameters $w_c = 1$, $w_s = 3$, $t = 0.80$; the central claim is falsified if the stroke fails to reach the second flower, or reaches the similar-colored distractor.","supporting_citations":[{"cited_title":"Unmixing-based soft color segmentation for image manipulation,","cited_arxiv_id":null,"evidence_quote":"Supplies the per-pixel semantic features that are PCA-compressed and fused with RGB to form the 6D feature space; without this network the palette would have no semantic component."},{"cited_title":"Slic superpixels compared to state- of-the-art superpixel methods,","cited_arxiv_id":null,"evidence_quote":"Provides SLIC superpixels whose centroid pixels are used as sampling points for the modified k-means palette extraction."},{"cited_title":"Palette-based photo recoloring.,","cited_arxiv_id":null,"evidence_quote":"Supplies the palette-based recoloring machinery, including the linear system for the RBF similarity coefficients and the color-transfer equation that applies palette changes to pixels."},{"cited_title":"An improved geometric approach for palette-based image decomposition and recoloring,","cited_arxiv_id":null,"evidence_quote":"Compared method; an RGB-space geometric palette approach that lacks semantic information and serves as a baseline the method must beat on content-aware propagation."},{"cited_title":"Edit propagation via color palettes,","cited_arxiv_id":null,"evidence_quote":"Compared method; an edit-propagation-via-color-palettes approach that provides a palette-based baseline without semantics."},{"cited_title":"Palette-based content-aware image recoloring,","cited_arxiv_id":null,"evidence_quote":"Compared method; a content-aware palette recoloring approach that neither adaptively determines palette size nor supports pixel-level editing, the baseline closest in aim."},{"cited_title":"Joint learning of visual and spatial features for edit propagation from a single image,","cited_arxiv_id":null,"evidence_quote":"Compared method; a deep learning edit-propagation approach that represents the dense-stroke requirement and limited distant-pixel propagation the method aims to overcome."},{"cited_title":"LoCoPalettes: Local control for palette-based image editing,","cited_arxiv_id":null,"evidence_quote":"Compared method; a hierarchical semantic-segmentation-based palette editing approach whose segmentation dependence the method contrasts with its own continuous semantic-feature propagation."}],"review_version":1}