{"id":"941ac623-6a43-48fd-94d0-ad4cc9680cba","arxiv_id":"2607.08185","paper_version":1,"verdict":"ACCEPT","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"low","formal_verification":"none","parameter_count":4,"one_line_summary":"NamedCurves+ conditions Bezier tone curves on color-naming probability maps and fuses them with a transposed-attention transformer, yielding SOTA interpretable enhancement on MIT-5K, PPR10K, MSEC and SICE.","lead":"NamedCurves+ enhances photos by learning Bezier tone curves for each named color (red, green, blue, etc.) after a standardization backbone, then fuses the results with a transformer. It beats prior methods on retouching, tone mapping and exposure correction while letting users edit the curves by hand.","discovery_kind":"extension","skeptic_critique":{"model":"grok-4.5","headline":"No significant objection identified","rationale":"The reader correctly identifies the fixed color-naming model as the softest modeling choice, yet the paper already quantifies its impact (Table VII) and shows that the rest of the pipeline remains robust. Gains are modest but consistent, ablations isolate each component, and the interactivity claim is demonstrated rather than merely asserted. No stronger load-bearing flaw (e.g., data leakage, metric gaming, or architectural circularity) appears. Therefore the ACCEPT verdict stands; the only remaining action is the post-release verification above.","tokens_in":20515,"tokens_out":420,"duration_ms":4178,"concrete_test":"Once the promised code is released, re-train and re-evaluate the full NamedCurves+ pipeline on the official MIT-Adobe-5K Expert-C split (4500/500) with the exact hyper-parameters of §IV-A; if the reported PSNR of 25.75 falls by more than 0.15 dB or the ranking versus BGLUT reverses, the headline quantitative claim would need revision.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper’s central claim—that NamedCurves+ outperforms SOTA on retouching, tone mapping and exposure correction while remaining explainable/interactive via per-color-name Bezier curves—is supported by consistent gains across four public datasets (Tables I–IV), module ablations (Table V), color-naming variants (Table VII), backbone swaps (Table VI), a 2AFC user study, and qualitative examples of curve editing. The reader’s weakest assumption (fixed Van de Weijer maps + ad-hoc 6-way grouping) is real but already stress-tested: grouping 11\to6 improves metrics, random GMM/RGB-cube partitions underperform, and the Limitations section acknowledges boundary leakage. No internal inconsistency, circular derivation, or untested leap undermines the empirical claim. Code is promised but not yet public—the only material reproducibility caveat.","agreement_with_reader":"agree"},"referee_report":{"model":"grok-4.5","summary":"NamedCurves+ is a supervised image-enhancement framework that first standardizes an input with a UNet-like backbone, decomposes the result into six color-name probability maps (Van de Weijer et al. with an ad-hoc grouping of the classic 11 terms), estimates a set of Bezier-parametrized global tone curves conditioned on each map, and fuses the six globally adjusted images with a transposed-attention transformer block. The design is intended to be both competitive with SOTA and interactive: users can edit the learned control points of individual color-name curves at inference time without retraining. The method is evaluated on image retouching (MIT-Adobe-5K, PPR10K), tone mapping (MIT-Adobe-5K XYZ\to sRGB), and exposure correction (MSEC, SICE), with systematic ablations of backbone, naming scheme, fusion module, control-point count and loss weight α, plus a small 2AFC user study.","tokens_in":20782,"tokens_out":844,"duration_ms":8995,"significance":"If the reported gains hold, the paper supplies a practical middle ground between black-box image-to-image networks and classical parametric editors: competitive quantitative performance on four standard benchmarks together with an explicit, user-editable representation (per-color-name Bezier curves) that matches the color-naming vocabulary already used in commercial software. The extension from the conference NamedCurves paper—transformer fusion that reduces halo artifacts, multi-task evaluation, and clearer interactivity demos—is incremental but useful. Strengths that should be credited include the breadth of the experimental suite (Tables I–IV), the module and naming ablations (Tables V–VII), the controlled 2AFC study, and the explicit Limitations section that already flags boundary leakage of the fixed color-naming prior.","major_comments":[{"comment":"The central empirical claim is adequately supported; no load-bearing technical error was found. The only material caveat is reproducibility: source code and models are promised (Abstract) but not yet released, so independent verification of the reported numbers and of the interactive editing interface is currently impossible. This should be remedied before final acceptance.","section":null}],"minor_comments":[{"comment":"Tables I–IV report point estimates only; even a single-run standard deviation or a note that results are from one seed would strengthen confidence.","section":null},{"comment":"Inference-time figures (Table I, Table VIII) are given for a single hardware pair; a short note on batch size and whether the color-naming step is included would help readers compare fairly with LUT-based methods.","section":null},{"comment":"Section III-B and Figure 5 motivate the 11\to6 grouping by intensity-dependent hue overlap; a one-sentence quantitative check (e.g., average probability mass retained after grouping) would make the design choice more transparent.","section":null},{"comment":"Figure 2 caption and the surrounding text correctly highlight halo reduction, but the visual difference is subtle; a zoomed inset or residual map would make the improvement clearer.","section":null},{"comment":"A few typographical issues remain (e.g., “F orum” in the references, inconsistent spacing around “ΔE”). A final proof-reading pass is recommended.","section":null}],"recommendation":"minor_revision","confidential_remarks":"The manuscript is a solid journal extension of a recent ECCV paper. Novelty relative to the conference version is real but modest (mainly the fusion redesign and multi-task results). Fit for a top-tier journal is acceptable given the breadth of evaluation and the practical interactivity angle; I would not block on novelty alone. Code release is the only item I would treat as a hard condition."},"author_rebuttal":null,"desk_editor":{"model":"grok-4.5","letter":"This is a clean, well-executed extension of the authors’ conference paper. The new pieces that matter are the transposed-attention fusion block (which kills the halo artifacts of the old weighted-average design), the expansion to tone mapping and exposure correction, and the demonstration that users can actually edit the per-color-name Bezier curves at inference time without retraining. Those are real, measurable improvements, not just packaging.\n\nWhat the paper does well is the empirical hygiene. Four public benchmarks, systematic ablations on backbone, naming scheme, fusion, control-point count and loss weight, a properly controlled 2AFC user study, and qualitative examples that match the claims. The color-naming ablation is especially useful: Van de Weijer + the 6-way grouping beats both the 11-name version and random/RGB-cube partitions, so the semantic prior is doing real work rather than just adding channels. Metrics move in the right direction across the board (MIT-5K 25.75 dB, PPR10K, MSEC/SICE), and the inference-time numbers are honest.\n\nSoft spots are minor and already flagged by the authors. The fixed Van de Weijer maps plus ad-hoc grouping can leak at chromatic boundaries or under weird illuminants; the Limitations section owns this. No error bars, code still only promised, and the free parameters (α, N, C) are tuned rather than derived. None of that undermines the central claim. The math is just standard Bezier + transposed attention; nothing circular or load-bearing-wrong.\n\nThis is for people who care about practical, controllable photo enhancement and low-level vision systems. It will not rewrite the field, but it is a better engineering answer than most of the LUT and curve papers it beats. I would send it to referees without hesitation; the evidence is already strong enough that the review conversation will be about polish and generality, not about whether the result is real. Worth reading if you work in this area; worth citing if you need a strong, interactive baseline.","headline":"Solid systems extension of the authors’ own NamedCurves work: real SOTA gains on three tasks plus genuine interactivity, with no load-bearing flaws.","tokens_in":21308,"tokens_out":502,"would_cite":true,"duration_ms":6396,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"grok-4.5","headline":"Color-name tone curves make learned image enhancement both better and editable by hand.","keywords":["image enhancement","color naming","tone curves","image retouching","tone mapping","exposure correction","user interaction","transformer fusion"],"falsifier":"Replace the color-naming stage with random RGB partitions or pure intensity curves on the same training data; if the quantitative gains and the absence of boundary halos disappear, the color-naming claim fails.","tokens_in":21443,"feed_emoji":"🎨","tokens_out":544,"duration_ms":9433,"temperature":0.7,"pith_summary":"Most deep image enhancers copy an expert's style but leave users with an opaque black box they cannot tweak. NamedCurves+ claims the remedy is to first standardize the photo, then split it into six universal color-name maps (red, green, blue, orange-brown-yellow, pink-purple, achromatic), learn a smooth Bezier tone curve for each map, and fuse the six globally adjusted versions with a lightweight transformer that restores local context. Because every edit is still a familiar color curve, the model is both stronger than prior LUT- and curve-based systems on retouching, tone-mapping and exposure-correction benchmarks and open to live user adjustment: drag a curve and only the matching color regions change, without retraining. The paper therefore argues that color naming is the missing inductive bias that reconciles automatic quality with human control.","feed_headline":"Color-name curves beat black-box enhancers and stay editable","feed_subtitle":"Six familiar tone curves plus a transformer top retouching, tone-mapping and exposure scores.","key_machinery":"NamedCurves+: a pipeline that maps an input through a standardization backbone, six color-name probability maps, per-name Bezier tone curves, and a single multi-head transposed-attention fusion block.","core_discovery":"Conditioning learned Bezier tone curves on a fixed, perceptually grounded color-naming decomposition, then fusing the resulting globally adjusted images with a transposed-attention transformer, yields state-of-the-art enhancement while remaining fully interpretable and user-editable.","pith_inferences":[],"forward_implications":[],"fun_headline_variants":["Color-name tone curves beat black-box enhancers yet stay fully editable","NamedCurves+ pairs named-color Bezier curves with transformer for SOTA retouch","Interpretable enhancement via color-name curves plus spatial transformer fusion","Six familiar color-name curves deliver editable top scores on retouch and tone map","User-editable image enhancement from curves conditioned on fixed color names"],"cache_read_input_tokens":16512,"weakest_assumption_plain":"The fixed color-naming model (and the authors' grouping of eleven names into six) must produce stable probability maps even under unusual cameras and lighting, or the later curves and fusion will mis-edit the image.","fun_headline_variants_meta":{"raw":{"variants":["Color-name tone curves beat black-box enhancers yet stay fully editable","NamedCurves+ pairs named-color Bezier curves with transformer for SOTA retouch","Interpretable enhancement via color-name curves plus spatial transformer fusion","Six familiar color-name curves deliver editable top scores on retouch and tone map","User-editable image enhancement from curves conditioned on fixed color names"]},"model":"grok-4.5","effort":"low","cost_usd":0.005708,"raw_usage":{"total_tokens":1514,"prompt_tokens":748,"num_sources_used":0,"completion_tokens":101,"cost_in_usd_ticks":57080000,"prompt_tokens_details":{"text_tokens":748,"audio_tokens":0,"image_tokens":0,"cached_tokens":256},"completion_tokens_details":{"audio_tokens":0,"reasoning_tokens":665,"accepted_prediction_tokens":0,"rejected_prediction_tokens":0}},"tokens_in":748,"tokens_out":101,"duration_ms":7110,"temperature":1.0,"reasoning_tokens":665,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-07-10T11:41:21.673168+00:00","model_set":{"reader":"grok-4.5"},"falsifier":"Replace the color-naming stage with random RGB partitions or pure intensity curves on the same training data; if the quantitative gains and the absence of boundary halos disappear, the color-naming claim fails.","supporting_citations":[],"review_version":1}