{"id":"4969d2fe-6c62-4819-a730-ae25d149036e","arxiv_id":"2412.07203","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"FCNet decouples facial colorization into per-component color codes, enabling controllable reference-based, automatic, and diverse colorization of face images.","lead":"A new system, FCNet, colorizes grayscale face photos by learning separate color codes for lips, skin, eyes, hair, and background. It is designed for old photo restoration and lets users control the color of each facial feature, with or without reference photos.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Quantitative support is provided only for automatic colorization; the claimed favorable performance in single- and multi-reference scenarios is not tested, leaving the central claim unsubstantiated.","rationale":"The reader's verdict is CONDITIONAL, and my concern also points to a condition that must be satisfied before the central claim is accepted: quantitative results for the single- and multi-reference application scenarios are missing. The reader's weakest_assumption focuses on face-parser reliability, which is a real but secondary issue because the experiments are limited to aligned datasets (FFHQ/CelebA-HQ) where a parser is more likely to succeed. The more load-bearing problem is that the paper's headline claim spans all three scenarios, yet only the automatic mode is numerically evaluated. This is a direct evidentiary gap: if the reference-guided modes performed poorly, the central claim would be false, and the current paper provides no numbers to rule this out. The concrete test I propose would settle the issue by running the missing quantitative comparison. I therefore agree with the CONDITIONAL verdict, but my emphasis differs from the reader's weakest_assumption, hence 'partial' agreement. The verdict should remain CONDITIONAL until the missing evaluations are provided.","tokens_in":10685,"tokens_out":4301,"duration_ms":94277,"concrete_test":"Run the official implementations and pre-trained weights of FCNet and the reference-based baselines (WCT2, Gray2ColorNet, TFColor, Unicolor, PDNLA-Net) on the CelebA-HQ test set, using the same reference images as in Fig. 5. Report FID, CF, PSNR, and SSIM for single- and multi-reference colorization; if FCNet does not outperform or match the strongest baseline on at least the primary metric, the 'favorably' claim for those modes is unsupported. Additionally, compute a component-perturbation test: swap only the lips color representation from a different reference and measure the mean absolute error inside vs. outside the lips mask; a large outside change would show the representation is not decoupled.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Table 1 reports FID/CF/PSNR/SSIM only for the automatic colorization mode (Section 'Comparison with Previous Methods'). The single- and multi-reference modes, which are core to the paper's title and to the abstract's claim of 'favorably against existing methods in various application scenarios', are compared only via qualitative figures (Fig. 5 and supplementary). No quantitative comparison against WCT2, Gray2ColorNet, TFColor, Unicolor, or PDNLA-Net is given. Similarly, diverse colorization is only shown visually. Without these numbers, the central claim is an assertion, not a demonstrated result. The paper's own ablation study (Table 2) also does not include a metric that measures per-component color control, so the 'decoupled representation' is supported mainly by anecdotal examples. This is an evidentiary gap, not an internal inconsistency, but it is load-bearing: if the missing evaluations do not beat the baselines, the paper's primary contribution and abstract claim fail. The promised code and pre-trained models are not yet available, so the numbers in Table 1 cannot be independently checked either.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes FCNet, a facial image colorization framework that learns a spatially decoupled color representation for each facial component (lips, skin, eyes, hair, background) under the guidance of face parsing maps. A chromatic and spatial augmentation strategy is introduced to enforce decoupling using only grayscale/color pairs. The trained model is extended to no-reference scenarios by training an automatic encoder g_auto and a normalizing-flow-based diverse colorization module g_flow. Experiments on CelebA-HQ report quantitative results for automatic colorization against ten baselines (Table 1), qualitative comparisons for single-/multi-reference colorization (Fig. 5), and an ablation study (Table 2). The abstract claims favorable performance in no-, single-, and multi-reference scenarios, but quantitative support is provided only for the automatic mode.","tokens_in":10897,"tokens_out":3151,"duration_ms":38229,"significance":"If substantiated, the paper would offer a practical face colorization system with per-component color control and multiple application modes. The automatic colorization experiments are a strength: they use official baseline implementations, a held-out test set, and a common set of metrics. The proposed augmentation-based decoupling is a plausible and interesting mechanism. However, the central claim of 'favorably against existing methods in various application scenarios' is not quantitatively demonstrated for single-reference, multi-reference, or diverse colorization, and the face parser—on which the whole pipeline depends—is not specified or evaluated. The paper also lacks error bars or significance tests, and the promised code and pre-trained models are not yet available, limiting independent verification. The core idea is defensible, but the evidence is uneven.","major_comments":[{"comment":"Table 1 reports FID/CF/PSNR/SSIM only for the automatic colorization mode. The single- and multi-reference modes, which are core to the title and to the abstract's claim of favorable performance 'in various application scenarios', are compared only via qualitative figures (Fig. 5 and supplementary). No quantitative comparison against WCT2, Gray2ColorNet, TFColor, Unicolor, or PDNLA-Net is given. Since these baselines are central to the claimed contribution, the paper must add quantitative results (e.g., FID, CF, and component-wise color accuracy) for reference-guided colorization before the central claim can be accepted.","section":"Comparison with Previous Methods; Fig. 5"},{"comment":"The entire decoupling and the automatic module rely on a face parser P, used as m_ref = P(x_ref) for references and m_l for grayscale inputs. The paper never names the parser, reports its architecture, or evaluates its accuracy on grayscale inputs and on the aligned faces used in the experiments. If the parser mislabels components (especially on grayscale or less-aligned faces), the color representations w and the predicted w from g_auto are wrong by construction. Please specify the parser and provide a quantitative analysis of parsing quality on both color and grayscale inputs, or at least show representative parsing maps and failure cases.","section":"Facial Component Priors; Automatic Colorization"},{"comment":"Table 1 reports a single run for each method without error bars, confidence intervals, or significance tests. For example, the FID gap between Ours (3.60) and DDColor (3.80) is small and could be within run-to-run variation, especially since GAN-based methods are sensitive to seed and initialization. Please report results over multiple seeds (at least three) with mean and standard deviation, or provide a statistical significance test, so the reader can judge whether the reported improvements are meaningful.","section":"Experimental Setting"},{"comment":"The ablation study measures only global metrics (FID, CF, PSNR, SSIM) and does not include any metric that quantifies per-component color control or decoupling strength. The paper's key claim is that the representation is 'spatially decoupled' and enables independent control of each facial component, yet this is supported only by qualitative examples. Please add a quantitative evaluation of controllability, for example by measuring component-wise color transfer accuracy (e.g., color error on each parsing region when a reference is swapped) or by reporting a user study on controllability.","section":"Ablation Study; Table 2"},{"comment":"The diverse colorization mode is introduced as one of the three application paradigms, but it has no quantitative evaluation: no diversity metric (e.g., LPIPS distance, color histogram variance), no comparison to other diverse colorization methods, and no measure of whether the sampled colors match the learned distribution. Also, the training details and loss functions for g_flow and g_auto are deferred entirely to the supplementary material. Please include a quantitative diversity evaluation and state the full training objective in the main text or clearly reference a specific section of the supplementary that provides it.","section":"Diverse Colorization; Extension to No-reference Scenarios"}],"minor_comments":[{"comment":"The email address 'csmliu@oulook.com' appears to contain a typo; it should likely be 'outlook.com'.","section":"Contact information"},{"comment":"The notation is inconsistent: the grayscale input is x_l with one channel, the reference is x_ab with two channels in one place, but later x_ref and m_ref are used. Please define all symbols consistently in one place, especially since the face parser P presumably operates on an RGB or Lab image, while the reference is described as containing only the ab channels.","section":"Notation in Method"},{"comment":"The presentation of Table 2 is confusing: the checkmarks and row labels are not aligned in the text, making it difficult to determine which configuration each row corresponds to. For example, the row labeled 'w/o Data Augmentaion' shows two checkmarks, which appears to contradict its label. Please reformat the table so each row's configuration is unambiguous.","section":"Table 2"},{"comment":"Some references are introduced by nickname (e.g., 'Colorful', 'InstColor', 'DisColor') without giving the paper title at first mention; please provide full citations in Table 1 to help readers identify the exact works.","section":"Related Work and References"},{"comment":"In the qualitative evaluation of Fig. 4, the statement that 'our results demonstrate the minimal degree of color bleeding' is subjective; please support such claims with a quantitative metric (e.g., edge color-bleeding measure or component-wise color consistency) or temper the phrasing.","section":"Qualitative claims"}],"recommendation":"major_revision","confidential_remarks":"The manuscript has a solid automatic-colorization experiment with official baselines, but the paper's broader claims—component-wise decoupling and favorable performance in single-/multi-reference and diverse settings—are not quantitatively supported. The face parser is a critical dependency that is left unspecified. I would be willing to reconsider after the authors add the missing evaluations and parser details. The promised public code would also help independent verification."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The short version: the core idea is real, and the automatic colorization results are the strongest part of the paper. The composite-augmentation trick—augmenting a color image chromatically and spatially, then assembling the ground truth from one component of each augmented version—is a simple and effective way to force per-component latent codes to decouple. The grouped decoder design is a sensible accompaniment. The automatic benchmark uses official implementations of ten baselines on a held-out test set, which is better than the field's norm, and the FID/PSNR gains over DDColor are plausible if not yet convincing without error bars.\n\nThe stress-test note is on target. Table 1 covers automatic colorization only. The single- and multi-reference modes, which are half of the title and central to the abstract's \"various application scenarios\" claim, get Fig. 5 and no quantitative comparison against WCT2, Gray2ColorNet, TFColor, Unicolor, or PDNLA-Net. The diverse mode gets even less. If those missing numbers do not beat the baselines, the paper's primary contribution is substantially weakened. This is an evidentiary gap, not an internal contradiction—the method is coherent and the qualitative figures look competitive—but it is load-bearing.\n\nTwo further soft spots are worth flagging. The face parser is never named, and no parsing accuracy on grayscale inputs is reported; the entire decoupling depends on this module. And many implementation details, including the network architecture and the normalizing flow for diverse colorization, are deferred to a missing supplementary file. No code or pretrained models are linked, so Table 1 cannot be independently checked yet.\n\nNone of this is fatal. The ablation study shows each design choice contributes, the automatic results are strong, and the missing evaluations are straightforward to supply. Who is this for? Researchers working on face-specific colorization or on disentangled color representations for image editing. A serious referee should see it, with the requirement that the revision adds quantitative results for reference-based and diverse modes, names the parser and reports its accuracy, and releases code.\n\nRecommendation: send to peer review. The evidentiary gaps are real but fixable, and the underlying idea deserves engagement.","headline":"A genuinely clever decoupling trick and strong automatic colorization results, but the paper's headline claim spans three modes and only one is quantified.","tokens_in":11406,"tokens_out":2077,"would_cite":true,"duration_ms":26458,"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":"This paper establishes that separating color into per-component codes for lips, skin, eyes, and hair makes facial image colorization more natural and controllable, and that one network then covers no-reference, single-reference, and…","keywords":["facial image colorization","face parsing","decoupled color representation","reference-based colorization","automatic colorization","normalizing flow","spatial and chromatic augmentation","component-wise color control"],"falsifier":"Colorize a fixed grayscale face twice using the same reference except with the lips slot of w taken from a different reference photo (e.g., red vs brown lipstick on the same face layout). If pixels outside the lip mask change between the two results, the representation is not spatially decoupled and the paper's central mechanism is falsified. A second decisive observation: run the face parser on grayscale versions of images and compare its component masks with the masks on the original color images; a systematic drop in lip/eye boundary agreement on grayscale input would predict exactly where the automatic mode fails.","tokens_in":10483,"feed_emoji":"🎨","tokens_out":9874,"duration_ms":90897,"temperature":0.7,"pith_summary":"This paper argues that existing image colorizers produce unnatural, uneven results on faces because they lack an understanding of facial components. It proposes FCNet, which uses a face parsing model to learn a decoupled color representation for each major face region, so the color of the lips cannot contaminate the skin or eyes. A training strategy that augments each color image chromatically and spatially, then assembles a composite ground truth from different augmented versions, forces each part of the representation to control exactly one component. After training, the same pipeline colorizes from a single reference image, mixes colors from several references, or runs without any reference by predicting codes from the grayscale input or sampling them from a learned distribution. The paper reports lower FID and higher PSNR/SSIM than prior methods on an aligned face test set, with more controllable and visually natural results.","feed_headline":"Face colorizer splits colors by lips, skin, eyes, and hair","feed_subtitle":"One trained network handles no-reference, single-reference, and multi-reference facial colorization.","key_machinery":"The load-bearing mechanism is the decoupled component color representation w, a low-dimensional latent code with one slot per facial component (lips, skin, eyes, hair, background) plus the training objective that keeps the slots independent. A pre-trained face parser provides the masks that define the components; chromatic and spatial augmentation creates a composite supervision signal in which each component of the target color comes from a different augmented copy of the same image, so the encoder is rewarded only when each code slot controls its own region. The colorization network f expands w through a grouped decoder and uses an affine transform (multiplication and addition) to inject color into grayscale features.","core_discovery":"The central claim is that the unnatural color bleeding in face colorization comes from entangled color information across facial components, and that this entanglement can be broken by structured supervision. FCNet extracts a grouped color code w = [w_lips, w_skin, w_eyes, w_hair, w_background] from reference images; during training the ground truth image is assembled from five differently augmented versions so that each w_i is the only source of color for its own component. Because the encoder is shared across augmented versions, it cannot fall back on spatial shortcuts and must learn which code entries map to which facial regions. The same w then feeds a colorization network whose decoder uses grouped convolutions and affine modulation, giving component-level control in single- and multi-reference modes and enabling two no-reference modes via an encoder that predicts w and a normalizing flow that samples it.","pith_inferences":["Editorial extension: the same decoupling recipe—parse into parts, assign each part a latent color slot, supervise with cross-augmented composites—should transfer to other objects with stable component colors, such as clothing, cars, or animals, where color bleeding is also a visible artifact.","Editorial extension: the automatic mode's ceiling is set by the face parser's accuracy on grayscale, unaligned, or occluded faces; the paper does not report parsing metrics there, so the practical gain over prior methods is likely to shrink on such inputs.","Editorial extension: the per-component codes look like editable palettes, so one could build a 'lipstick preview' tool that fixes the skin and hair slots and only varies the lips slot, which the paper demonstrates qualitatively but does not develop into an interface.","Editorial extension: a direct test of decoupling would be to swap the lips slot of w between two identities and verify visually and in pixel terms that only the lips change; reporting such an experiment would make the controllability claim quantitative."],"forward_implications":["A user can recolor only the lips, only the skin, or only the eyes of a grayscale face by editing or replacing the corresponding slot in w, without retraining.","The same trained colorization network serves single-reference, multi-reference, and no-reference modes, so a face colorizer can be deployed without collecting reference images.","Automatic mode requires only the grayscale photo plus the parser's masks, making black-and-white photo restoration a one-pass procedure.","Color diversity becomes a sampling operation on a learned distribution, so multiple plausible colorizations of the same face are produced at inference time.","Because the color representation branch is the only module that changes between modes, the paper's architecture lets the branch be replaced with a better encoder or flow without retraining the colorization network."],"supporting_citations":[{"why":"Supplies the FFHQ dataset used to train the colorization network and its submodules.","marker":"(Karras, Laine, and Aila 2019)"},{"why":"Supplies the CelebA-HQ dataset used to evaluate all methods in the quantitative comparison.","marker":"(Karras et al. 2017)"},{"why":"Provides the FID metric and the adversarial loss used in the learning objective.","marker":"(Heusel et al. 2017)"},{"why":"Provides the perceptual loss term in the learning objective.","marker":"(Johnson, Alahi, and Fei-Fei 2016)"},{"why":"Provides the cycle-consistency loss term in the learning objective.","marker":"(Zhu et al. 2017)"},{"why":"Supplies the normalizing flow architecture used to sample diverse color representations in the no-reference diverse mode.","marker":"(Kingma and Dhariwal 2018)"},{"why":"A comparison baseline (Colorful) that the automatic colorization results are tested against.","marker":"(Zhang, Isola, and Efros 2016)"},{"why":"A comparison baseline (InstColor) that the automatic colorization results are tested against.","marker":"(Su, Chu, and Huang 2020)"},{"why":"A comparison baseline (DDColor) that the automatic colorization results are tested against.","marker":"(Kang et al. 2023)"},{"why":"A comparison baseline (L-CAD) that the automatic colorization results are tested against.","marker":"(Weng et al. 2024)"}],"fun_headline_variants":["Split face colors by component: lips, skin, eyes, hair","Decoupled color codes give natural facial colorization","FCNet separates color per face part to stop uneven tones","Per-region color codes fix unnatural facial colorization"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The entire pipeline assumes that the pre-trained face parser—the model that labels which pixels are lips, skin, eyes, and hair—works reliably on both color and grayscale faces; if it mislabels those regions, the decoupled color codes land on the wrong pixels and the quality collapses.","fun_headline_variants_meta":{"raw":{"variants":["Split face colors by component: lips, skin, eyes, hair","Decoupled color codes give natural facial colorization","FCNet separates color per face part to stop uneven tones","Per-region color codes fix unnatural facial colorization"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000862,"raw_usage":{"total_tokens":3738,"prompt_tokens":945,"completion_tokens":2793,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":561,"completion_tokens_details":{"reasoning_tokens":2727}},"tokens_in":561,"tokens_out":2793,"duration_ms":20025,"temperature":1.0,"reasoning_tokens":2727,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T19:00:34.747484+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Colorize a fixed grayscale face twice using the same reference except with the lips slot of w taken from a different reference photo (e.g., red vs brown lipstick on the same face layout). If pixels outside the lip mask change between the two results, the representation is not spatially decoupled and the paper's central mechanism is falsified. A second decisive observation: run the face parser on grayscale versions of images and compare its component masks with the masks on the original color images; a systematic drop in lip/eye boundary agreement on grayscale input would predict exactly where the automatic mode fails.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"A comparison baseline (Colorful) that the automatic colorization results are tested against."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"A comparison baseline (DDColor) that the automatic colorization results are tested against."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"A comparison baseline (L-CAD) that the automatic colorization results are tested against."}],"review_version":1}