{"id":"75fee1b2-9ebd-4385-89c3-97fd094ff4af","arxiv_id":"2608.10798","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":1,"one_line_summary":"A luminance-agnostic colorizer fine-tuned on standard and red-insensitive grayscale inputs stays competitive on normal photos and handles orthochromatic photos better than fixed-luminance baselines.","lead":"This paper introduces a colorization method that does not lock the brightness of the input, so it can genuinely change how light or dark a region looks. It is aimed at restoring old red-insensitive photographs, where standard methods cannot brighten dark red objects.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Ortho robustness is shown only on the same (B+G)/2 distribution used in training; the only real-archive evidence is qualitative, so external validity is the key unverified condition.","rationale":"The paper's central claim is conditional on the orthochromatic simulation being representative enough that gains on simulated ortho inputs transfer to real red-insensitive archives. That condition is least secure not merely because Eq. (2) is an approximation, but because the quantitative evaluation uses the exact same transform to generate both the training supervision and the test inputs, while real-archive evidence is only qualitative. The Limitations section explicitly concedes the simulation is a simplified approximation and lists blue-biased sensitivity, fading, and chemical degradation as unmodeled factors. A concrete way to settle this is to evaluate on real glass plates with color ground truth, or to generate held-out test sets from measured spectral sensitivity curves distinct from Eq. (2). If the robustness margin persists, the central claim stands; if it collapses, Eq. (2) is load-bearing and the contribution becomes robustness to a synthetic blue-green grayscale rather than to orthochromatic imagery. This concern does not attack the fixed-luminance analysis, which is sound; it targets external validity. The reader identified the same assumption, and I agree. The CONDITIONAL verdict remains appropriate, so no change to the reader's verdict is needed.","tokens_in":11166,"tokens_out":6113,"duration_ms":58746,"concrete_test":"Evaluate Ours and all baselines on a held-out set of real orthochromatic glass-plate photographs with paired color references or expert colorized ground truth, using the same FID, sFID, FID-DINO, ColorNet, and human artifact-free protocols. If the ortho margin over baselines disappears, the claim should be narrowed to the specific (B+G)/2 formation. A cheaper computational version is to generate a held-out test set from several measured ortho spectral sensitivity curves (e.g., Kodak blue-green and blue-only responses) rather than Eq. (2), with no retraining, and check whether Ours still beats the best baseline by a significant margin.","verdict_should_be":"UNCHANGED","load_bearing_attack":"For the central claim to hold, the method must be robust to orthochromatic inputs in general, not only to the synthetic grayscale formation used during training. The quantitative support (Tables 1-3) is generated from Eq. (2), I_ortho_gray = (I_B + I_G)/2, which is exactly the transform used to create the ortho training distribution with the prompt c_ortho. This creates a train/test distributional coupling: the model is trained to invert this specific blue-green average, while all baselines are evaluated on a formation they never saw. The paper's only real-archive evaluation (Figure 4) is qualitative, with no FID, color metrics, or artifact-free rates, and the human study in Section 4.4 was run on simulated inputs, not real glass plates. The Limitations section explicitly concedes: 'our orthochromatic simulation is a simplified approximation of historical film response, while real archival imagery exhibits diverse capture and aging effects (e.g., blue-biased spectral sensitivity, color fading, and chemical degradation).' Thus the central claim of 'substantially more robust on orthochromatic inputs' is currently supported only on a self-generated distribution. If real orthochromatic stocks respond differently from (B+G)/2, the robustness may not transfer, and the fixed-luminance bottleneck story would not be validated on actual archives.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a luminance-agnostic image colorization framework built on a foundation image-editing model (FLUX.2-klein). Instead of predicting chroma under a fixed luminance channel, the method fine-tunes the model with LoRA to generate full RGB output directly from grayscale input, conditioned on a text prompt. To handle both modern panchromatic and historical orthochromatic imagery, the authors introduce a mixed grayscale objective that trains on two synthetic grayscale formations: standard luminance (Eq. 1) and a simplified orthochromatic formation (Eq. 2), I_ortho_gray = (I_B + I_G)/2. Experiments on COCO, ImageNet, and a multi-instance benchmark report FID-family and colorfulness metrics for pan and ortho settings, structural-fidelity checks via grayscale reprojection, a small human study on artifacts, and qualitative results on real historical glass-plate photographs. The central claim is that the method is competitive on standard grayscale inputs and substantially more robust on orthochromatic inputs.","tokens_in":11542,"tokens_out":3943,"duration_ms":35006,"significance":"If substantiated, the paper addresses a genuine limitation of the fixed-luminance paradigm in colorization: such methods cannot brighten dark regions when the input grayscale deviates from natural luminance, which is exactly the situation for red-insensitive orthochromatic film. The proposed full-RGB editing formulation is a sensible alternative, and the mixed-objective training idea is a reasonable way to expose the model to multiple grayscale formations. The paper's strengths include a clear articulation of the bottleneck, a concrete training recipe with publicly described configuration, structural-fidelity verification that fine-tuning preserves input structure, and the inclusion of real archival images as qualitative evidence. However, the quantitative support for the central orthochromatic-robustness claim is currently limited to the same synthetic transform used in training, and the contribution of the mixed objective is not ablated on the color metrics that drive the headline results. The human study is small and lacks statistical analysis.","major_comments":[{"comment":"The quantitative orthochromatic evaluation uses exactly the same synthetic grayscale formation, I_ortho_gray = (I_B + I_G)/2, that was used to construct the training distribution. This creates a train/test distributional coupling: the model is trained to invert this specific blue-green average, while all baselines are evaluated on a formation they never saw. The real-archive evaluation in Figure 4 is qualitative only, with no FID, color metrics, or artifact-free rates. The Limitations section explicitly concedes that the simulation is a simplified approximation of historical film response. As a result, the central claim of 'substantially more robust on orthochromatic inputs' is currently supported only on a self-generated distribution. I recommend either adding quantitative evaluation on real orthochromatic photographs (or on a diverse set of physically motivated spectral-sensitivity models), or explicitly scoping the claim to robustness under the simulated (B+G)/2 formation rather than to historical orthochromatic imagery in general.","section":"Section 3, Eq. (2), Section 4.3, Tables 1-3"},{"comment":"The Mixed Grayscale Objective is the paper's claimed core contribution, but it is never ablated on the color metrics that appear in Tables 1-3. Tables 4-6 compare Not-FT, Normal-FT, and Mixed-FT only on grayscale reprojection fidelity (L1/L2/SSIM), and on those structural metrics Mixed-FT is not consistently better than Normal-FT (e.g., Table 4, ortho SSIM: Mixed-FT 0.811 vs Normal-FT 0.828; Table 5, ortho SSIM: 0.842 vs 0.857). Without reporting FID, sFID, FID-DINO, Colorfulness, Col-diverse, Saturation, or ColorNet for the Normal-FT variant, it is impossible to attribute the ortho robustness gains in Tables 1-3 to the mixed objective rather than to the full-RGB fine-tuning alone. Please add the Normal-FT model to the ortho rows of Tables 1-3 (or provide a dedicated ablation table with these color metrics).","section":"Section 4.3, Tables 4-6"},{"comment":"The human study is based on only 30 inputs per method (5 images × 3 datasets × 2 grayscale settings) and 25 participants, and it reports only a binary artifact-free rate without any statistical test or confidence interval. The claim that 'our method attains the highest Artifact-Free Rate' is therefore not established beyond chance. Given that this evidence is cited in the abstract and conclusion, please either add a statistical analysis (e.g., bootstrap confidence intervals, a mixed-effects model with participant as a random effect) or soften the wording to indicate an exploratory result.","section":"Section 4.4, Eq. (8)"}],"minor_comments":[{"comment":"In the text, ImageNet is described as 'strongest overall in the ortho setting,' but on COCO the method does not achieve the best FID or sFID; please make the per-benchmark descriptions precise to avoid over-generalization.","section":"Section 4.3, Tables 1-3"},{"comment":"There is a typo in the HistoryNet sentence: 'andalargeold-moviedataset' should read 'and a large old-movie dataset.'","section":"Section 2.1"},{"comment":"The Artifact-Free Rate equation uses N for total valid responses, but the preceding text says '30 evaluation inputs per method'; please clarify whether N is the number of images or the number of participant responses.","section":"Section 4.4"},{"comment":"The notation for prompts, c_pan and c_ortho, is clear, but it is not stated explicitly how the caption component is obtained for the Multi-Instance dataset versus the other datasets; a sentence clarifying the captioning pipeline would improve reproducibility.","section":"Section 3"},{"comment":"Please define 'Normal-FT' and 'Mixed-FT' unambiguously in a caption or in the text: specifically, that Normal-FT is fine-tuned only on panchromatic grayscale with the pan prompt, while Mixed-FT is fine-tuned on both modes.","section":"Tables 4-6"}],"recommendation":"major_revision","confidential_remarks":"The paper's core direction is sound and the fixed-luminance bottleneck is a real issue. The main risk is external validity: without quantitative evidence on real or diverse simulated orthochromatic inputs, and without an ablation of the mixed objective on color metrics, the headline claim is not fully supported. These concerns are addressable with additional experiments, so I recommend major revision rather than rejection. The novelty relative to existing diffusion-based colorization work is incremental, but the orthochromatic focus is a useful angle for this venue."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Dear colleague,\n\nRead the IITH colorization paper. In one line: it identifies a real bottleneck—Lab colorizers freeze L, which fails when grayscale formation deviates from natural luminance, as in orthochromatic film—and attacks it with a sensible new recipe: fine-tune FLUX.2-klein as a full-RGB image editor with a mixed panchromatic/orthochromatic grayscale objective. That combination is genuinely new in the cited literature, and the fixed-L critique is well argued.\n\nThe paper earns credit on several counts. The math is straightforward; the objective is standard rectified flow, no fitted constants or self-referential metrics. They also checked the obvious failure mode for a generative full-RGB model: structural fidelity via grayscale reprojection. Mixed-FT does not collapse relative to Normal-FT in that check, which is reassuring. The real glass-plate examples in Figure 4 show outputs that visibly lift red-darkened regions, exactly what the fixed-L baselines cannot do. So the direction is right and the qualitative evidence is suggestive, not empty.\n\nNow the soft spots, in proportion. The main one is external validity of the ortho claim. Equation (2), I_ortho_gray = (I_B + I_G)/2, is the exact transform used to generate the training distribution, and all quantitative ortho results in Tables 1–3 are evaluated on that same synthetic distribution. The paper itself concedes in the Limitations section that this is a simplified approximation of historical film response, and real archival imagery has diverse spectral sensitivity, fading, and degradation. So \"substantially more robust on orthochromatic inputs\" is solid only for the simulated distribution; the real-archive evidence is qualitative only, with no FID, color metrics, or artifact rates. That is the load-bearing gap, and I think the stress-test note gets it right.\n\nSecond, the mixed-objective contribution is not actually ablated on color metrics. Normal-FT appears only in the structural fidelity tables, so you cannot tell whether the ortho color gains come from the mixed grayscale objective or just from fine-tuning an image-editing model on any grayscale inputs. That is a fixable experiment, but it needs doing. Third, the human study is a small convenience sample—30 inputs, 25 participants, a binary question, no statistics—and it runs on simulated inputs rather than real plates. Minor relative to the first issue, but worth reporting as minor. No code or weights are released, so the empirical claims are not independently reproducible.\n\nWho gets value: someone working on historical photo restoration or on adapting foundation editing models to grayscale. This is an engineering contribution with a new recipe, not a change to a major branch of science. It deserves a serious referee. I would send it out, expect major revision, and ask for an ortho evaluation on real archives with at least a small quantitative or artifact-rate study, an ablation of the mixed objective on color metrics, and ideally code and weights. The authors are honest about their limitations, and the direction is worth engaging.","headline":"A credible, well-motivated recipe for luminance-agnostic colorization, with one load-bearing gap: the claimed ortho robustness is only demonstrated on the same synthetic grayscale transform used in training.","tokens_in":11924,"tokens_out":1976,"would_cite":false,"duration_ms":20649,"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":"The paper's central claim is that colorization should predict full RGB output from grayscale input rather than preserve an input-derived luminance channel, because fixed luminance cannot brighten the dark red regions produced by…","keywords":["image colorization","orthochromatic photography","luminance-agnostic colorization","fixed-luminance bottleneck","diffusion image editing","mixed grayscale objective","historical photo restoration","rectified flow"],"falsifier":"Photograph a color chart containing saturated red, green, and blue patches on real orthochromatic glass plates, scan the plates, and run the model on both the scans and on synthetic $(I_B+I_G)/2$ versions of the same chart. If the model brightens and correctly colorizes the dark red patch in the synthetic case but fails on the real plate, then the simplified simulation, not the luminance-agnostic architecture, is producing the reported gain.","tokens_in":10970,"feed_emoji":"🎨","tokens_out":9145,"duration_ms":101854,"temperature":0.7,"pith_summary":"Most colorizers operate in Lab space: they keep the input grayscale as the L channel and predict only the a and b chroma values. This paper argues that this fixed-luminance design is a structural bottleneck, because it makes the model mathematically incapable of brightening dark regions, and it becomes especially problematic when the grayscale was formed by orthochromatic film, which suppresses red. The proposed remedy is a luminance-agnostic formulation that treats colorization as full-RGB image editing with a foundation image-editing model, trained with a mixed grayscale objective: each training image appears both as standard luminance grayscale and as a red-insensitive simulation $(I_B+I_G)/2$, with distinct text prompts. The claimed payoff, supported by experiments on COCO, ImageNet, and a multi-instance benchmark plus a human study, is that the model stays competitive on ordinary grayscale inputs while brightening and recoloring orthochromatic inputs more naturally than fixed-luminance baselines.","feed_headline":"Full-RGB colorization lifts brightness that fixed-luminance models cannot","feed_subtitle":"Fine-tuned on standard and red-insensitive grayscale, it stays competitive on both and rescues dark archive images.","key_machinery":"The load-bearing machinery is the Mixed Grayscale Objective combined with a full-RGB diffusion backbone. For each ground-truth image, the paper constructs two conditioning images, standard panchromatic grayscale $I^{pan}_{gray}=0.299I_R+0.587I_G+0.114I_B$ with prompt 'colorize', and simulated orthochromatic grayscale $I^{ortho}_{gray}=(I_B+I_G)/2$ with prompt 'colorize ortho'. The backbone is a rectified-flow diffusion transformer (the 4-billion-parameter FLUX.2-klein model) fine-tuned with LoRA on the DiT blocks only, with a frozen VAE and text encoder; it is trained with the velocity-matching loss $L=\\mathbb{E}_{z_0,z_1,t}\\|v_\\theta(z_t,t,c_{text},I_{gray})-(z_1-z_0)\\|_2^2$. This setup does the work: the text prompt tells the model which grayscale formation produced the input, and because the output is full RGB rather than chroma conditioned on fixed L, the model can change brightness as well as hue. LoRA keeps trainable parameters at roughly 23.1 million, about 0.5% of the backbone.","core_discovery":"The central discovery is that the failure mode of existing colorizers on historical photography is not a data problem but a formulation problem. Because the luminance channel is treated as immutable ground truth, a Lab-based model cannot assign a bright color to a dark input pixel; and orthochromatic film, being insensitive to red, produces grayscale values that are darker than human-perceived luminance exactly where the scene was red. The paper's luminance-agnostic framework removes that constraint by having a diffusion image-editing model regress the entire RGB image, conditioned on the grayscale input and a text prompt. Trained on the mixed objective, the model learns to reinterpret the same luminance values differently depending on whether the prompt marks the input as panchromatic or orthochromatic, which lets it lift dark red regions. The paper reports that on simulated orthochromatic inputs the method leads most metrics on ImageNet and is best or second-best on COCO and Multi-Instance, while on standard inputs it remains competitive with the strongest baselines; a human evaluation records fewer artifact-flagged outputs, and grayscale-reprojection checks show structure is preserved.","pith_inferences":["Beyond the paper, the mixed grayscale objective is effectively a domain-randomization strategy over grayscale formations; a straightforward extension would train on a continuum of spectral-sensitivity curves rather than two fixed modes, which the paper's limitations section already hints at.","Beyond the paper, because the method follows multi-color text prompts correctly (e.g., assigning blue, yellow, and cyan to three cups), full-RGB colorization could become a controllable editing tool where a user specifies object-level colors, not just a global palette.","Beyond the paper, the fixed-luminance argument implies a diagnostic that can be applied to any colorizer: take a dark grayscale patch that should be a bright red or yellow surface and check whether the output's RGB luminance can exceed the input's; for Lab-based models this is impossible.","Beyond the paper, if the simulation $(I_B+I_G)/2$ is representative, the same training recipe should transfer to other non-standard inputs such as infrared or channel-drop imagery, though the paper does not test this."],"forward_implications":["On standard panchromatic grayscale inputs, abandoning fixed luminance does not cost reconstruction quality: the paper reports FID-family scores that are competitive with strong baselines and often second-best.","On orthochromatic inputs, fixed-luminance baselines degrade sharply while the proposed method stays close to its panchromatic performance, so the approach targets the actual cause of red-region failures rather than patching colors.","Because the model outputs full RGB, it can brighten dark red regions in scanned historical glass-plate negatives, something Lab-based methods cannot do by construction.","The human study indicates fewer visible color artifacts—bleeding, unnatural hue shifts, inconsistent colorization—under both input types, so the full-RGB formulation does not come at the cost of perceptual cleanliness.","Grayscale-reprojection checks show the fine-tuned model preserves input structure (SSIM around 0.81–0.88 versus 0.36 for the non-fine-tuned backbone), suggesting the generative prior is not hallucinating content."],"supporting_citations":[{"why":"Introduces FLUX.2-klein, the foundation image-editing model that the method fine-tunes with LoRA.","marker":"[24]"},{"why":"Defines the CIE L*a*b* color space, the fixed-luminance formulation the paper identifies as the bottleneck.","marker":"[8]"},{"why":"Models historical negative film types, motivating the simplified orthochromatic grayscale simulation in Eq. (2).","marker":"[30]"},{"why":"Established the Lab-space chroma-prediction paradigm with fixed luminance that the paper argues against.","marker":"[43]"},{"why":"Provides ImageNet, used both as the training-data source and as an evaluation benchmark.","marker":"[33]"},{"why":"Presents DDColor, the baseline that typically achieves the best FID-family metrics on standard inputs and must be matched.","marker":"[19]"},{"why":"Supplies ColorNet, the learned perceptual colorfulness metric used in quantitative evaluation.","marker":"[41]"},{"why":"Generates the captions used as text prompts for ImageNet and COCO experiments.","marker":"[26]"},{"why":"Provides the COCO evaluation split used for comparisons.","marker":"[4]"},{"why":"Provides the multi-instance benchmark with human-annotated captions used in evaluation.","marker":"[6]"}],"fun_headline_variants":["Luminance-agnostic colorization fixes orthochromatic photo darkness","Full-RGB editing restores dark red areas in historic photos","Mixed grayscale training beats fixed-luminance on orthochromatic input","Diffusion model reinterprets luminance to revive red-blind photos","Luminance-agnostic full-RGB colorization rescues orthochromatic images"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that averaging the blue and green channels, $(I_B+I_G)/2$, faithfully represents how real orthochromatic film darkened red regions; real film has varied spectral sensitivity, blue bias, fading, and chemical aging, so if the simulation is unrepresentative, the robustness results may not transfer to actual archives.","fun_headline_variants_meta":{"raw":{"variants":["Luminance-agnostic colorization fixes orthochromatic photo darkness","Full-RGB editing restores dark red areas in historic photos","Mixed grayscale training beats fixed-luminance on orthochromatic input","Diffusion model reinterprets luminance to revive red-blind photos","Luminance-agnostic full-RGB colorization rescues orthochromatic images"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001405,"raw_usage":{"total_tokens":5665,"prompt_tokens":921,"completion_tokens":4744,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":537,"completion_tokens_details":{"reasoning_tokens":4651}},"tokens_in":537,"tokens_out":4744,"duration_ms":29078,"temperature":1.0,"reasoning_tokens":4651,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T14:12:51.763882+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Photograph a color chart containing saturated red, green, and blue patches on real orthochromatic glass plates, scan the plates, and run the model on both the scans and on synthetic $(I_B+I_G)/2$ versions of the same chart. If the model brightens and correctly colorizes the dark red patch in the synthetic case but fails on the real plate, then the simplified simulation, not the luminance-agnostic architecture, is producing the reported gain.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Introduces FLUX.2-klein, the foundation image-editing model that the method fine-tunes with LoRA."},{"cited_title":"In: CVPR (2023)","cited_arxiv_id":null,"evidence_quote":"Established the Lab-space chroma-prediction paradigm with fixed luminance that the paper argues against."},{"cited_title":"In: SIGGRAPH Asia 2023 Conference Papers (2023)","cited_arxiv_id":null,"evidence_quote":"Supplies ColorNet, the learned perceptual colorfulness metric used in quantitative evaluation."},{"cited_title":"In:ProceedingsoftheIEEEconferenceoncomputervisionandpatternrecognition","cited_arxiv_id":null,"evidence_quote":"Provides the COCO evaluation split used for comparisons."}],"review_version":2}