{"id":"7a6965dd-1c17-4eaf-82a6-abb4d0817a70","arxiv_id":"2505.22291","paper_version":2,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"A modified ChaIR restoration network trained on synthetic greening defects removes green discoloration from autochrome photos, outperforming Photoshop's generative fill in qualitative tests.","lead":"This paper trains a neural network to remove green discoloration defects from scanned historical autochrome photographs, using only synthetically generated defect examples. It is relevant to cultural heritage preservation, where such defects are common and no annotated real datasets exist.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Real-world effectiveness rests on a hand-set simulator fitted to 7 autochromes; all quantitative metrics are synthetic, so the central restoration claim is not yet demonstrated on real greening defects.","rationale":"I read the paper as making two related claims: novelty ('first approach') and effectiveness ('efficient and effective restoration ... overcoming limitations of alternative techniques'). The novelty claim is plausible and not the main risk. The effectiveness claim is where the load-bearing weakness sits: every quantitative result is generated by the same synthetic model that defines the training distribution. The reader's verdict correctly identifies the simulator's faithfulness as the weakest assumption. My reading agrees with that assessment, and I found no additional internal inconsistency that would change the verdict. I deliberately considered whether the missing code/data release or the small number of real qualitative examples should push the verdict to REJECT, but those are reproducibility and evidence-strength concerns rather than demonstrations of incorrectness. The paper's own conclusion discloses relevant limitations (bluish tones, missing small defects, need for expansion to other collections), and the qualitative results on real autochromes are consistent with a plausible but unproven method. A conditional acceptance is therefore the right level: the central claim should be accepted only if the simulator's representative power is demonstrated on real data, ideally through the independent simulator and expert-preference test proposed above. No ad hominem considerations apply; the critique is about the evidence chain, not the authors. I also credit the authors for transparently discussing limitations and for attempting comparisons to strong baselines, but those comparisons inherit the same synthetic-data weakness, so they do not independently support the real-world claim.","tokens_in":9489,"tokens_out":4065,"duration_ms":53499,"concrete_test":"Fit a second, independent simulator: have a conservator segment 30-50 real greening regions from at least two archives other than the Harold Taylor collection, measure per-ring RGB ratio distributions directly from those pixels, and use these measured distributions instead of Table 1's hand-set dictionary to generate a held-out synthetic test set. Then fine-tune ChaIRFinetunedV2 on this measured-parameter simulator and compare it against the current model on 20 held-out real greening images in a blind expert preference test covering greening removal, color plausibility, and grain preservation. If the measured-parameter model outperforms the current model on real images, the quantitative claims in Tables 2-4 are artifacts of the hand-set simulator; if it does not, the simulator is not the bottleneck and the real-world restoration claim is strengthened.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central claim is automatic removal of greening in real digitized autochromes, but the only quantitative evidence is PSNR/MS-SSIM computed on synthetic images generated by the very simulator used for training (Tables 2 and 3). The simulator is not fitted from data: Section 3 states that defects were characterized by visual analysis of just 7 autochromes from the Harold Taylor collection, and Table 1 is a hand-set dictionary of per-ring multiplicative RGB coefficients, with a random perturbation factor of 0.2 chosen 'empirically'. Real greening is a physical photochemical process in which green starch-grain dyes bleed into neighboring layers; a smooth radial field of per-channel multipliers may not capture transparency, orange borders, dark cores, spatially correlated color shifts, or dependence on the underlying image content. If the simulator misses these properties, the network may simply learn to invert synthetic radial green blobs, and the qualitative real-image results (Fig. 7) remain anecdotal. The conclusion reinforces this concern by conceding that the method 'sometimes struggles to accurately reproduce the original colors', often producing bluish tones and missing small defects, and the future-work section states that the synthetic dataset must be expanded to support digitized autochromes from other collections. The comparisons against Photoshop and InstructIR (Tables 4, Figs. 9-10) are also demonstrated on the same synthetic distribution, so they do not independently validate real-world restoration. Thus the load-bearing assumption is that Table 1's model is representative of real greening defects; that assumption is currently supported only by a small, single-collection visual inspection and by a few qualitative examples.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper targets automatic removal of \"greening\" color defects in digitized autochrome photographs. Since no paired real-world data with defect-free ground truth exists, the authors design a synthetic degradation simulator based on visual inspection of seven autochromes from the Harold Taylor collection, generate defected/undamaged training pairs from 306 defect-free autochromes, and train a ChaIR-based restoration network with a modified loss that upweights defect regions. The method is compared quantitatively against Pix2Pix, CycleGAN, and a pretrained ChaIR variant on synthetic test images, and qualitatively against Photoshop Generative Fill and InstructIR-based restoration on real autochromes. The paper claims to be the first learning-based approach for automatic greening removal and to provide efficient and effective restoration.","tokens_in":9922,"tokens_out":4128,"duration_ms":46098,"significance":"If the central assumption holds—that the synthetic defect simulator faithfully represents real greening in autochromes—the paper would offer a practical tool for cultural heritage preservation and a reusable synthetic-data recipe. The authors provide a concrete pipeline, a publicly promised dataset and code release, a modified loss for color-imbalanced defects, and a useful discussion of failure modes. However, the quantitative evidence is currently self-referential: the network is trained and evaluated on outputs of the same simulator, so the reported PSNR/MS-SSIM values largely measure the model's ability to invert its own synthetic corruption model. The real-image evaluation is qualitative and anecdotal, with no metrics, statistical tests, or independent validation. The significance of the contribution therefore depends on future evidence that the simulator generalizes, which the paper itself flags as open future work.","major_comments":[{"comment":"The central quantitative evaluation is circular in a practical sense. The test images in Tables 2 and 3 are generated by the same synthetic degradation algorithm used to create the training data, so the reported PSNR and MS-SSIM scores measure self-consistency with the simulator rather than restoration performance on real greening defects. The paper's abstract and conclusion claim \"efficient and effective restoration\" of the considered defects, but this claim is not supported by the quantitative evidence. The authors should add a quantitative real-image evaluation protocol—for example, expert ratings on a defined scale, no-reference image quality metrics, or paired scans where available—or at minimum provide a cross-validation study that fits the simulator on a subset of real defects and tests on held-out real defect samples.","section":"§5, Tables 2–3"},{"comment":"The synthetic simulator is fitted to an extremely small and homogeneous evidence base: visual analysis of seven autochromes from one collection (Harold Taylor). The per-ring RGB corruption percentages in Table 1 are hand-set, the random adjustment factor 0.2 is chosen \"empirically,\" and there is no quantitative comparison of the simulated defect statistics (color distributions, ring geometry, transparency, edge behavior) against a held-out set of real greening defects. Because the real-image claims rest entirely on this simulator, the authors should provide evidence that the synthetic defect distribution matches real greening beyond the seven training examples, and should test the trained model on autochromes from other collections or under different digitization conditions. The future-work statement that the synthetic dataset \"must be expanded to support digitized autochromes from other collections\" implicitly concedes that this generalization is currently unverified.","section":"§3, Table 1"},{"comment":"The comparison against Photoshop Generative Fill is reported only on synthetic test images, with two aggregate metrics and no sample size, variance, or statistical test. The claim that the method \"outperforms\" a state-of-the-art photo-editing tool is therefore not statistically established. The authors should specify the number of test images, report per-image or error-bar results, and ideally include a real-image comparison with a defined scoring protocol, since the practical claim is about real autochromes.","section":"§5, Table 4 and \"Quantitative results\""},{"comment":"The weighted loss identifies defect areas during training by thresholding |I_in − I_GT| at t = 0.1. In the synthetic setup, both I_in and I_GT are generated by the same pipeline, so the defect mask is known by construction and is perfectly aligned with the corruption. This does not demonstrate that the model can localize or correct defects in real images, where no ground-truth difference is available. The qualitative \"affected areas\" shown in Fig. 7(c) appear to rely on the same input-minus-output comparison and should not be presented as evidence of accurate defect detection on real images. The authors should clarify how the loss weighting transfers to inference and provide an independent real-image defect-localization evaluation if they wish to claim automatic detection.","section":"§3, Eqs. (3)–(4)"}],"minor_comments":[{"comment":"The phrase \"first approach\" appears twice in the abstract and again in the introduction; given the breadth of the restoration literature, the authors should either substantiate the novelty claim with a more exhaustive literature search or temper the wording.","section":"Abstract and §1"},{"comment":"The labeling procedure is described only as \"with assistance from an expert\"; the authors should specify the labeling protocol, the number of experts, and any inter-rater agreement measure, since the 306/95 split is used to justify the simulator design.","section":"§4"},{"comment":"The symbol S_i in the spatial and frequency losses is not defined. It should be stated whether S_i is the number of pixels, the spatial size, or a normalization constant.","section":"§3, Eqs. (1)–(2)"},{"comment":"Table 2's caption notes that the loss weights defect and non-defect areas equally (W(x,y)=1.0) for those results, which is inconsistent with the method section that defines w ∈ {0.1, 0.5}. Please clarify which configuration corresponds to which table and why W=1.0 is used for the baseline comparison.","section":"§5, Tables 2 and 3"},{"comment":"The first column \"Label\" contains entries such as '9', '1', '2', '3', '4', '99', and '20' with no explanation of what these labels mean or how they correspond to rings; the final row labeled \"Surface\" further confuses the schema. Please replace with a clear ring-index or name convention.","section":"§5, Table 1"},{"comment":"The conclusion states that \"AI based inpainting methods enable targeted defect removal in stationary regions in small areas\" while earlier sections argue Generative Fill is unsuitable; the authors should reconcile this apparent contradiction or clarify the intended workflow (e.g., using their model as a preprocessing step plus inpainting).","section":"§6"},{"comment":"References [10] and [11] appear to be the same arXiv preprint; please consolidate and use a consistent citation format. Also, the training/test split and the number of synthetic test images are not reported; these details are necessary for reproducibility.","section":"References"}],"recommendation":"major_revision","confidential_remarks":"The paper addresses a real and underserved problem, and the synthetic-data-plus-weighted-loss pipeline is a sensible starting point. However, the current manuscript's quantitative claims are built on a circular evaluation that cannot support the stated real-world restoration performance. I would like the authors to either add a credible real-image evaluation (even a small expert-rated study) or reframe the contribution as a synthetic benchmark and simulator study. The latter would lower the bar for acceptance, but the current abstract and conclusion overclaim."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: this is a sensible, narrowly scoped application paper. The genuinely new thing is the first learning-based attempt at removing greening defects in autochromes, built on a synthetic defect generator that encodes real observations of how greening looks. The loss tweak on ChaIR is minor, but the simulator is the real contribution.\n\nWhat the paper does well: the task is well motivated and previously unaddressed; the defect generation process is described concretely (ring-based color corruption, geometry from real examples); they compare against Pix2Pix, CycleGAN, ChaIR variants, Photoshop's Generative Fill, and InstructIR; and the conclusion is honest about the method's limits, noting bluish color casts and missed small defects. That kind of candor counts.\n\nWhere it is soft: the quantitative results are all measured on synthetic test images generated by the very simulator used for training, so they mostly show the model can invert its own degradation model. The simulator itself rests on visual analysis of only 7 autochromes from one collection; Table 1 is hand-set, and the 0.2 random adjustment is empirical. Real-image evaluation is qualitative only, with no no-reference metric or user study. The paper itself concedes the dataset must be expanded for other collections. The ablation shows the weighted loss helps a bit, but it is incremental.\n\nThe circularity concern is fair, but I would not call it fatal. For a first step on a niche restoration task, synthetic evaluation is an acceptable starting point if clearly labeled. The problem is the abstract and Section 1 claim 'effective restoration' of real defects, which overstates what the evidence supports. The title says 'purely synthetic data,' which is accurate, but the claim of restoration of greening defects goes beyond that.\n\nWho it is for: people working on cultural-heritage image restoration, especially with synthetic training data. They will find the defect model useful. General computer vision readers can skip.\n\nRecommendation: send it to peer review. A serious referee can push for real-image quantitative evaluation (e.g., no-reference metrics, or a small user study with curators) and for clarification that the current evidence is synthetic-only. With those revisions, it would be a solid workshop or short-conference paper. I would not desk reject it.","headline":"A sensible, narrowly scoped first step on autochrome greening restoration, with an honest synthetic-data pipeline but real-world effect shown only qualitatively.","tokens_in":10328,"tokens_out":2116,"would_cite":false,"duration_ms":22271,"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 greening defects in autochrome photographs can be removed automatically by a restoration network trained purely on synthetic defect simulations, with a modified loss that focuses color correction on…","keywords":["autochrome restoration","greening defect","synthetic data generation","image restoration","defect-aware loss","cultural heritage","color correction","generative AI"],"falsifier":"Run the trained model on autochromes from a different archive or digitization setup and inspect whether green stains disappear without bluish color casts, smoothed grain texture, or missed small spots; if the simulator's per-channel ratios are specific to one collection's aging and scanning conditions, defects from other collections will either persist or be over-corrected.","tokens_in":9315,"feed_emoji":"🎨","tokens_out":5958,"duration_ms":58289,"temperature":0.7,"pith_summary":"The paper claims to be the first to remove 'greening' color defects from digitized autochrome photographs automatically, without manual masking. Because no annotated dataset of real damaged autochromes exists, the authors build a simulator that reproduces the characteristic ring-shaped green stains from measured per-channel color changes, then train the ChaIR restoration network on synthetic defect/clean pairs. A modified loss weights defective regions two to ten times more heavily, forcing the network to correct colors where it matters. If the simulator is faithful, the method offers an efficient, automatic alternative to manual expert restoration and outperforms general-purpose tools such as generative inpainting and prompt-based image restoration.","feed_headline":"First automatic method removes greening defects from old autochromes","feed_subtitle":"Trained on purely synthetic stains, it restores original colors and beats manual editing tools.","key_machinery":"The load-bearing mechanism is the synthetic defect generator paired with a defect-weighted loss on the ChaIR channel-interaction restoration network, a model that combines spatial and frequency-domain losses with dual-domain channel attention. The simulator turns clean autochromes into defected/ground-truth pairs by drawing irregular ellipses (origins inside the image for spot defects, outside for large liquid-leak defects), interpolating intensity by $I = -d^2 + 1$, applying a per-ring dictionary of percent changes to red, green, and blue channels, and smoothing with a Gaussian filter; the resulting masks annotate the defect regions. The loss modification replaces ChaIR's uniform spatial L1 term with a weighted version in which $W(x,y)=1.0$ where input and ground truth differ by more than threshold $t=0.1$ and $w\\in\\{0.1,0.5\\}$ elsewhere, so the model is penalized 2–10 times more for color errors inside defects while preserving the frequency-domain loss.","core_discovery":"Greening in autochromes is not a uniform green overlay: analysis of real defects shows the green channel is least affected, often increasing in intensity, while the red and blue channels drop sharply, in concentric rings with dark cores and orange borders. The paper's central claim is that a restoration network trained purely on synthetic reproductions of this channel-wise corruption can automatically de-green real autochromes. The authors simulate defects as irregular ellipses with intensity falloff $I = -d^2 + 1$, per-ring RGB corruption percentages from a hand-built dictionary, and Gaussian smoothing, and they fine-tune ChaIR on these synthetic pairs using a weighted spatial loss $l_s = \\frac{1}{N}\\sum_{x,y} W(x,y)|\\hat{Y} - Y|$ with weights two to ten times higher in defect areas. Their evaluation reports higher PSNR and MS-SSIM than the baseline network, and qualitative removal of large and spotting defects on real images, where generic restoration and inpainting tools fail.","pith_inferences":["The same per-ring channel-corruption dictionary could be re-fit for other systematic dye defects, such as the 'oranging' the paper names as future work, giving a parameterized family of synthetic defect generators for historical color photography.","Because the weighted loss only needs input/ground-truth pairs during training, the method could be extended to semi-supervised settings: a learned defect detector could supply masks for real images and let the same loss train on unpaired data.","A direct testable extension would be to train on synthetic defects derived from several collections with different scanners and compare per-collection restoration quality; this would quantify how tightly the simulator is tied to the original seven-image analysis.","The code and dataset release promised by the paper would let conservators generate bespoke training sets for their own digitized autochromes, effectively turning the simulator into a reusable restoration tool."],"forward_implications":["Autochrome collections without defect annotations can still be restored by training exclusively on synthetically generated defects from defect-free scans.","The defect-weighted loss improves color fidelity inside damaged regions compared to a uniform loss, with larger gains at weight 10 than at weight 2.","The approach preserves the characteristic autochrome grain texture, unlike histogram matching or generative inpainting, so restored images remain plausible for curators.","The degreened output can serve as a starting point for further targeted inpainting of very small remaining defects."],"supporting_citations":[{"why":"Supplies the ChaIR base restoration network, including its spatial and frequency losses, which the paper fine-tunes with a modified defect-weighted loss.","marker":"[8]"},{"why":"Documents the greening mechanism and defect types in autochromes, providing the domain knowledge used to design the synthetic defect simulator.","marker":"[16]"},{"why":"Provides the 420-image public collection from which the 306 defect-free autochromes were taken to generate synthetic defect/clean training pairs.","marker":"[29]"},{"why":"Serves as a Pix2Pix image-to-image baseline compared against the proposed approach.","marker":"[13]"},{"why":"Serves as a CycleGAN unpaired image-to-image baseline compared against the proposed approach.","marker":"[36]"},{"why":"Serves as the prompt-based InstructIR restoration baseline that the paper shows fails to remove greening defects.","marker":"[7]"},{"why":"Serves as the commercial Generative Fill inpainting baseline that the paper demonstrates is unsuitable for large greening defects.","marker":"[12]"},{"why":"Provides the RESIDE dehazing dataset used for the pretrained ChaIR model that is then fine-tuned on the synthetic ChannelGreeningDefects data.","marker":"[17]"}],"fun_headline_variants":["Synthetic-data AI removes autochrome greening automatically","Purely synthetic training restores colors in damaged autochromes","Neural fix for greening defects needs no real defect examples","Autochrome de-greening: AI trained on simulated stains only"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The entire pipeline rests on the assumption that the synthetic defect simulator faithfully reproduces real greening; its geometry and per-channel color changes come from visual analysis of only seven autochromes from a single collection, with the per-ring RGB percentages in the dictionary hand-set to fit those observations.","fun_headline_variants_meta":{"raw":{"variants":["Synthetic-data AI removes autochrome greening automatically","Purely synthetic training restores colors in damaged autochromes","Neural fix for greening defects needs no real defect examples","Autochrome de-greening: AI trained on simulated stains only"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000549,"raw_usage":{"total_tokens":2628,"prompt_tokens":955,"completion_tokens":1673,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":571,"completion_tokens_details":{"reasoning_tokens":1601}},"tokens_in":571,"tokens_out":1673,"duration_ms":15492,"temperature":1.0,"reasoning_tokens":1601,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T13:10:16.459482+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the trained model on autochromes from a different archive or digitization setup and inspect whether green stains disappear without bluish color casts, smoothed grain texture, or missed small spots; if the simulator's per-channel ratios are specific to one collection's aging and scanning conditions, defects from other collections will either persist or be over-corrected.","supporting_citations":[{"cited_title":"Knowledge-Based Systems 282 (2023)","cited_arxiv_id":null,"evidence_quote":"Supplies the ChaIR base restoration network, including its spatial and frequency losses, which the paper fine-tunes with a modified defect-weighted loss."},{"cited_title":"Getty Publications (2013)","cited_arxiv_id":null,"evidence_quote":"Documents the greening mechanism and defect types in autochromes, providing the domain knowledge used to design the synthetic defect simulator."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the 420-image public collection from which the 306 defect-free autochromes were taken to generate synthetic defect/clean training pairs."},{"cited_title":"In: CVPR","cited_arxiv_id":null,"evidence_quote":"Serves as a Pix2Pix image-to-image baseline compared against the proposed approach."},{"cited_title":"In: ICCV","cited_arxiv_id":null,"evidence_quote":"Serves as a CycleGAN unpaired image-to-image baseline compared against the proposed approach."},{"cited_title":"In: Proceedings of the European Conference on Com- puter Vision (ECCV) (2024)","cited_arxiv_id":null,"evidence_quote":"Serves as the prompt-based InstructIR restoration baseline that the paper shows fails to remove greening defects."},{"cited_title":"adobe.com/uk/products/photoshop/generative-fill.html, visited on 10/30/2024","cited_arxiv_id":null,"evidence_quote":"Serves as the commercial Generative Fill inpainting baseline that the paper demonstrates is unsuitable for large greening defects."},{"cited_title":"IEEE Trans","cited_arxiv_id":null,"evidence_quote":"Provides the RESIDE dehazing dataset used for the pretrained ChaIR model that is then fine-tuned on the synthetic ChannelGreeningDefects data."}],"review_version":1}