{"id":"13a5c0bd-585e-4fc1-b5df-a57b4910a76c","arxiv_id":"2412.11634","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"A diffusion model and a 28K-pair synthetic dataset are introduced for repairing damaged historical document images, but the method requires the original character content and damage locations as user inputs.","lead":"This paper introduces Historical Document Repair, a task that reconstructs damaged historical documents, along with a large synthetic dataset and a diffusion-based model. The model repairs images when the original text content and damaged locations are provided as inputs, which limits its use on unknown damage.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The central claim is conditional on user-supplied content and mask cues (Eq. 2); without them DiffHDR cannot repair severe damage, so 'predicting original appearance' overstates the demonstrated capability. Same concern as the reader; quantifying the unconditional gap would settle it.","rationale":"The reader's weakest assumption is the same one I find most load-bearing: the method requires xc and xm at test time, so it does not autonomously predict original appearance. The formal model in the paper makes this explicit, and the Limitation section admits failure under severe damage without those cues. I considered whether a different issue—such as the FID/LPIPS protocol replacing non-repaired pixels with the target—might be more damaging, but that affects all compared methods equally and is secondary to the conditional-input issue. The dataset and controlled synthetic evaluation are genuine contributions, and the comparison to adapted baselines is reasonably fair, so a conditional verdict rather than rejection is appropriate. The proposed check would quantify how much of the reported performance, especially Rec-ACC, depends on the supplied content and mask; until that number is reported, the abstract's phrase 'predict the original appearance' remains an overclaim.","tokens_in":11888,"tokens_out":3672,"duration_ms":37436,"concrete_test":"Run the released DiffHDR checkpoint on the HDR28K test set in the unconditional mode (xc and xm set to blank, as in the Limitation's Figure 10) and compute FID, LPIPS, and Rec-ACC under the same region-replacement protocol used in Table 1. Compare with the conditional numbers (FID 0.7499, LPIPS 0.0384, Rec-ACC 81.9180). If Rec-ACC drops substantially (e.g., more than 20 points) or FID/LPIPS deteriorate, the central claim must be restated as 'repair given user-supplied content and location,' and the abstract should be adjusted accordingly.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's stated task (Eq. 1) is xr = FHDR(xd), but the proposed method (Eq. 2) is xr = F(xd, xc, xm), where xc and xm are parsed from user-provided content and location. The section 'DiffHDR: Diffusion-based HDR Network', under condition parsing, explicitly says 'the user provides the content and location of the damaged characters.' The Limitation section concedes that without semantic/spatial information and under severe damage, the method cannot repair. Thus the headline claim 'predict the original appearance' is not demonstrated for the autonomous setting; the model is a guided inpainter that is handed the answer text and region. This is not internal inconsistency, but it means the advertised capability is narrower than the abstract and title imply. A concrete consequence is that Rec-ACC, the main character-level metric, is measured with xc supplied, so it measures how faithfully the model renders given content, not how well it recovers unknown content. The real-document section (Fig. 9) does not report what cues were given; if the ground-truth transcription was provided, that demonstration also supports guided repair only. The held-out synthetic comparison is otherwise reasonably designed, and the dataset is a real contribution, so the correct verdict is conditional, not rejection.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces Historical Document Repair (HDR), a task of reconstructing the original appearance of damaged historical document images. It contributes HDR28K, a dataset of 28,552 synthetic damaged-repaired pairs with character-level annotations, and DiffHDR, a diffusion-based model conditioned on the damaged image xd, a content image xc, and a mask image xm, together with a character perceptual loss. Experiments on HDR28K compare against nine baselines and report lower FID/LPIPS and higher character-recognition accuracy, along with qualitative results on real damaged documents and extensions to document editing and text-block generation.","tokens_in":12119,"tokens_out":3478,"duration_ms":31035,"significance":"The dataset and the guided-repair formulation are potentially useful contributions: HDR28K is large, multi-style, and accompanied by character-level annotations; the held-out evaluation is reasonably designed with no same-document leakage between train and test; the comparison is broad; and the announced code and data release would support reproducibility. The reported gains on the guided task are substantial. However, the significance as currently framed is overstated: the method does not predict unknown damaged content but renders user-specified content into user-specified locations. If repositioned as an interactive repair tool, the work is solid; the autonomous 'predicting the original appearance' claim requires an unconditional model or a demonstration without user-supplied cues.","major_comments":[{"comment":"The central claim of predicting the original appearance of damaged documents is not supported for the autonomous setting. The task is defined in Eq. (1) as xr = FHDR(xd), but the proposed method in Eq. (2) is xr = F(xd, xc, xm), where xc and xm are parsed from user-provided content and location. The condition-parsing paragraph explicitly states that 'the user provides the content and location of the damaged characters,' and the Limitation section concedes that without semantic and spatial information, severe damage cannot be repaired. The abstract and title should be qualified, or the paper should report results without xc and xm and quantify the gap.","section":"Abstract; Eq. (2); Section 'DiffHDR: Diffusion-based HDR Network' (Condition Parsing); Limitation"},{"comment":"Rec-ACC is measured with xc supplied, because the recognizer is applied to repaired regions that were generated from the ground-truth content image. This metric therefore measures how faithfully the model renders given content, not how well it recovers unknown content. The 'character correctness' claim in Table 1 should be stated in conditional terms; an unconditional evaluation (for example, recognition accuracy on outputs produced without xc) is needed to support the autonomous framing.","section":"Evaluation Metrics; Table 1"},{"comment":"The real-document experiment does not report what inputs were provided: whether xc and xm were supplied and, if so, whether they came from manual transcription or an automated recognizer. Without this information, the qualitative results support only guided repair, not autonomous damage recovery. Please state the protocol and, if feasible, include a failure analysis when the cues are withheld.","section":"Real Damaged Document Image Repair (Figure 9)"}],"minor_comments":[{"comment":"The heading 'Historcial Document Image Processing' contains a typo; it should be 'Historical.'","section":"Related Work heading"},{"comment":"The word 'senmantic' in the explanation of xc should be 'semantic.'","section":"Eq. (2) explanation"},{"comment":"The notation 'sc.m' should be 'sc,m' to match the text describing the guidance scales.","section":"Eq. (6)"},{"comment":"The caption 'without semantic and visual cues' is ambiguous because the first row shows successful repair; the text clarifies this, but the figure should distinguish the moderate case from the severe failure case.","section":"Figure 10 caption"},{"comment":"The sentence describing the 8% dropout is ambiguous about the probabilities for the three conditions; please specify the exact sampling schedule (for example, each condition with 8% independently or jointly).","section":"Attribute-Sensitive Repair Strategy"}],"recommendation":"major_revision","confidential_remarks":"The paper is a reasonable candidate for a document-analysis venue if the claims are scaled to the conditional method. The main risk is the mismatch between the title/abstract and the actual input requirements; the authors should be pushed to either add unconditional experiments or explicitly reframe the contribution as interactive repair. I do not see grounds for rejection, because the dataset and the guided model are likely useful contributions to the community."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The thing to know: HDR28K is a real contribution, and DiffHDR is a solid conditional repair baseline, but the paper sells the result as autonomous reconstruction when the method is actually handed the damaged characters' content and location. The stress-test note and the reader's report land; I agree with the conditional verdict.\n\nWhat is actually new and good: HDR28K gives 28,552 damaged-repaired pairs with character-level annotations, three degradation types built from MTHv2 and M5HisDoc, and a clean source-document train/test split. That is a useful resource for document analysis and digital humanities. DiffHDR is a reasonable baseline: a diffusion denoiser conditioned on content image x_c and mask x_m, with classifier-free guidance and a character perceptual loss. The comparison against 9 methods is broad, and the gains in FID, LPIPS, and Rec-ACC are large. The real-document demos, document editing, and text block font generation are welcome extras. The paper is also honest in its Limitation section: with severe damage and no semantic/spatial cues, it cannot repair.\n\nSoft spots, in proportion. The central framing problem is real. Eq. (1) defines the HDR task as x_r = F_HDR(x_d), but the method is Eq. (2) F(x_d,x_c,x_m), where the user supplies content and location. The abstract says the model can \"predict the original appearance,\" which overstates the demonstrated capability. Rec-ACC is measured with x_c supplied, so it tests how faithfully the model renders given content, not how well it recovers unknown content. Figure 9 does not say what cues were used for the real documents; if the transcription was provided, that demo also shows guided repair. Missing error bars are a smaller issue, but should be fixed. I do not think the evaluation is circular—the held-out set and independently trained recognizer prevent that—but the task definition should be reframed as guided historical document repair, or the authors should add an unconditional baseline and quantify the gap.\n\nWho this is for: people working on document image restoration, historical document analysis, and digital humanities infrastructure. It deserves a serious referee; the dataset alone justifies reviewer time. My recommendation: send to peer review, but ask for major revision that aligns the abstract and title with the conditional inputs, adds unconditional results or explicitly reframes the contribution, and reports variability across seeds. The conditional claim holds; the advertised autonomous claim does not.","headline":"A genuinely useful benchmark dataset and a strong conditional inpainter, but the title and abstract promise autonomous reconstruction the method doesn't deliver.","tokens_in":12666,"tokens_out":2626,"would_cite":true,"duration_ms":26118,"reading_group":"yes","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 claims that a diffusion-based network, DiffHDR, can predict the original appearance of damaged historical documents when given the damaged image plus a content image and mask of the missing characters, and that it outperforms…","keywords":["historical document repair","diffusion models","document image restoration","HDR28K dataset","character perceptual loss","cultural heritage digitization","image inpainting","classifier-free guidance"],"falsifier":"Feed DiffHDR real damaged document patches with severe ink erosion or large missing regions while suppressing the content and mask cues (as in the paper's limitation figure) and measure the repaired-region character recognition accuracy; the paper's own limitation predicts it will remain near chance. A complementary test is to supply a deliberately wrong transcript for one damaged character: if the model faithfully renders the wrong character, that confirms the repair is copying the supplied cue rather than inferring the original content, which is exactly the conditional behavior the paper claims.","tokens_in":11667,"feed_emoji":"📜","tokens_out":7176,"duration_ms":59451,"temperature":0.7,"pith_summary":"This paper introduces Historical Document Repair as a new task and argues that the original appearance of a damaged historical document can be predicted, not just binarized or enhanced. It contributes HDR28K, a dataset of 28,552 damaged-repaired image pairs built from two Chinese historical document collections with three synthetic degradations: character missing, paper damage, and ink erosion. The proposed DiffHDR is a diffusion model that takes the damaged patch together with a user-supplied content image and damage mask as conditions, and directly produces the repaired image. On the HDR28K test set, the authors report that DiffHDR outperforms nine existing restoration and generation methods on FID, LPIPS, and repaired-character recognition accuracy, and they show qualitative repairs of real damaged scans. The broader promise is a practical tool for cultural-heritage conservation that can reconstruct missing text in context.","feed_headline":"A diffusion model restores missing characters in historical documents","feed_subtitle":"Trained on 28,552 synthetic damage pairs, DiffHDR also repairs real scans and edits text blocks in style.","key_machinery":"The load-bearing mechanism is an 8-channel conditional diffusion denoiser that concatenates the noised image, the damaged image, the content image, and the mask image, and outputs the repaired image in a single forward pass. The content image supplies the semantics of what should appear and the mask supplies the spatial location, so the model does not need to infer damaged content from pixels alone. Randomly dropping these conditions during training enables classifier-free guidance at inference, with separate guidance scales for the damaged image and for the semantic/spatial pair. The character perceptual loss, computed as a masked multi-scale VGG feature distance between the repaired and target regions, focuses training on producing correct character content and style within the repaired area.","core_discovery":"The central claim is that a conditional diffusion process can reconstruct the original appearance of damaged historical documents when it is told what the missing characters say and where they sit. Formally, the model learns $x_r = F(x_d, x_c, x_m)$, where $x_d$ is the damaged image, $x_c$ is a content image encoding the transcript of the damaged characters, and $x_m$ is a mask marking the damaged regions. DiffHDR's denoiser predicts the repaired image directly, rather than the added noise, and is trained with a diffusion loss plus a character perceptual loss that penalizes feature mismatches inside the repaired regions. The authors report that this design outperforms the compared methods on the HDR28K benchmark and that the model trained only on synthetic degradations still repairs real damaged documents, edits text content, and generates new text blocks in the document's style.","pith_inferences":["The method's reliance on user-provided transcripts means an end-to-end repair tool would need an upstream module that reads partially damaged text; the paper names this as future work but does not implement it.","Because HDR28K is built from Chinese historical documents and three synthetic damage types, transfer to other scripts, inks, and damage modes is untested; analogous datasets for other traditions would be a direct extension.","The conditional formulation suggests a controlled-use application: by deliberately altering the content image, the same model could serve as a text-editing or reconstruction tool for digital heritage, a capability the paper only begins to explore."],"forward_implications":["HDR28K gives future work a common benchmark of 28,552 damaged-repaired pairs with character-level annotations, so new repair methods can be trained and compared on the same ground.","A model trained only on synthetic degradations transfers to real damaged scans, which means large-scale synthetic data can substitute for scarce real damaged-repaired pairs.","The same conditional diffusion framework performs document editing and text block font generation, so one model can rewrite or extend a historical document in its own style.","Repaired-region character recognition accuracy is substantially higher with DiffHDR than with the compared baselines, implying fewer hallucinated or wrong characters in the output."],"supporting_citations":[{"why":"Supplies the MTHv2 source document images and character-level location annotations used to build HDR28K.","marker":"(Ma et al. 2020)"},{"why":"Supplies the M5HisDoc source document images and annotations that form the other half of HDR28K.","marker":"(Shi et al. 2023)"},{"why":"Provides the large-mask inpainting model used to erase content for the character-missing degradation.","marker":"(Suvorov et al. 2022)"},{"why":"Defines the DDPM forward noising process and diffusion training objective that DiffHDR extends.","marker":"(Ho, Jain, and Abbeel 2020)"},{"why":"Provides the instructive image-editing formulation that motivates conditioning diffusion on task-specific attributes.","marker":"(Brooks, Holynski, and Efros 2023)"},{"why":"Supplies the classifier-free guidance technique used to control sensitivity to the damaged image and semantic/spatial cues.","marker":"(Ho and Salimans 2022)"},{"why":"Provides the VGG feature network used in the character perceptual loss and in the recognizer that measures repaired-character accuracy.","marker":"(Simonyan and Zisserman 2014)"},{"why":"Provides the DPM-Solver++ fast sampler used to generate repaired images at inference.","marker":"(Lu et al. 2022)"}],"fun_headline_variants":["Diffusion model predicts original look of damaged texts","AI restores missing characters using diffusion","DiffHDR: diffusion repairs historical documents","Reconstructing damaged manuscripts with diffusion","Diffusion predicts original appearance of damaged documents"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"At test time, someone must already know and supply the content and exact location of every damaged character, and if the damage is severe the method cannot repair the image without those cues.","fun_headline_variants_meta":{"raw":{"variants":["Diffusion model predicts original look of damaged texts","AI restores missing characters using diffusion","DiffHDR: diffusion repairs historical documents","Reconstructing damaged manuscripts with diffusion","Diffusion predicts original appearance of damaged documents"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000591,"raw_usage":{"total_tokens":2777,"prompt_tokens":954,"completion_tokens":1823,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":570,"completion_tokens_details":{"reasoning_tokens":1758}},"tokens_in":570,"tokens_out":1823,"duration_ms":12648,"temperature":1.0,"reasoning_tokens":1758,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T14:44:11.894826+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Feed DiffHDR real damaged document patches with severe ink erosion or large missing regions while suppressing the content and mask cues (as in the paper's limitation figure) and measure the repaired-region character recognition accuracy; the paper's own limitation predicts it will remain near chance. A complementary test is to supply a deliberately wrong transcript for one damaged character: if the model faithfully renders the wrong character, that confirms the repair is copying the supplied cue rather than inferring the original content, which is exactly the conditional behavior the paper claims.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the M5HisDoc source document images and annotations that form the other half of HDR28K."}],"review_version":1}