{"id":"5f6c75b7-a7dd-4ec7-a5e0-17fe39d3c051","arxiv_id":"2506.04641","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":9,"one_line_summary":"TADiSR is a diffusion-based super-resolution model that jointly predicts a text segmentation mask from cross-attention maps, improving text legibility in restored full images.","lead":"This paper introduces TADiSR, a diffusion model for image super-resolution that adds a text segmentation decoder driven by the model's own attention maps, so text in low-resolution photos is restored more accurately. The work matters because scene text is everywhere, and existing super-resolution models often corrupt letters and strokes.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The 18.7% OCR-A gain on Real-CE is not established as genuine text-fidelity improvement because the test set is author-filtered and the training pipeline is curated by the same OCR/segmentation model family used to define the task; an independent OCR and unfiltered test-set check is needed.","rationale":"The reader's weakest assumption identified the pseudo-label and attention-localization risk, and I agree that this is a real concern. However, the most load-bearing issue for the central claim is more specific: the headlined quantitative result on Real-CE is measured on an author-filtered test set and scored with an OCR engine that was also used to filter the training data. This creates a concrete route by which the 18.7% OCR-A gain could be an artifact of pipeline alignment rather than a general improvement in text legibility. The paper is otherwise methodologically plausible: the ablations in Table 2 show each component contributes, the FTSR synthesis is described in detail, and the authors include a limitation appendix with a failure case. No fatal flaw is apparent, but the external validity of the main quantitative claim cannot be assessed without an independent evaluation protocol. I therefore keep the reader's CONDITIONAL verdict unchanged rather than moving to ACCEPT or REJECT.","tokens_in":12654,"tokens_out":5520,"duration_ms":72689,"concrete_test":"Re-run the comparison on the full Real-CE test set without the authors' manual misalignment filtering, using at least two independent OCR engines (e.g., PP-OCR and TrOCR) and computing OCR-A with standard errors; if the 18.7-point gap over HAT shrinks below significance or reverses under either condition, the claimed state-of-the-art text-fidelity improvement is not robust.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central quantitative claim is the OCR-A gain on Real-CE-val in Table 1, where TADiSR beats HAT by 18.7 points. Two properties make this number hard to interpret. First, Real-CE-val is not the standard benchmark split: Section 4.1 states that the authors manually filtered out misaligned pairs, yielding 189 test images. The filtering criterion is subjective and applied only by the authors; no results on the full, unfiltered Real-CE set are reported, so hard cases that might disproportionately hurt baselines can be silently removed. Second, the training pipeline is itself OCR-filtered: Section 3.5 filters pseudo-GT text patches by comparing OCR recognition results between the original CTR images and their segmentation maps, and Section 4.2 measures OCR-A using an OCR model (PP-OCR [7]) with a Levenshtein ratio. If the same OCR engine is used to curate training data and to score test outputs, the model is implicitly trained to produce text that this OCR engine reads as matching the GT, so the metric can reflect alignment with the training-time OCR filter rather than generic legibility. In addition, the segmentation supervision is generated by SAM-TS/Hi-SAM, and any systematic bias in those masks (e.g., over-merging dense strokes, smoothing complex characters) is absorbed into the reconstruction. The supplementary Hi-SAM comparison does not break this loop because Hi-SAM is the same model family that produced the training labels. The paper reports no error bars, significance tests, or human readability evaluation. Therefore the load-bearing assumption that the 18.7% gain reflects genuine text fidelity is untested against an independent OCR engine and an unfiltered, third-party-aligned test set.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper proposes TADiSR, a diffusion-based real-world image super-resolution framework that augments a latent diffusion model with text-aware cross-attention and joint segmentation decoders. The authors also introduce the FTSR dataset, constructed by pasting segmentation-masked text patches onto high-quality backgrounds, and report state-of-the-art results on both the synthetic FTSR test set and a manually filtered validation split of Real-CE, including an 18.7% OCR-A improvement over HAT on Real-CE-val. The work addresses the practically important problem of preserving text legibility in super-resolved images and provides extensive qualitative comparisons, including challenging real-world cases.","tokens_in":13043,"tokens_out":4329,"duration_ms":50017,"significance":"If the quantitative claims hold, TADiSR would be a valuable step toward practical, single-pass text-aware super-resolution, offering an alternative to OCR-based pipeline approaches. The paper is clearly written, the architecture is plausible, and the authors include a thorough empirical study with several baselines. However, the evaluation protocol contains weaknesses that currently prevent full confidence in the headline numbers: the Real-CE test set is author-filtered, the OCR metric may be circular with the training data curation, no statistical uncertainty is reported, and the ablations are confined to the synthetic FTSR set. These issues are fixable and do not invalidate the core idea, but they must be addressed before the claimed gains can be accepted.","major_comments":[{"comment":"The Real-CE evaluation uses only the 189 manually filtered 'aligned' validation pairs, and no results are reported on the full, unfiltered Real-CE validation or test set. Because the filtering criterion is subjective and performed only by the authors, the 18.7% OCR-A gain over HAT could be driven by the removal of hard cases that disproportionately affect baselines. The authors should report metrics on the complete Real-CE test split (or at least the unfiltered validation set), state the exact filtering rule, and release the list of retained sample indices for reproducibility.","section":"§4.1, Table 1"},{"comment":"The training-data curation and the OCR-A evaluation rely on the same OCR engine (PP-OCR [7]). Section 3.5 filters pseudo-GT text patches by comparing OCR recognition results between original images and their segmentation maps, and Section 4.2 defines OCR-A using PP-OCR recognition with Levenshtein ratio. This creates a training-evaluation circularity that can inflate the reported OCR gains without improving generic text legibility. The authors should evaluate with at least one independent OCR engine (e.g., Tesseract or a different commercial OCR) and explicitly state whether the evaluation OCR is identical to the one used in data filtering.","section":"§3.5 and §4.2"},{"comment":"All metrics are reported as point estimates without error bars, confidence intervals, or significance tests. The Real-CE-val set contains only 189 images, yet the paper claims a large 18.7% OCR-A improvement. The authors should provide bootstrap confidence intervals or a paired statistical test (e.g., Wilcoxon signed-rank over images) to demonstrate that the differences are not due to a few outliers.","section":"Table 1 and §4.2"},{"comment":"The ablation study (Table 2) is conducted only on the synthetic FTSR-TE set, so it does not establish which components are responsible for the real-world gains on Real-CE. The authors should include the same ablations (w/o JSD, w/o TACA, w/o MF Loss) on Real-CE-val, at least for OCR-A, to support the claim that each module contributes to real-world text fidelity.","section":"§4.3"}],"minor_comments":[{"comment":"The symbols alpha_t and beta_t are used without definition; please define them by referencing the diffusion noise schedule (e.g., DDPM notation) or by giving the formulas in the text.","section":"Eq. (1)"},{"comment":"The OCR-A protocol should clarify whether the same OCR detector/recognizer is used for all methods and whether the recognizer is PP-OCR; the metric definition would be easier to interpret with this information.","section":"§4.2"},{"comment":"The paper states that baselines are fine-tuned on FTSR and Real-CE using official code 'if provided,' but it does not list which baselines were fine-tuned and which were not. A supplementary table with training details for each baseline would improve reproducibility.","section":"§4.2"},{"comment":"The caption acknowledges that 'part of our training segmentation labels were selected from Hi-SAM outputs.' This limits the strength of the segmentation comparison against Hi-SAM and should be mentioned in the main text as a known limitation.","section":"Supplementary Fig. 7"},{"comment":"Minor typographical issues include a stray space in 'V AE,' a missing space in 'Wuet al.,' and 'Flicker2K' should be 'Flickr2K.'","section":"Throughout"}],"recommendation":"major_revision","confidential_remarks":"The paper is technically interesting and the proposed method is plausible, but the central quantitative claim of a large OCR-A gain on Real-CE rests on a non-standard, author-filtered test split and an OCR metric that shares the same engine as the training data curation. I recommend asking the authors to provide results on the full Real-CE set or an independently filtered split, to use an independent OCR engine for evaluation, and to include ablations on Real-CE. The industrial affiliation and the release of code (if included in the camera-ready) would be assets, but the benchmark filtering should be transparent to the community."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Worth a serious refereeing, but the headline OCR-A number on Real-CE needs independent verification before anyone builds on it. The actual new thing here is architectural: taking the cross-attention response to the single token 'text' from a LoRA-tuned diffusion U-Net, linearly projecting it, and feeding it together with the denoised latent into a jointly trained text segmentation decoder with cross-decoder interactions. That gives single-pass full-image text-aware SR without the OCR-crop-paste pipeline that MARCONet and DiffTSR require, and the paper makes a plausible case that the two tasks reinforce each other. The FTSR synthesis pipeline is also a real contribution: it produces full-image triplets with text masks, which didn't exist at scale, and the filtering steps are described in enough detail to be reproduced. The ablations are clean: every removed component drops OCR-A, with the joint decoders giving the largest drop. I also credit the honest limitation section showing a genuinely hard failure case.\n\nNow the soft spots, in proportion. The 18.7% OCR-A gain on Real-CE-val is the load-bearing number, and it is not fully established. The test set is manually filtered to 189 aligned pairs, with no results on the unfiltered split; that leaves room for favorable selection. More importantly, the same OCR model family (PP-OCR) is used both to filter pseudo-GT text patches during training and to score OCR-A at evaluation, and the segmentation pseudolabels come from the same Hi-SAM family used in the supplementary comparison. So part of the gain could reflect the model learning to please the training-time filter rather than generic legibility. The stress-test note is right about this. I would not call it fatal—the qualitative results and the consistency of the ablations suggest something real is happening—but an independent OCR engine and an unfiltered, third-party-aligned test evaluation would settle it. There are also no error bars or significance tests anywhere, and the code/data links are unverified, though code is claimed.\n\nFor whom: anyone working on real-world SR or text restoration will want to read this; it is stronger as a systems paper than as a benchmark paper. It deserves a serious referee—send it out, but ask for the independent OCR check and unfiltered Real-CE numbers before final acceptance.","headline":"Worth refereeing; the Real-CE OCR-A gain looks real but is not yet established independently.","tokens_in":13626,"tokens_out":2431,"would_cite":false,"duration_ms":27786,"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":"TADiSR, a diffusion SR model with text-aware cross-attention and joint segmentation decoders, claims to substantially improve text legibility, reporting an 18.7% OCR-A gain over HAT on Real-CE.","keywords":["text-aware super-resolution","diffusion models","text segmentation","cross-attention fine-tuning","joint segmentation decoders","synthetic text SR dataset","real-world image super-resolution","OCR accuracy"],"falsifier":"On a held-out set of real-world text images from unseen domains with human-annotated masks, compute the IoU between TADiSR's predicted text masks and the human masks; if the IoU is no better than that of a model trained without the joint segmentation decoders, the claimed structural benefit of JSD is not supported. Alternatively, retrain the pipeline using masks from a different segmentation model and check whether the OCR-A advantage over HAT persists; a large sensitivity to the mask source would imply the results are an artifact of the specific pseudo-label generator.","tokens_in":12417,"feed_emoji":"📝","tokens_out":7376,"duration_ms":77862,"temperature":0.7,"pith_summary":"This paper presents TADiSR, a diffusion-based framework for real-world image super-resolution that treats text as a first-class citizen. The authors claim that by fine-tuning the cross-attention layers of a latent diffusion model so they focus on the word 'text' in a fixed prompt, and by decoding image and text segmentation jointly through interacting decoder branches, the model recovers character-level details that generic GAN- and diffusion-based super-resolvers distort. They support this claim with a new synthetic dataset, FTSR, built by pasting filtered, super-resolved text patches onto high-quality backgrounds, and with experiments showing improved PSNR, SSIM, LPIPS, FID, and OCR accuracy over existing methods on both synthetic and real paired data. The strongest reported result is an 18.7% improvement in OCR-A over the HAT baseline on the real-world Real-CE validation set. If true, this shows that text-aware attention is a practical route to preserving structural fidelity in generative super-resolution.","feed_headline":"Text-aware diffusion SR lifts OCR accuracy by 18.7% on real scenes","feed_subtitle":"A single-pass diffusion model with joint text-segmentation decoders restores legible characters in blurred, real-world images.","key_machinery":"The two load-bearing components are Text-Aware Cross-Attention (TACA) and the Joint Segmentation Decoders (JSD). TACA extracts, from every cross-attention layer of the U-Net, the response slice corresponding to the single token 'text' in the prompt, concatenates these slices, and projects them with a learned matrix; after LoRA fine-tuning these maps concentrate on textual regions. JSD consists of the VAE image decoder and a symmetric text segmentation decoder that exchange features through Cross-Decoder Interaction Blocks (CDIB), which split each branch's features, exchange half across branches through a sigmoid-gated Hadamard product, and use zero-initialized residual scaling to stabilize training. The segmentation decoder turns the attention maps into text masks, giving the image decoder an explicit structural target.","core_discovery":"The central discovery is that the cross-attention maps of a pre-trained latent diffusion model, when fine-tuned with LoRA on a joint super-resolution and text-segmentation task, become a reliable spatial indicator of text regions for the token 'text'. These attention maps are linearly projected and fed, together with the denoised latent, into a pair of interacting decoders: the original VAE image decoder and a newly added text segmentation decoder, connected by Cross-Decoder Interaction Blocks. This design lets structural text information flow between the two tasks, and a modified focal loss that weights edge-gradient mismatch by segmentation-confidence sharpens character boundaries. The paper further claims that the accompanying FTSR data-synthesis pipeline, which combines OCR-filtered text patches from recognition datasets with high-quality background images, provides the fine-grained, full-image text masks needed to train such a model at scale.","pith_inferences":["The attention-to-token mechanism could be probed as a general spatial localizer: the same LoRA-tuned attention maps might serve as candidate text priors for text inpainting, text removal, or scene text editing without retraining the segmentation head.","The 18.7% OCR-A gain over HAT suggests that the bottleneck in current generative super-resolution is not raw detail synthesis but the model's failure to 'know' where text is; TADiSR's mechanism is one way to inject that knowledge, and a text-recognition loss at the character level might push accuracy further.","A natural stress test is domain shift: because the fixed prompt contains only the word 'text', the method may saturate for unusual typefaces, decorative fonts, or low-contrast text; extending the prompt with style tokens (e.g., 'neon sign', 'handwriting') could specialize attention without architectural changes.","If the FTSR dataset is adopted as a benchmark, its pseudo-GT masks could themselves become a bottleneck; a comparison with human-annotated masks on a subset would tell whether label noise is capping the achievable text fidelity."],"forward_implications":["A single forward pass of TADiSR produces a full-image super-resolved output with intact text, removing the need for OCR-based detection, cropped text super-resolution, and subsequent fusion used by prior text-SR methods.","Because the text-awareness comes from tuning attention to a token rather than adding a recognition module, the same recipe can in principle be applied to other under-represented semantic categories by selecting their tokens.","The FTSR synthesis pipeline is scalable: any text-segmentation or detection dataset can be filtered with OCR and pasted onto high-quality backgrounds, enabling large-scale training data for text-aware restoration.","Joint training strengthens both tasks: the paper reports that the text masks predicted by TADiSR on degraded real-world photos are finer than those of a dedicated text segmentation model, implying the SR path helps segmentation under degradation.","Across synthetic FTSR and real Real-CE benchmarks, TADiSR reports the best quantitative scores on all reported metrics, with the largest margin on OCR-A, indicating that the improvements transfer from synthetic training to real capture conditions."],"supporting_citations":[{"why":"Supplies the latent diffusion backbone whose cross-attention layers are fine-tuned.","marker":"[23]"},{"why":"Provides the parameter-efficient LoRA adapters used to fine-tune U-Net and VAE decoder.","marker":"[11]"},{"why":"Supplies the synthetic degradation pipeline used to generate LR images and serves as the primary GAN baseline.","marker":"[32]"},{"why":"Provides the real paired scene-text dataset (Real-CE) used for additional training and evaluation.","marker":"[18]"},{"why":"The SAM-based text segmentation model fine-tuned to produce pseudo-GT masks for FTSR and Real-CE.","marker":"[40]"},{"why":"One of the two text segmentation datasets used to fine-tune SAM-TS for mask generation.","marker":"[37]"},{"why":"The bilingual text segmentation dataset used alongside TextSeg to fine-tune SAM-TS.","marker":"[38]"}],"fun_headline_variants":["Diffusion SR with joint decoders sharpens text in real scenes","Text-aware attention and segmentation boost super-resolved text","TADiSR: diffusion model recovers legible text in degraded images","Joint segmentation decoders improve text fidelity in diffusion SR","Diffusion super-resolution with text masks beats prior art on text"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The method works only if, after LoRA fine-tuning, the cross-attention map for the token 'text' reliably marks the location of text in images the model has not seen, and if the pseudo-ground-truth text masks generated by the fine-tuned SAM-TS model (trained on TextSeg/BTS and applied to CTR and Real-CE) are accurate enough to supervise joint training without systematic bias.","fun_headline_variants_meta":{"raw":{"variants":["Diffusion SR with joint decoders sharpens text in real scenes","Text-aware attention and segmentation boost super-resolved text","TADiSR: diffusion model recovers legible text in degraded images","Joint segmentation decoders improve text fidelity in diffusion SR","Diffusion super-resolution with text masks beats prior art on text"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000187,"raw_usage":{"total_tokens":1292,"prompt_tokens":873,"completion_tokens":419,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":489,"completion_tokens_details":{"reasoning_tokens":333}},"tokens_in":489,"tokens_out":419,"duration_ms":4764,"temperature":1.0,"reasoning_tokens":333,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T10:36:47.081165+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"On a held-out set of real-world text images from unseen domains with human-annotated masks, compute the IoU between TADiSR's predicted text masks and the human masks; if the IoU is no better than that of a model trained without the joint segmentation decoders, the claimed structural benefit of JSD is not supported. Alternatively, retrain the pipeline using masks from a different segmentation model and check whether the OCR-A advantage over HAT persists; a large sensitivity to the mask source would imply the results are an artifact of the specific pseudo-label generator.","supporting_citations":[{"cited_title":"Lora: Low-rank adaptation of large language models.ICLR, page 3, 2022","cited_arxiv_id":null,"evidence_quote":"Provides the parameter-efficient LoRA adapters used to fine-tune U-Net and VAE decoder."},{"cited_title":"Real-esrgan: Training real-world blind super-resolution with pure synthetic data","cited_arxiv_id":null,"evidence_quote":"Supplies the synthetic degradation pipeline used to generate LR images and serves as the primary GAN baseline."},{"cited_title":"A benchmark for chinese-english scene text image super-resolution","cited_arxiv_id":null,"evidence_quote":"Provides the real paired scene-text dataset (Real-CE) used for additional training and evaluation."},{"cited_title":"Hi-sam: Marrying segment anything model for hierarchical text segmentation.TPAMI, 2024","cited_arxiv_id":null,"evidence_quote":"The SAM-based text segmentation model fine-tuned to produce pseudo-GT masks for FTSR and Real-CE."},{"cited_title":"Rethinking text segmentation: A novel dataset and a text-specific refinement approach","cited_arxiv_id":null,"evidence_quote":"One of the two text segmentation datasets used to fine-tune SAM-TS for mask generation."},{"cited_title":"Bts: a bi-lingual benchmark for text segmentation in the wild","cited_arxiv_id":null,"evidence_quote":"The bilingual text segmentation dataset used alongside TextSeg to fine-tune SAM-TS."}],"review_version":1}