{"id":"4353ff66-f2f0-48bb-b5d8-3f87beb68f5b","arxiv_id":"2510.21590","paper_version":3,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"A two-stage 'text-first, image-later' super-resolution framework restores glyph structures before enhancing the whole image, improving OCR accuracy and visual quality on a new extreme-zoom Chinese text dataset.","lead":"TIGER is a two-stage super-resolution method that first reconstructs blurred text, then uses that text to guide full-image enhancement. It also introduces a new Chinese scene-text dataset with extreme zoom, reporting higher readability and image quality than prior models.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"UZ-ST alignment is the load-bearing risk: the central SOTA claim rests on a new, manually cleaned benchmark whose residual pixel alignment is never independently verified.","rationale":"The paper has real strengths: the two-stage idea is clearly motivated, the ablations isolate the text-mask contribution, and the null/random-OCR experiments show the method does not simply memorize OCR predictions. I am not raising a methodological objection to the comparison protocol; all methods appear to be fine-tuned on the same filtered benchmark. The most fragile point is the integrity of UZ-ST alignment. The paper itself acknowledges the difficulty ('severe misalignment', 'manual filtering') but does not provide an independent, human-verified error analysis. Aggregate alignment metrics (PSNR/NCC) conflate blur/resolution differences with misalignment and cannot detect local warp errors. This is not an external-consensus dispute; it is an internal-support gap: the dataset's validity is asserted rather than demonstrated. The reader's conditional verdict already identifies this; my analysis agrees. The proposed annotator-based residual test would either vindicate the benchmark or reveal that the quantitative claims need reinterpretation. I therefore recommend no change to the conditional verdict: accept after the alignment check (and ideally release of the test split/code) is performed.","tokens_in":18887,"tokens_out":8753,"duration_ms":87402,"concrete_test":"Select a random sample of ~100 UZ-ST pairs stratified by zoom mode (include 35 mm and 14 mm). For each pair, independently annotate 5–10 corresponding text stroke corners/edges in the aligned LR and the 200 mm GT; compute residual displacement after best-fit similarity/projective transform, and report median and 95th percentile residual per focal-length subset. A meaningful pass threshold is median residual ≤1 GT pixel (or ≤0.5 median stroke width). If residuals exceed this, re-run Table 2 and Table 3 on a human-verified aligned subset and check whether TIGER's OCR-A and image-quality margins persist; if they shrink or flip on that subset, the central claim needs to be downgraded. This test directly settles whether the UZ-ST benchmark, and the results derived from it, are trustworthy.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's strongest claim — TIGER breaks the readability/quality trade-off — is supported mainly by tables on Real-CE and the new UZ-ST benchmark. The UZ-ST leg is load-bearing because the dataset is both the training source and the evaluation instrument. Section 4 and Figure 3b describe a cascade coarse-to-fine alignment using homographies between focal lengths plus manual filtering/realignment; Table 8 reports PSNR, MSE, SSIM, NCC, and AKD. However, none of these metrics validates pixel alignment. Homography warping assumes a planar scene or zero-parallax camera motion; a handheld multi-lens phone capturing scenes with depth (signs, book covers, posters) violates this, producing locally varying residual misalignment. Manual filtering removes the worst cases but the retained pairs are not independently verified, and the supplementary alignment comparison is run on the 35mm subset only, not the hardest 14mm (×14.29) mode. Because Stage 2 is trained with pixel-level MSE/LPIPS losses against the 200mm GT, even subpixel misalignment can blur supervision and bias PSNR/SSIM; OCR-A is less sensitive but still affected by crop misalignment. If residual misalignment is non-negligible, the quantitative superiority in Tables 2–3 on UZ-ST, and the claim that the dataset is a reliable harder benchmark, may not hold.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes TIGER, a two-stage scene-text image super-resolution framework built on a 'text-first, image-later' paradigm. Stage 1 uses an OCR-conditioned latent diffusion model to restore glyph structure from cropped text regions and produces a text mask; Stage 2 uses a ControlNet-like network conditioned on the LR image and the predicted text mask to perform full-image super-resolution. The paper also introduces UZ-ST, a Chinese scene-text dataset with extreme zoom factors up to ×14.29, collected with a multi-focal-length phone camera and aligned by a cascade coarse-to-fine homography pipeline. Experiments on Real-CE and UZ-ST report state-of-the-art results in both image fidelity and OCR accuracy, and ablations study the role of the text mask, the OCR condition, and the UZ-ST training data. The central claim is that the two-stage decoupling breaks the usual trade-off between text readability and overall image quality.","tokens_in":19257,"tokens_out":6708,"duration_ms":69423,"significance":"If validated, TIGER would be a practically useful architecture for real-world scene-text SR, and UZ-ST would fill a clear gap as the first extreme-zoom, Chinese, multi-line text SR benchmark. The two-stage idea is well motivated, and the ablations are informative. In particular, the null-text / random-text experiments (Tables 5 and 10) are a strong and welcome response to the natural circularity concern: OCR-A remains high even when the OCR condition is uninformative, showing the model is not simply reading out the answer. The paper also reports extensive comparisons and states that code and data will be released. However, the central quantitative claims rest on the UZ-ST benchmark and on a modified Real-CE benchmark, and both raise methodological concerns that need to be resolved before the SOTA claim can be accepted.","major_comments":[{"comment":"The UZ-ST split, as described, appears to be at the image-pair level rather than the scene level: 'Among them, we randomly select 470, 589, and 581 pairs for evaluation under each zoom mode.' Since the same 200 mm image is the GT for the 14 mm, 35 mm, and 85 mm pairs of the same physical scene, a random split per zoom mode can place the same scene (and the same GT image) in both training and test sets. This would leak the exact target into training and inflate all UZ-ST numbers, especially OCR-A and PSNR. The paper must either demonstrate a scene-disjoint split or re-run the experiments on a properly split benchmark.","section":"§4, Dataset split"},{"comment":"The alignment validation is not sufficient for the load it carries. Table 8 reports PSNR/MSE/SSIM/NCC/AKD, but these are global similarity measures and do not verify pixel-level alignment; AKD=241.44 still suggests substantial residual keypoint displacement. The supplementary comparison is run only on the 35 mm subset, not on the hardest 14 mm (×14.29) mode. Because Stage 2 is trained with pixel-level MSE/LPIPS losses against the 200 mm GT, residual misalignment—especially depth-dependent parallax that homography cannot model—would blur supervision and bias every UZ-ST metric. Please provide per-zoom-mode residual alignment statistics, a human-verified subset, or an independent alignment check before using UZ-ST as the basis for the SOTA claim.","section":"§4, Figure 3b, Table 8"},{"comment":"The paper filters and reannotates Real-CE, using only 337 training and 188 testing pairs, but the main text says 'Evaluations are performed on the Real-CE Benchmark [31]' without qualification. This changed benchmark is not comparable to the original Real-CE test set used in prior work. The authors should clearly state in the main text that all comparisons use their filtered subset, report the filtering criteria, and preferably also report results on the original Real-CE test split so that readers can relate the numbers to prior publications.","section":"§5.1, §11"},{"comment":"All quantitative results are reported as single runs without error bars or significance tests. Some of the decisive margins are small (e.g., 0.41 dB PSNR on UZ-ST, 2.6 pp OCR-A on Real-CE), and OCR recognition is stochastic (PP-OCRv3 is applied to generated images, which may also be stochastic). Reporting mean ± std over at least a few runs, or a paired significance test for the main comparisons, would materially strengthen the 'state-of-the-art' claim.","section":"Tables 2–3"}],"minor_comments":[{"comment":"Focal length is inconsistently reported: the main text and Table 7 use 85 mm, while Section 10 says 84 mm. Please harmonize.","section":"§4, §10"},{"comment":"The dataset name is misspelled as 'UltraZoom-Sence Text' in the Section 4 heading; should be 'Scene'.","section":"§4"},{"comment":"ΔOCR-A is used in Table 3 but is never defined. Please define it explicitly, and state whether the LR baseline is the original LR input or the aligned LR used in the pipeline.","section":"Eq. (8), Table 3"},{"comment":"The text around Eq. (4) says the network 'denoises the zL only at the specific timestep t', but the formula is a single-step residual update. Please clarify the inference schedule and the role of the timestep t.","section":"Eq. (4)"},{"comment":"The sentence 'we integrate the outputs of MARCONet and DiffTSR with HAT-generated results to simulate real-world application scenarios' is important for fairness but appears only in the experimental section. Please state this in the protocol description and justify that the integration is not disadvantageous to those baselines.","section":"§5.2.1"}],"recommendation":"major_revision","confidential_remarks":"The paper has a sensible core idea and the OCR-dependence ablations are genuinely reassuring. The main risk is benchmark hygiene: the UZ-ST dataset is both the training source and the evaluation instrument, and the current description does not rule out scene-level train/test leakage or residual misalignment. Both issues are fixable within the manuscript's scope, so I do not recommend rejection, but the SOTA claim should not be accepted until the benchmark is shown to be sound. The filtered Real-CE comparison also needs to be clearly qualified in the main text."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: this is a solid paper, not a breakthrough. The two-stage \"restore text, then enhance image\" idea is genuinely different from TADiSR, DiffTSR, and MARCONet, and the paper does the work to show the text mask actually carries the weight. The ablation with null and random OCR text is the right check and the result is reassuring: even with garbage OCR conditioning, the model beats TADiSR. So the central mechanism isn't just reading the answer off the OCR branch.\n\nThe strongest evidence is the collection of ablations: empty mask, standard font, SAM-TS, and LDM guidance all hurt OCR-A relative to the full pipeline, and the gains are not tiny. The comparisons on Real-CE are extensive. I'd want error bars or at least multiple seeds before believing the margins against TADiSR, but that's a standard weakness in this subfield, not a reason to doubt the direction.\n\nThe real soft spot is UZ-ST. The dataset is the main contribution beyond the method, and its value depends on the LR-HR alignment being genuinely pixel-accurate. The paper describes a cascade homography plus manual filtering, and Table 8 shows their alignment beats RealSR and single-shot SIFT on PSNR/NCC/AKD. That helps, but those are aggregate measures against the same GT; they don't prove there's no residual local misalignment, especially in scenes with depth and at the 14mm extreme where homography assumptions stretch. Since stage 2 is trained with pixel-level losses against the 200mm GT, any residual blur or shift would leak into the supervision and inflate the quality gap. I'd want a small human-verified sample, or a release of the raw pairs so the community can check. The supplementary alignment comparison is only run on 35mm, not 14mm, which is the worst case. That's a real gap, and the paper should address it.\n\nAlso note the Real-CE filtering to 337 train / 188 test: reasonable and honestly reported, but it means the comparison is on a curated subset. Fine for a paper, worth flagging in review.\n\nWho's this for: people working on scene-text SR or real-world SR evaluation. It deserves a serious referee. I'd send it, with a request to verify UZ-ST alignment and add variance or a human eval. The method itself I'd cite once the dataset is out.","headline":"Two-stage text-first SR is a real idea with honest ablations; the new UZ-ST benchmark is the thing to pressure before trusting the SOTA numbers.","tokens_in":19743,"tokens_out":2000,"would_cite":true,"duration_ms":20889,"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":"Scene text super-resolution can restore readable text and high image quality at once by first reconstructing glyph structures, then using them to guide full-image enhancement.","keywords":["scene text super-resolution","glyph structure guidance","text-first image-later","two-stage diffusion","UZ-ST dataset","extreme zoom","OCR accuracy","image quality"],"falsifier":"Randomly sample UZ-ST pairs, warp the LR image to the HR frame using the paper's alignment, and have human annotators mark corresponding control points; if the average residual error exceeds a few pixels on a meaningful fraction of samples, the dataset's alignment guarantee—and thus the quantitative comparison—would need revision. A simpler check: retrain TIGER without UZ-ST and compare the OCR-A delta against the full version to isolate the dataset's contribution.","tokens_in":18817,"feed_emoji":"🔤","tokens_out":5967,"duration_ms":106780,"temperature":0.7,"pith_summary":"The paper tries to establish that a two-stage 'text-first, image-later' super-resolution framework can break the usual trade-off between readable text and overall image quality. TIGER first reconstructs precise glyph structures from low-resolution text regions, then feeds the recovered text mask as guidance into a full-image enhancement stage. On the existing Real-CE benchmark and their new UZ-ST dataset—the first Chinese scene-text benchmark with extreme zoom up to ×14.29—the authors report state-of-the-art OCR accuracy and image-quality scores. A sympathetic reader would care because real-world scene text is often severely degraded, and current generative methods tend to turn small Chinese characters into gibberish while making the background look plausible.","feed_headline":"Text-first super-resolution yields sharper glyphs and cleaner images","feed_subtitle":"TIGER restores glyph structure before enhancing the full image, raising OCR accuracy and image quality on extreme-zoom Chinese scene text.","key_machinery":"The load-bearing mechanism is the two-stage cascade with a glyph structure mask as the intermediate representation. Stage 1 (text restoration) is a latent diffusion model with two output branches—one for appearance, one for text mask—whose mask branch is supervised by a combined segmentation loss and refined through a two-phase training strategy that first learns from synthetic plus real data, then freezes the output heads and trains only on synthetic data to clean up noisy real-world masks. Stage 2 (image enhancement) is a ControlNet-like diffusion model conditioned on the low-resolution image and the restored text mask, trained with MSE, LPIPS, and Sobel edge losses. The paper also builds","core_discovery":"The central claim is that text and non-text should not be restored by one shared generative prior. TIGER explicitly decouples glyph restoration from image enhancement: a diffusion-based text restoration stage recovers stroke geometry in text regions, and the resulting glyph structure mask is injected as conditional guidance into a ControlNet-style full-image enhancer. The authors assert this ordering—text structure first, whole-image enhancement second—allows the enhancer to harmonize text with background without destroying strokes. They also present UZ-ST, a real-world dataset collected at four focal lengths (14mm to 200mm) with ×14.29 maximum zoom, containing 5,036 LR-HR pairs and 49,675 a","pith_inferences":["The 'text-first, image-later' ordering may transfer to other structure-critical restoration tasks—such as face or document enhancement—where a dedicated structure-recovery module's output is used as conditioning for a general enhancer.","The UZ-ST alignment pipeline's claim of accurate pixel alignment is the least supported link: an independent human-verified alignment sample would be needed to rule out misalignment inflating the reported gains.","The OCR-A improvements shrink dramatically at ×14.29 (16.0% vs 63.2% at ×2.35), so practical deployments should expect the method's advantage to fade at the most extreme zoom levels.","Because evaluation relies on a single OCR engine, readability may be under- or over-estimated; testing with multiple recognizers or human readers would give a more robust measure of the claimed readability gains."],"forward_implications":["Super-resolution systems for text-heavy scenes can adopt an explicit structure-restoration step rather than relying on a single generative prior, yielding both higher OCR accuracy and better perceptual image quality.","The UZ-ST dataset provides a benchmark for extreme zoom (×14.29) and multi-line Chinese text, enabling evaluation of SR methods under real-world degradations far beyond existing ×4 datasets.","The two-phase training schedule offers a recipe for using noisy real-world masks together with clean synthetic masks to train a structure-restoring diffusion model without costly pixel-level manual segmentation of degraded text.","Ablations show the method remains effective even with random or null OCR text input, implying the text restoration stage learns glyph structure from the low-resolution image itself rather than depending on the semantic condition."],"fun_headline_variants":["Two-stage SR: restore text first, then enhance image","Text-first SR resolves trade-off: crisp glyphs, clear images","Glyph structure first: TIGER sharpens text and image","Decouple text restoration from enhancement for sharper SR","First restore glyphs, then enhance: TIGER's two-stage SR"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The UZ-ST dataset is accurately pixel-aligned across all extreme-zoom pairs, so that each low-resolution image and its high-resolution ground truth truly depict the same scene; if residual misalignment remains, every UZ-ST training and evaluation number is compromised.","fun_headline_variants_meta":{"raw":{"variants":["Two-stage SR: restore text first, then enhance image","Text-first SR resolves trade-off: crisp glyphs, clear images","Glyph structure first: TIGER sharpens text and image","Decouple text restoration from enhancement for sharper SR","First restore glyphs, then enhance: TIGER's two-stage SR"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000711,"raw_usage":{"total_tokens":3001,"prompt_tokens":674,"completion_tokens":2327,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":418,"completion_tokens_details":{"reasoning_tokens":2240}},"tokens_in":418,"tokens_out":2327,"duration_ms":15384,"temperature":1.0,"reasoning_tokens":2240,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-04T08:13:35.581776+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Randomly sample UZ-ST pairs, warp the LR image to the HR frame using the paper's alignment, and have human annotators mark corresponding control points; if the average residual error exceeds a few pixels on a meaningful fraction of samples, the dataset's alignment guarantee—and thus the quantitative comparison—would need revision. A simpler check: retrain TIGER without UZ-ST and compare the OCR-A delta against the full version to isolate the dataset's contribution.","supporting_citations":[],"review_version":1}