{"id":"2e0edf02-7804-4cec-a0de-52b499aa0a16","arxiv_id":"2608.04525","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"A shared transformer trained with continuous flow matching for images and discrete diffusion for text jointly restores scene text images and reads out their characters, removing the external OCR prior.","lead":"DualTSR is a unified model for scene text image super-resolution that restores sharp images and predicts the text at the same time, using one shared transformer instead of external OCR modules. The authors report the best perceptual and text fidelity scores among compared methods on synthetic and real Chinese text benchmarks, with far lower latency and parameter count than DiffTSR.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The SOTA comparison rests on a re-implemented CTR-TSR split and a self-curated 300-pair RealCE subset with no released sample lists; until official splits are used or artifacts are released, the headline gains are not independently verifiable.","rationale":"The reader's weakest assumption is exactly the benchmark comparability issue, and I agree. The paper is internally coherent in its architecture and loss design; the equations describe a plausible coupled flow-matching/discrete-diffusion framework. However, the empirical payoff, the SOTA claim, depends on comparisons that are not yet independently checkable. The paper itself flags the missing official splits in Appendices A and E, so this is a self-acknowledged limitation, not a hidden one. I also noticed a secondary internal inconsistency: in Algorithm 1, the rectified target u_target is computed once with text conditioning and then used for both conditional and dropout samples, whereas Eq. (10) implies the unconditional target should be the plain velocity u_t. This is worth correcting and could affect reproducibility of the training procedure, but the benchmark comparability issue is more load-bearing because it directly bears on whether the reported numbers support the headline claim. My verdict therefore remains CONDITIONAL; the paper should be accepted only if the authors release the exact splits and sample lists and ideally re-verify on the original DiffTSR splits.","tokens_in":17373,"tokens_out":11015,"duration_ms":124929,"concrete_test":"Obtain the official DiffTSR CTR-TSR test split and the official RealCE subset from the DiffTSR authors, or ask them to run their released checkpoints on their original splits; then evaluate DualTSR on those exact splits and compare. Additionally, bootstrap the RealCE FID over the 300 pairs (e.g., 1000 resamples) and report confidence intervals; if the FID gap between DualTSR and DiffTSR is within the bootstrap interval, the RealCE perceptual-quality claim is not significant.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The headline claim that DualTSR outperforms all compared methods is only as strong as the benchmarks it is measured on, and both benchmarks are non-standard. For CTR-TSR (Appendix A), the authors re-implemented the DiffTSR construction pipeline because the official split is not public, then evaluate DiffTSR, MARCONet, and MARCONet++ official checkpoints on their re-created split. Any deviation in filtering criteria, canonical resizing, or the blind degradation recipe can shift ACC/FID for the baselines and alter the reported 12.78 pp ACC advantage. For RealCE (Appendix E), the evaluation uses a self-curated 300-pair 'clean, well-aligned' subset; the exact sample list is not released, and FID on 300 images is high-variance, so the reported RealCE FID/ACC gaps could be within subset-selection noise. The paper provides no overlap check with the original DiffTSR splits and no code, checkpoints, or sample lists. This is an empirical verification gap, not an internal inconsistency: if the artifacts are released and the numbers hold on official splits, the central claim would be supported.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes DualTSR, a scene text image super-resolution method that couples conditional flow matching for continuous image latents with absorbing-state discrete diffusion for text prediction inside a single multimodal transformer. The training objective combines image flow matching with model guidance, a text NELBO term, and a synchronized joint corruption loss; inference updates image and text states jointly without an external OCR prior. The authors report state-of-the-art FID, LPIPS, ACC, and NED on a re-implemented CTR-TSR benchmark, the best FID/ACC/NED on a self-curated 300-pair RealCE subset, and large efficiency gains over DiffTSR.","tokens_in":17543,"tokens_out":8926,"duration_ms":82607,"significance":"The conceptual contribution—unifying image restoration and text prediction in one shared transformer without an external OCR prior—is timely and the formulation is internally consistent. The two generative objectives are standard, the loss design is supported by the ablation in Table 5, and the efficiency comparison in Table 3 is clearly specified. However, the empirical headline rests on two non-standard benchmarks whose exact composition is not released, so the reported gains cannot currently be independently verified. If the artifacts are released and the numbers hold on the official splits, this would be a solid contribution to STISR.","major_comments":[{"comment":"The CTR-TSR benchmark used for the headline SOTA claim is a re-implementation of DiffTSR's private construction pipeline, and official DiffTSR/MARCONet/MARCONet++ checkpoints are evaluated on this re-created split. The paper states in Appendix A that the exact DiffTSR split is not publicly released, and while the filtering/resizing/degradation criteria are listed, no verification is provided that the resulting 64,139/8,690 split matches the original in image identities and degradation draws. Any deviation can shift ACC and FID for the baselines, so the 12.78 pp ACC advantage (Table 3) and the FID comparisons in Table 1 are not independently verifiable. The authors should release the exact sample list and construction code as part of the revision, and provide an overlap or distributional check against the original DiffTSR split if it can be obtained.","section":"Appendix A, Section 4"},{"comment":"The RealCE evaluation is performed on a self-curated subset of 300 clean, well-aligned LR-HR pairs with no released sample list. FID computed on 300 images is high-variance, and the reported ACC advantages are small (e.g., Table 2, x4 ACC 62.20% for DualTSR vs. 61.52% for MARCONet), so subset-selection noise could change the ranking. The claim of best FID, ACC, and NED on RealCE is therefore not robustly established. Please release the exact sample list, add bootstrap confidence intervals for the 300-image metrics, and compare the subset with the one used by DiffTSR if it can be obtained.","section":"Appendix E, Table 2"},{"comment":"For MARCONet++ on RealCE, Appendix D states that when the model fails to detect text, inference raises an error and the LR image is used directly as the model output, and that this issue is especially common in real-world samples. This fallback can substantially deflate MARCONet++'s RealCE numbers and inflate DualTSR's relative advantage. The paper should report the number of fallback cases per benchmark and either exclude them from the comparison or handle them in a way that does not bias the results.","section":"Appendix D, Table 2"},{"comment":"The full loss configuration in Table 5 row (c), trained for 300k iterations at batch size 128, reports FID 9.92 on CTR-TSR x4, whereas the main model in Table 1, trained for 700k iterations at batch size 32, reports FID 16.42 for the same benchmark. ACC/NED also differ (53.71/73.60 vs. 57.65/76.64). Please specify the evaluation protocol used for Table 5 (sampling steps, random seed, evaluation set) so that the discrepancy can be reconciled; as written, the absolute FID values appear inconsistent with the main table.","section":"Tables 1 and 5"}],"minor_comments":[{"comment":"Algorithm 2 has default steps=50, but Section 4 and Table 3 describe a 4-step ODE sampler as the default; please align the pseudocode with the main text.","section":"Algorithm 2, Section 4"},{"comment":"Figure 1 contains two panels labeled '(b)' (DiffTSR and Ours); the second panel should be labeled '(c)'.","section":"Figure 1"},{"comment":"There are several language issues: 'A detailed description can be find at Appendix B' and 'We also provide a effect of sampling on Appendix C' should be corrected.","section":"Section 3, Appendix C"},{"comment":"Appendix C selects 4 sampling steps as the default operating point, but Figure 5 indicates FID continues to improve up to about 40 steps; please report the FID/NED values at 4 steps in the figure or table so that the chosen trade-off is quantified.","section":"Appendix C"},{"comment":"Appendix E says the exact RealCE sample list will be released; if it is available at revision time, include it as supplementary material rather than deferring the release.","section":"Appendix E"}],"recommendation":"major_revision","confidential_remarks":"The paper is transparent about the provenance of its benchmarks, and the core method is plausible; the main issues are verification gaps rather than technical errors. I recommend major revision rather than rejection, because releasing the sample lists and clarifying the evaluation protocol would directly address the concerns."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: this is a genuinely new joint formulation for STISR, and the efficiency gains are real, but the headline numbers sit on a re-implemented benchmark and a self-curated subset that haven't been independently checked.\n\nThe new thing: DualTSR couples conditional flow matching for image latents and absorbing-state discrete diffusion for text tokens in a single shared MM-DiT-style transformer, with synchronized corruption and joint inference. That specific combination is not in prior work. The paper is straightforward about the ingredients—each is standard—but the integration is the contribution, and the ablations in Table 5 make a credible case that the joint loss and the modality-specific losses each earn their place.\n\nWhat's done well: the method is clearly specified, the training and inference algorithms are spelled out, and the efficiency comparison is striking: 203M parameters, 132ms latency, 4 steps, versus DiffTSR's 1.23B, 13.3s, 200 NFE, while improving ACC by 12.78 points at x4. The PSNR trade-off is acknowledged as perception-distortion, which is fine for this kind of generative method.\n\nThe soft spot is the evaluation protocol. The CTR-TSR split is re-implemented because the official one isn't public; the RealCE evaluation uses a self-curated 300-pair subset. Official checkpoints of the baselines are run on these self-built sets, so the comparison is only as fair as the re-implementation is faithful. The paper gives the filtering criteria and degradation recipe, but there is no overlap check with the original splits, and the exact sample list is not yet released. FID on 300 images is high-variance, so the RealCE gaps could be within subset-selection noise. None of this is an internal inconsistency; it's an empirical verification gap. The architecture and losses are internally consistent and the reported gains are large enough to matter.\n\nOne minor thing: for MARCONet++ failure cases, the authors use the LR input directly as output when detection fails, which could understate that baseline; the paper notes it, but it's still a small unfairness.\n\nBottom line: this is a serious, well-executed paper with a novel technical contribution. The central claim is plausible but not yet independently verifiable. It deserves a careful referee. If the artifacts are released and the numbers hold on official splits or on a release of their exact splits, it will be a strong contribution. I would send this to review with a request for the authors to provide the sample lists and any compatibility checks with the original DiffTSR split.","headline":"A novel joint continuous-discrete generation framework for STISR with real efficiency gains, but the headline SOTA numbers depend on benchmark splits that need independent verification.","tokens_in":18147,"tokens_out":2573,"would_cite":true,"duration_ms":24140,"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":"A single shared multimodal transformer that jointly runs conditional flow matching for the image and absorbing-state discrete diffusion for the text achieves the best recognition and perceptual scores on Chinese scene text…","keywords":["scene text image super-resolution","conditional flow matching","discrete diffusion","multimodal transformer","Chinese scene text","OCR-free text prior","generative super-resolution","joint image-text generation"],"falsifier":"Re-running the official DiffTSR, MARCONet, and MARCONet++ checkpoints on the exact CTR-TSR test set and the exact 300 RealCE pairs that the authors release, with the same TransOCR evaluator, and checking whether DualTSR's reported ACC and FID margins, including the 12.78-point $\\times 4$ ACC gain, survive, would settle the central claim.","tokens_in":17111,"feed_emoji":"🔤","tokens_out":16762,"duration_ms":144322,"temperature":0.7,"pith_summary":"Scene text super-resolution must produce an image that looks like real text and a character sequence that can be read, and today's systems usually meet that need by attaching a pretrained OCR recognizer to a super-resolution network, so recognition mistakes leak into the restored glyphs and the pipeline has many stages. DualTSR is an attempt to show that one shared multimodal transformer can do both jobs: a conditional flow matching head restores the continuous image latent while an absorbing-state discrete diffusion head reconstructs the text tokens, and the two processes condition each other at every layer instead of meeting at a single fusion point. At inference no external OCR prior is used, and the paper reports the best perceptual scores (FID, LPIPS) and text-fidelity scores (ACC, NED) among the compared methods on the Chinese CTR-TSR benchmark at $\\times 2$ and $\\times 4$, plus the best FID, ACC, and NED on a curated RealCE subset. If the comparisons hold, the practical consequence is a much cheaper pipeline: at $\\times 4$, 203M parameters and 132 ms latency replace DiffTSR's 1.23B parameters and 13.3 s while recognition accuracy rises by 12.78 percentage points.","feed_headline":"DualTSR lifts scene text super-resolution accuracy","feed_subtitle":"Restores Chinese text at 4x in 132ms with higher accuracy and far fewer parameters than OCR-prior diffusion models.","key_machinery":"The load-bearing mechanism is a shared multimodal transformer with a joint attention block in the style of MM-DiT: image latent tokens and text tokens are projected into modality-specific streams, concatenated into shared query, key, and value attention at every layer, and then split back, so neither modality passes through a separate fusion module. Two coupled generative processes attach to this backbone. Conditional flow matching treats the high-resolution image latent as the endpoint of a linear interpolation $x_t = (1-t)x_0 + t x_1$ and trains a velocity head to predict the constant velocity $x_1 - x_0$, with a rectified target supplied by an EMA teacher through model-guided training. Absorbing-state discrete diffusion treats text as tokens in a continuous-time Markov chain whose forward process progressively replaces tokens with a mask, and the text head is trained on a simplified negative variational lower bound to predict the clean token from a corrupted sequence using a log-linear noise schedule $\\alpha_t = 1 - t$. Synchronized corruption presents image and text at the same timestep in the joint loss, forcing the two heads to cooperate; at inference a four-step Euler ODE for the image alternates with iterative unmasking for the text, keeping already generated tokens fixed.","core_discovery":"DualTSR's central claim is that scene text image super-resolution is better formulated as coupled continuous-discrete generation than as image restoration steered by an external text prior. The paper learns the conditional joint distribution $p(x_{\\mathrm{hr}}, x_{\\mathrm{txt}} | x_{\\mathrm{lr}})$ with a single multimodal transformer: conditional flow matching generates the high-resolution image latent by integrating a velocity field along a linear interpolation from noise to data, and absorbing-state discrete diffusion reconstructs the character sequence by progressively unmasking tokens. Image and text share the same timestep during synchronized corruption, so the model must recover both modalities at once and each can guide the other; at inference, alternating Euler steps for the image and reverse absorbing transitions for the text let the evolving text hypothesis refine glyph structure while the evolving image improves text prediction. The paper reports that this unified model, without any external OCR prior, achieves the best FID, LPIPS, ACC, and NED among compared methods on CTR-TSR at $\\times 2$ and $\\times 4$, and the best FID, ACC, and NED with competitive LPIPS on the aligned RealCE subset, while cutting DiffTSR's $\\times 4$ inference cost from 1.23B parameters and 13.3 s to 203M parameters and 132 ms.","pith_inferences":["Beyond the paper: the headline comparison is only as clean as the re-implemented CTR-TSR split and the 300-pair RealCE subset, so the released sample lists deserve independent scrutiny; a distributional overlap check against the original DiffTSR splits would settle how much of the reported margin is benchmark construction rather than model behavior.","Beyond the paper: the guidance-scale result, where stronger classifier-free guidance monotonically worsens FID and eventually NED, suggests stroke-sensitive tasks need weaker guidance than natural-image synthesis; a testable extension is a modality-specific or time-varying guidance schedule that preserves fine strokes.","Beyond the paper: because the same backbone both reads and renders text, the coupled continuous-discrete design is a natural template for other joint perception-generation tasks, such as document restoration, deblurring, or multilingual text enhancement, where the recognition stream can regularize the image stream."],"forward_implications":["External OCR priors and separate image-text backbones are not necessary for strong scene text super-resolution; the internal text hypothesis generated by the shared transformer supplies the semantic constraint, and the ablation shows recognition accuracy climbing from 49.85% to 53.71% ACC when the joint and modality-specific losses are combined.","At $\\times 4$, generative scene text super-resolution can run inside feed-forward latency (132 ms, four network function evaluations) instead of the 13.3 s reported for DiffTSR, so diffusion-based restoration becomes usable in latency-sensitive settings.","The text branch is a usable recognition signal on its own, not just a training regularizer: on CTR-TSR at $\\times 4$ its internal predictions reach 45.44% ACC versus 26.09% for DiffTSR's dedicated text module, and OCR-on-restored-image ACC reaches 57.65%.","The perception-distortion trade-off is explicit: PSNR is lower than regression-based baselines, so applications that require pixel-level fidelity rather than perceptual realism and legibility should expect a different operating point."],"supporting_citations":[{"why":"Supplies the DiffTSR baseline, the CTR-TSR construction recipe the authors re-implement, and the official checkpoint compared at $\\times 2$ and $\\times 4$.","marker":"(Zhang et al. 2024)"},{"why":"Provides conditional flow matching, the image-generation objective with the linear interpolation path used in the velocity prediction.","marker":"(Lipman et al. 2023)"},{"why":"Introduces absorbing-state discrete diffusion in continuous time, the text-corruption process used for token unmasking.","marker":"(Lou, Meng, and Ermon 2023)"},{"why":"Gives the simplified NELBO objective and the log-linear noise schedule $\\alpha_t = 1 - t$ used for text denoising.","marker":"(Sahoo et al. 2024)"},{"why":"Supplies the MM-DiT joint attention block that lets image and text tokens share query, key, and value attention inside the unified backbone.","marker":"(Esser et al. 2024)"},{"why":"Provides the CTR corpus used to build CTR-TSR and the TransOCR recognizer used to compute ACC and NED.","marker":"(Yu et al. 2021)"},{"why":"Motivates the EMA-teacher model-guided training target used to rectify the flow-matching velocity.","marker":"(Tang et al. 2025)"},{"why":"MARCONet baseline with a learned character-structure codebook that DualTSR is compared against.","marker":"(Li, Zuo, and Loy 2023)"},{"why":"MARCONet++ baseline with enhanced character priors compared under official checkpoints.","marker":"(Li, Zuo, and Loy 2025)"},{"why":"BSRGAN degradation recipe that, together with Real-ESRGAN, defines the blind LR synthesis for CTR-TSR and RealCE evaluations.","marker":"(Zhang et al. 2021)"}],"fun_headline_variants":["DualTSR: unified text-image SR in 132ms","DualTSR without OCR priors tops scene text SR","Coupled diffusion: DualTSR speeds text SR 100x","DualTSR: 6x faster, 6x smaller, more accurate","No OCR prior: DualTSR redefines scene text SR"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The comparison hinges on the assumption that the authors' re-implementation of the DiffTSR CTR-TSR construction and their self-curated 300-pair RealCE subset are faithful enough to the original benchmarks that official checkpoints of DiffTSR, MARCONet, and MARCONet++ can be compared fairly against DualTSR.","fun_headline_variants_meta":{"raw":{"variants":["DualTSR: unified text-image SR in 132ms","DualTSR without OCR priors tops scene text SR","Coupled diffusion: DualTSR speeds text SR 100x","DualTSR: 6x faster, 6x smaller, more accurate","No OCR prior: DualTSR redefines scene text SR"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000189,"raw_usage":{"total_tokens":1383,"prompt_tokens":1038,"completion_tokens":345,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":654,"completion_tokens_details":{"reasoning_tokens":251}},"tokens_in":654,"tokens_out":345,"duration_ms":4039,"temperature":1.0,"reasoning_tokens":251,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T22:50:00.604158+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Re-running the official DiffTSR, MARCONet, and MARCONet++ checkpoints on the exact CTR-TSR test set and the exact 300 RealCE pairs that the authors release, with the same TransOCR evaluator, and checking whether DualTSR's reported ACC and FID margins, including the 12.78-point $\\times 4$ ACC gain, survive, would settle the central claim.","supporting_citations":[],"review_version":1}