{"id":"52d78116-aeb3-4231-ad53-3fd664ebbbfd","arxiv_id":"2507.04118","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"low","formal_verification":"none","parameter_count":4,"one_line_summary":"PromptSR uses cascaded cross-scale anchor prompts and category-based attention to enlarge the receptive field in lightweight image super-resolution, achieving state-of-the-art PSNR on most of five benchmarks.","lead":"PromptSR is a lightweight image super-resolution network that uses downscaled image features as \"anchor prompts\" to give every pixel access to global context without expensive full-resolution attention. It reports small but consistent PSNR gains over prior lightweight models, reaching performance close to a 12M-parameter model with only 0.78M parameters.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The claimed SOTA superiority is not established: PromptSR trails ATD-light on several benchmarks, and most wins are ≤0.03 dB with no code, seeds, or significance tests.","rationale":"The central claim is empirical: a 0.78M-parameter network outperforms existing lightweight SR methods. The evidence is Table I. Reading the table against the strongest baseline, ATD-light, shows the claim is not uniform: PromptSR is numerically worse on four dataset/scale cells (Set14 ×2, BSD100 ×2, BSD100 ×3, Set5 ×4) and essentially tied on several others. The larger gains appear only on Urban100 and Manga109. Since the paper does not report training variance or significance, and code is not yet public, the existing numbers are insufficient to support the strong 'outperforms state-of-the-art' phrasing. This is not a claim of dishonesty; it is a request to match the strength of the conclusion to the strength of the evidence. Eq. (8) is a concrete internal blocker: as written it is algebraically degenerate, so the anchor-update mechanism cannot be reimplemented exactly. Together these issues justify keeping the paper conditional, with the condition that the authors release code and provide significance or error-bar evidence, or soften the claim. This is closely related to, but slightly different from, the reader's weakest-assumption point about cross-scale similarity; the reader emphasized the mechanism, whereas the more load-bearing issue here is the robustness and verifiability of the reported quantitative superiority.","tokens_in":22546,"tokens_out":9689,"duration_ms":106659,"concrete_test":"Once the authors release code (as promised), retrain PromptSR and ATD-light under the identical protocol with 5 independent seeds and compute per-image PSNR/SSIM paired differences with a Wilcoxon signed-rank test on all five benchmarks. If PromptSR is not significantly better (p<0.05) on a majority of benchmarks, the abstract and conclusion should be revised to claim superiority only on self-similar datasets (Urban100, Manga109). Also ask the authors to replace Eq. (8) with the intended update P^i ← α P^{i-1} + (1−α) P_new^i and verify that the ablation in Table VII is reproduced under that reading.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's headline claim is that PromptSR 'outperforms state-of-the-art lightweight SR methods in quantitative, qualitative, and complexity evaluations.' Taking Table I at face value, this is not uniformly true: against the strongest published lightweight baseline, ATD-light, PromptSR is numerically worse on Set14 ×2 (34.10 vs 34.11), BSD100 ×2 (32.37 vs 32.39), BSD100 ×3 (29.30 vs 29.32), and Set5 ×4 (32.61 vs 32.62), and it ties on BSD100 ×4. Most favorable margins on other datasets are 0.01–0.05 dB; the larger gains are concentrated on Urban100 and Manga109 (e.g., 0.10–0.12 dB at ×2/×3), which are the two benchmarks with the strongest repetitive/self-similar structure and thus the most favorable setting for the cross-scale anchor assumption. Because no code, training seeds, or error bars are provided, a 0.01–0.03 dB advantage is within typical run-to-run variation and does not support a general SOTA claim. The self-referential error in Eq. (8), where P^i appears on both sides and algebraically forces P^i = P^{i-1}, further prevents faithful re-implementation of the anchor-update mechanism, so the reported numbers cannot currently be independently checked.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes PromptSR, a lightweight image super-resolution network built from cascade prompting blocks (CPBs). Each CPB contains a global anchor prompting layer (GAPL) that constructs low-dimensional anchor prompts from downscaled features via cross-scale attention, followed by two local prompting layers (LPLs) that use the GAPL's similarity maps to guide category-based and window-based self-attention. The authors claim state-of-the-art performance on five benchmarks with 0.78M parameters, and that the method matches a 12M-parameter SwinIR model. The paper includes complexity analysis, ablations, LAM visualizations, and a limitation discussion.","tokens_in":22854,"tokens_out":6974,"duration_ms":67060,"significance":"If the architecture works as described, the GAPL idea is a plausible way to achieve a global receptive field at near-window-attention cost, and the complexity bound in Table II is correctly derived. The ablation study is thorough and generally supports the contribution of each prompting component. However, the central empirical claim of state-of-the-art superiority is not substantiated by the reported numbers, and the anchor update formula in Eq. (8) is mathematically self-contradictory. These issues prevent me from recommending acceptance at this stage.","major_comments":[{"comment":"Equation (8) defines the anchor prompt update as P^i = αP^{i-1} + (1−α)P^i. Since P^i appears on both sides, subtracting (1−α)P^i from both sides gives αP^i = αP^{i-1}, and for α ≠ 0 this forces P^i = P^{i-1}. Thus the update as written is the identity and cannot compute the 'adaptive anchor prompt update' that the paper describes. Please rewrite with a distinct symbol for the newly constructed prompts (e.g., \\tilde P^i) and clarify the update schedule within each residual group.","section":"Section III-B, Eq. (8)"},{"comment":"The text states that 'our proposed model can consistently outperform existing methods on all five benchmark datasets,' but the table shows the opposite. For example, PromptSR is worse than ATD-light on Set14 ×2 (34.10 vs 34.11), BSD100 ×2 (32.37 vs 32.39), BSD100 ×3 (29.30 vs 29.32), and Set5 ×4 (32.61 vs 32.62), with a tie on BSD100 ×4. Because the favorable margins on other benchmarks are typically 0.01–0.03 dB, and no code, seeds, or repeated runs are provided, the claimed superiority is not established. I recommend either providing statistical significance tests (e.g., paired tests across multiple seeds) or revising the claim to 'comparable or better in most cases.'","section":"Section IV-B, Table I"}],"minor_comments":[{"comment":"The parameter counts for different downscale ratios d are unexpected: d=2 gives 534K, d=4 gives 540K, and d=8 gives 562K. Since the linear projections in the GAPL are applied with shared weights, the number of anchors should not change the parameter count. Please clarify why the parameter counts differ, or correct the table.","section":"Section IV-D, Table V"},{"comment":"The text says that α=0 'resulted in a significant drop in performance,' but the table shows only a 0.01 dB drop on Urban100 (26.62 vs 26.63) and a 0.12 dB drop on Manga109 (30.86 vs 30.98). The characterization 'significant' is an overstatement and should be adjusted.","section":"Section IV-D, Table VI"},{"comment":"The expression M_fine = Q(Kp + A)^T / √C has ambiguous parentheses; it should be written as M_fine = Q (Kp + A)^T / √C to make clear that the transpose applies to the sum.","section":"Section III-B, Eq. (10)"},{"comment":"The paper uses the names 'OminiSR' and 'OmniSR' inconsistently (e.g., in the quantitative results paragraph). Please standardize to 'OmniSR' as in the reference list.","section":"Section IV-B"},{"comment":"The paper mentions 'sub-category size in CSA' in the implementation details but never defines this concept in the method section. Please add a definition or explanation of how the sub-category size is used in category-based self-attention.","section":"Section III-C"}],"recommendation":"major_revision","confidential_remarks":"The manuscript's headline claim needs to be reconciled with its own Table I; the reported results do not support 'consistently outperform.' The self-referential update equation in Eq. (8) is a clear technical error that should be caught before publication. If the authors fix the equation and temper the empirical claims, the paper could be a reasonable contribution to lightweight SR. The t-SNE analysis and LAM visualizations are nice touches, but the quantitative evidence is currently too thin to support the asserted superiority."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The core idea here is worth taking seriously: downscaled features as anchors to build global prompts, then feeding the resulting attention maps back into local prompting layers. That composition is new, even though each piece borrows from prior work (downscaled anchors/cross-attention from [19], category-based attention from [41]). The complexity analysis is correct and the ablations are consistent with the design story. I also appreciate the t-SNE analysis and the honest limitations paragraph admitting that the anchor prompts filter out semantic information—that is the kind of self-assessment we want.\n\nThe soft spots are real but fixable. First, the headline claim that the method \"outperforms state-of-the-art lightweight SR methods\" is not supported by the paper's own Table I. Against ATD-light, PromptSR is numerically worse on Set14 ×2, BSD100 ×2, BSD100 ×3, and Set5 ×4, and ties on BSD100 ×4. Most favorable margins elsewhere are 0.01–0.05 dB, which without code, seeds, or error bars is within run-to-run variation. The larger gains appear precisely on Urban100 and Manga109, the benchmarks with the most repetitive structure—exactly where the cross-scale anchor assumption is strongest. That weakens the generality of the claim. Second, Eq. (8) is written as P_i = αP_{i-1} + (1−α)P_i, which is self-referential and algebraically forces P_i = P_{i-1} for any nonzero α. As written, the anchor update does nothing; presumably the intended right-hand side uses the freshly constructed prompts, but as printed the mechanism cannot be reimplemented. That is a load-bearing notation error. Minor issues: the text says \"OminiSR\" at one point, and Table III shows the method is not the cheapest in multiply-adds, so the efficiency edge is mainly parameter count.\n\nThis is a competent, well-structured paper, not a desk reject. The architecture is plausible and the ablations are informative. I would send it to a serious referee, with the expectation that the authors fix the update equation, soften the absolute superiority claim, and - ideally - release the code. Without those, the quantitative contributions cannot be independently checked. So: accept for review, major revision.","headline":"A solid lightweight SR architecture with a genuinely novel cascading-prompt idea, but the SOTA claim is not backed by its own Table I and Eq. (8) has a self-referential bug that must be fixed before the numbers can be trusted.","tokens_in":23362,"tokens_out":1696,"would_cite":false,"duration_ms":19434,"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":"A lightweight super-resolution network with 0.78M parameters can match the reconstruction quality of an 11.9M-parameter SwinIR by using downscaled image features as global anchor prompts.","keywords":["image super-resolution","lightweight model","anchor prompts","cross-scale attention","window-based self-attention","prompt learning","receptive field","transformer"],"falsifier":"Run PromptSR and a window-only baseline on images engineered to lack cross-scale similarity, such as white noise or aperiodic stochastic textures; if the anchor prompts stop helping, the cross-scale assumption is doing the work. A cleaner control is to replace anchors $A$ with random learned vectors of the same shape and check whether the downscaled-image anchors still win.","tokens_in":22381,"feed_emoji":"🖼️","tokens_out":7074,"duration_ms":64203,"temperature":0.7,"pith_summary":"The paper tries to show that a lightweight image super-resolution network can escape the small-receptive-field trap of window-based attention by using the image's own downscaled version as a set of anchors. Cross-attention to those anchors builds cheap global prompts, so every pixel can draw on the whole image for about one sixty-fourth of the cost of full self-attention. If correct, near-top-tier reconstruction quality no longer requires an 11.9-million-parameter model; a 0.78-million-parameter network can deliver comparable results, which matters for phones and embedded cameras.","feed_headline":"0.78M-parameter SR model matches 12M-parameter SwinIR","feed_subtitle":"Prompting with downscaled anchors widens the receptive field of lightweight super-resolution without quadratic attention cost.","key_machinery":"The cascade prompting block (CPB) is the unit that carries the argument. Its global anchor prompting layer (GAPL) builds anchor prompts from downscaled features and propagates global information through two cross-attention steps, while its two local prompting layers (LPLs) combine window-based self-attention with category-based self-attention guided by the GAPL's attention maps. An anchor-prompt update rule, $P_i = \\alpha P_{i-1} + (1-\\alpha) P_i$, carries consistent global priors across blocks inside a residual group while blocking cross-group mixing.","core_discovery":"The central claim is that natural-image cross-scale similarity is enough to carry global context in lightweight super-resolution. PromptSR downscales the input feature by a factor of $d=8$ to create anchors $A$, enhances them with cross-attention over the full key/value space, and uses the resulting anchor prompts $P$ to prompt the full-resolution features. The two local prompting layers then reuse the coarse and fine similarity maps from this global step to group tokens into irregular windows for category-based self-attention, refining details while staying globally informed. On five public benchmarks the model outperforms state-of-the-art lightweight methods with under 780K parameters, and the authors report that it performs comparably to SwinIR (11.9M parameters) at $\times4$ on Urban100 and Manga109.","pith_inferences":["If cross-scale similarity is doing the work, PromptSR should lose its edge on images with no self-similar structure, such as white noise or aperiodic stochastic textures; this is a testable consequence the paper does not run.","The paper's own t-SNE analysis shows the anchor prompts discard some semantic information, so coupling anchors with a semantic classifier could push lightweight super-resolution further.","Because the prompting module is a drop-in block around attention, the same CPB design could be transplanted to other transformer-based restoration tasks, such as denoising or deblurring, that also rely on global context.","The downscale ratio $d=8$ is a sweet spot in the paper's ablations; the behavior of larger ratios on higher-resolution inputs is unexplored and could sharpen the cost-performance trade-off."],"forward_implications":["With PromptSR, lightweight super-resolution models can claim a global receptive field at the same order of computational cost as window-based attention, not quadratic in image size.","A 0.78M-parameter network is reported to match SwinIR (11.9M parameters) on Urban100 $\times4$ and Manga109 $\times4$, so high-quality super-resolution becomes plausible on resource-constrained devices.","The gains are concentrated on structured, repetitive textures where global context matters most, such as Urban100 and Manga109, while gains on simpler natural-image benchmarks are smaller.","Ablations show the three prompting stages are additive: anchor prompting helps, coarse prompting helps more, and fine prompting adds the rest, so the coarse-to-fine cascade is load-bearing."],"supporting_citations":[{"why":"Supplies the window-based attention baseline and the large-model comparison (SwinIR at 11.9M parameters) that PromptSR claims to match.","marker":"[8]"},{"why":"Introduces the cross-scale hierarchy idea and the downscaled-features-as-anchors construction that the GAPL adapts.","marker":"[19]"},{"why":"OmniSR is a principal state-of-the-art lightweight baseline that PromptSR outperforms in PSNR/SSIM and computational cost.","marker":"[13]"},{"why":"HPINet is the matching-window lightweight baseline used for quantitative, visual, and local-attribution-map comparisons.","marker":"[14]"},{"why":"ATD-light is the strongest lightweight competitor, using an adaptive token dictionary for global information; PromptSR compares against it and reports failure cases against it.","marker":"[41]"},{"why":"Vision Transformer self-attention provides the vanilla complexity baseline in the theoretical GAPL analysis.","marker":"[34]"}],"fun_headline_variants":["PromptSR: cascaded prompts widen receptive field cheaply","Downscaled anchors give lightweight SR global context","PromptSR: 0.78M params, SwinIR-class SR via cascade prompts","Cross-scale anchors prompt lightweight SR to global reach","Cascade prompting lifts SR quality at fraction of cost"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The whole scheme rests on natural images being self-similar across scale, so that an 8x-downscaled version of the image still contains the structure needed to guide the full-resolution reconstruction.","fun_headline_variants_meta":{"raw":{"variants":["PromptSR: cascaded prompts widen receptive field cheaply","Downscaled anchors give lightweight SR global context","PromptSR: 0.78M params, SwinIR-class SR via cascade prompts","Cross-scale anchors prompt lightweight SR to global reach","Cascade prompting lifts SR quality at fraction of cost"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000389,"raw_usage":{"total_tokens":2080,"prompt_tokens":1006,"completion_tokens":1074,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":622,"completion_tokens_details":{"reasoning_tokens":991}},"tokens_in":622,"tokens_out":1074,"duration_ms":8692,"temperature":1.0,"reasoning_tokens":991,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T19:55:00.770386+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run PromptSR and a window-only baseline on images engineered to lack cross-scale similarity, such as white noise or aperiodic stochastic textures; if the anchor prompts stop helping, the cross-scale assumption is doing the work. A cleaner control is to replace anchors $A$ with random learned vectors of the same shape and check whether the downscaled-image anchors still win.","supporting_citations":[{"cited_title":"Swinir: Image restoration using swin transformer,","cited_arxiv_id":null,"evidence_quote":"Supplies the window-based attention baseline and the large-model comparison (SwinIR at 11.9M parameters) that PromptSR claims to match."},{"cited_title":"Efficient and explicit modelling of image hierarchies for image restoration,","cited_arxiv_id":null,"evidence_quote":"Introduces the cross-scale hierarchy idea and the downscaled-features-as-anchors construction that the GAPL adapts."},{"cited_title":"Omni aggregation networks for lightweight image super-resolution,","cited_arxiv_id":null,"evidence_quote":"OmniSR is a principal state-of-the-art lightweight baseline that PromptSR outperforms in PSNR/SSIM and computational cost."},{"cited_title":"From coarse to fine: Hierarchical pixel integration for lightweight image super-resolution,","cited_arxiv_id":null,"evidence_quote":"HPINet is the matching-window lightweight baseline used for quantitative, visual, and local-attribution-map comparisons."},{"cited_title":"Transcending the limit of local window: Advanced super-resolution transformer with adaptive token dictionary,","cited_arxiv_id":null,"evidence_quote":"ATD-light is the strongest lightweight competitor, using an adaptive token dictionary for global information; PromptSR compares against it and reports failure cases against it."}],"review_version":1}