{"id":"66bf6ca2-41cc-4038-85e0-43b7ad9a7f54","arxiv_id":"2506.12475","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"SDAN, a 0.41M-parameter super-resolution network pairing star-operation feature distillation with grouped multi-scale large kernel attention, reports small PSNR/SSIM improvements over recent lightweight SR baselines.","lead":"A lightweight image super-resolution network called SDAN combines a star-operation distillation module with a multi-shape multi-scale large kernel attention module, reporting PSNR gains on standard benchmarks with about 0.4 million parameters. The paper is an incremental architecture contribution to the crowded lightweight SR field, with modest gains over prior methods and no released code.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Uncontrolled comparison: SDAN's reported 0.02–0.20 dB gains over lightweight baselines (Table V) may stem from its stronger training recipe (ADAN, LR 5e-3, 1e6 iters) rather than the proposed modules; a same-protocol retraining of baselines is required.","rationale":"The reader's weakest assumption already identified the lack of controlled comparison; I agree and elevate it to the single most load-bearing issue because the claimed margins are so small that protocol differences can fully account for them. The paper's own §IV.A uses a distinctive training recipe, while the published baselines in Table V were not trained under that recipe. Without re-training baselines under identical conditions, the central claim is an uncontrolled comparison. The test-set-based hyperparameter selection in §IV.B compounds this by inflating SDAN's own numbers, but the decisive experiment is the controlled re-training. I therefore keep the reader's CONDITIONAL verdict: the paper is plausible and the tables are internally consistent, but acceptance should require the controlled comparison (and ideally a code release).","tokens_in":17923,"tokens_out":13360,"duration_ms":138605,"concrete_test":"Re-train the strongest lightweight baselines from Table V (OSFFNet and FIWHN) from scratch under the exact SDAN training protocol: DF2K training data, 64×64 random patches, ADAN with β1=0.98, β2=0.92, β3=0.99, EMA=0.999, constant LR=5e-3, 1e6 iterations, and identical evaluation (YCbCr PSNR/SSIM on the five benchmarks). If either baseline matches or exceeds SDAN's reported PSNR/SSIM, the claimed architectural superiority is not established; if the gap persists for both, the training-protocol concern is settled.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The most load-bearing concern is that the reported superiority of SDAN is not attributable to the proposed architecture because the comparison with prior methods is uncontrolled. In Table V, SDAN's largest margins over the strongest lightweight baselines at ×4 are only 0.02–0.20 dB PSNR (e.g., 32.54 vs. OSFFNet's 32.39 on Set5; 27.74 vs. PFFN/CFDN's 27.72 on BSD100). These margins are within the range that can be produced by training-protocol differences alone. Section IV.A specifies that SDAN uses the ADAN optimizer (β1=0.98, β2=0.92, β3=0.99), EMA=0.999, a constant learning rate of 5e-3, and 1e6 iterations, whereas the published baseline numbers were obtained with their own optimizers, schedules, and iteration counts. No baseline is re-trained under the SDAN recipe, and no ablation isolates the training protocol from the architectural modules. In addition, §IV.B selects the SDM kernel size on Manga109 (×3) and the MM-LKA kernel sizes on Set5 (×4), the same benchmarks on which final results are reported, so those numbers carry a positive selection bias. Consequently, the central claim that SDM and MM-LKA cause the gains remains untested; the evidence supports only that this training recipe plus this architecture reaches the reported numbers.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript proposes SDAN, a lightweight single-image super-resolution network built from a Star Distillation Module (SDM) and a Multi-shape Multi-scale Large Kernel Attention (MM-LKA) module, integrated into a Residual Star Distillation Attention Module (RSDAM). SDM applies the 'star operation' inside a multi-branch channel-splitting distillation structure, and MM-LKA computes attention from groups of 1D and 2D depth-wise convolutions of different kernel shapes and sizes. The network is trained from scratch with the ADAN optimizer on DF2K and evaluated on Set5, Set14, BSD100, Urban100, and Manga109 at scales ×2, ×3, and ×4. The paper reports best or second-best PSNR/SSIM among the compared lightweight methods on all five datasets with roughly 405–410K parameters, and results comparable to heavy networks such as EDSR and RDN with far fewer parameters.","tokens_in":18267,"tokens_out":10308,"duration_ms":108791,"significance":"If the reported gains are attributable to the proposed architecture, this is a meaningful contribution: it combines two currently active ideas (star operation and decomposed large-kernel attention) in a low-parameter regime and reports consistent gains over recent lightweight SR methods. The paper includes extensive benchmarking, and the reported numbers are direct measurements rather than predictions from fitted constants, so there is no circularity in the empirical claims. However, the central attribution claim is currently under-supported because the ablations lack a specified baseline, hyperparameters are chosen on test benchmarks, and prior methods are not retrained under the same protocol; these are fixable experimental issues rather than logical inconsistencies. No code or trained models are provided, which further limits independent verification.","major_comments":[{"comment":"The comparison with prior lightweight methods is uncontrolled. SDAN is trained with the ADAN optimizer (β1=0.98, β2=0.92, β3=0.99), EMA 0.999, constant learning rate 5×10^-3, and 1×10^6 iterations, whereas the baseline numbers are taken from their original papers with their own optimizers, schedules, and iteration counts. Given that the PSNR margins over the strongest baselines are only 0.02–0.20 dB (e.g., 32.54 vs. 32.39 dB on Set5 ×4 and 27.74 vs. 27.72 dB on BSD100 ×4), the reported superiority may be a training-recipe effect rather than an architectural one. Please retrain all baselines under the SDAN training protocol, or provide an ablation that fixes the training protocol and varies only the architecture.","section":"Section IV.A / Table V"},{"comment":"The ablations select architectural hyperparameters on the same test benchmarks used for final comparison: the SDM kernel size is chosen on Manga109 ×3 and the MM-LKA kernel sizes on Set5 ×4, and these benchmarks reappear in Table V. This induces selection bias in the final reported numbers. Please move hyperparameter selection to a held-out validation set, or use a nested evaluation that does not reuse the test split.","section":"Section IV.B / Tables II and III"},{"comment":"The baseline for the break-down ablation is not architecturally specified. The text says 'the model without the SDM and MM-LKA' but does not describe the remaining layers, channel widths, number of RSDAM-like blocks, or parameter count. Without this specification, the 0.12, 0.22, and 0.33 dB increments cannot be attributed to SDM and MM-LKA rather than to changes in capacity. Please give a precise baseline configuration and, ideally, parameter-matched variants.","section":"Section IV.B / Table I"},{"comment":"The parameter counts are inconsistent. The SDM kernel-size ablation reports 451K parameters for kz=7 on ×3 (Table II), and the text indicates that larger kernels perform better, yet the final ×3 SDAN in Table V has only 408K parameters. The paper does not explain how the selected kernel is reconciled with the final model's parameter count. Please clarify which kernel size is used in the final model and why the counts differ.","section":"Section IV.B / Tables II and V"},{"comment":"The presentation of the MM-LKA ablation conflicts with the text. The text states that the configuration with 7×7 square convolution and strip size 11 gives the best 32.54 dB, but as typeset, Table III does not unambiguously associate 32.54 dB with that configuration; the column markers appear to place the value elsewhere. The table should be re-typeset so that each column is a named configuration, and the text and table must agree.","section":"Section IV.B / Table III"},{"comment":"The claim that MM-LKA has lower computational and memory footprints than LKA is asserted but not demonstrated. No FLOPs or parameter comparison between LKA and MM-LKA is given, and Table VI reports only absolute FLOPs for whole networks. Please provide a formal derivation of the complexity reduction, or an ablation that measures the complexity of LKA versus MM-LKA under matched settings.","section":"Section III.C / Table VI"},{"comment":"The comparison table omits several recent LKA-based lightweight SR methods discussed in the Related Work, including LKASR, LKDN, MSID, and LCAN. Since the paper's contribution is specifically an improvement over LKA, these omissions weaken the state-of-the-art claim. Please include these methods or justify their exclusion.","section":"Table V / Related Work"}],"minor_comments":[{"comment":"The text refers to Table IV for lightweight SR results and Table V for heavy-model comparison, but the captions indicate the reverse; the cross-references should be corrected.","section":"Section IV.C / Tables IV and V"},{"comment":"LBRN is mentioned in the visual comparison but is not listed among the compared methods in Table V and is not described anywhere in the paper.","section":"Section IV.C / Visual results"},{"comment":"BSConv is used without definition or citation; please specify what this convolution is and why it is chosen.","section":"Section III.A"},{"comment":"Luo et al.'s LatticeNet is cited as reference [39], but reference [39] is the LKDN paper; LatticeNet appears to be reference [36].","section":"Section II.A / References"},{"comment":"The phrase 'FLOPs [43]' cites ECA-Net; a standard reference for FLOPs should be used, or the bracket removed.","section":"Section IV.D"},{"comment":"There are garbled symbols in the loss-function paragraph: 'Where' is capitalized mid-sentence and the parameters symbol appears as '9'; the formatting should be fixed.","section":"Equation (5) / Section III.A"},{"comment":"The parameter counts across the MM-LKA ablation and the final model should be made consistent; currently the best configuration selected in Table III is not clearly tied to the 410K parameter count reported for ×4 in Table V.","section":"Table III / Table V"}],"recommendation":"major_revision","confidential_remarks":"The paper is a standard lightweight-SR empirical contribution, and the core idea is plausible. The main risk is that the headline gains could disappear once the training protocol is controlled and the test-set hyperparameter selection is removed; this is testable and fixable. I do not see grounds for rejection, provided the authors can supply the controlled comparison and correct the internal inconsistencies. The omission of recent LKA-based SR baselines, including some from the authors' own group, should be addressed in revision."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: this is a competent, incremental lightweight SR paper. The specific package—StarNet-style star operation inside a channel-splitting distillation module, plus grouped strip-and-square large-kernel attention—is new as a combination, and on the standard five benchmarks it beats the listed lightweight methods at 405–410K parameters. The write-up is clear, the ablations are extensive, and the complexity table is reasonable.\n\nCredit where due: the internal ablation shows both components help (SDM +0.12 dB, MM-LKA +0.22 dB, together +0.33 dB on Set5 ×4). The idea of doing distillation in a high-dimensional non-linear space via element-wise products is a fresh twist on a known structure, and grouping the LKA depth-wise convolutions into 1D strips and 2D squares is a sensible way to cut the quadratic cost.\n\nThe soft spots are real. Most load-bearing: the comparison with prior methods is uncontrolled. The authors train with Adan, constant LR 5e-3, 1M iterations, EMA 0.999; the published numbers for RFDN, OSFFNet, FIWHN etc. were produced with their own optimizers and schedules. The margins over the strongest baselines are 0.02–0.20 dB PSNR—exactly the size range where training-protocol differences dominate. Without retraining at least the closest baselines under the same recipe, the claim that the architecture causes the gains is not established.\n\nSecond, hyperparameters are picked on test sets: SDM kernel size on Manga109 ×3, MM-LKA kernel sizes on Set5 ×4. That selection bias inflates the final numbers. Third, the ablation baseline is under-specified: 'model without SDM and MM-LKA' is not architecturally described, so the +0.33 dB attribution can't be independently checked.\n\nMinor but worth fixing: the text in Section IV.B says the best MM-LKA configuration is 7×7 square plus strip 11, but Table III appears to show 7×7 plus strip 7 giving 32.54 dB and strip 11 giving 32.47 dB. The text and table need to match. Also, no code or pretrained models are released; for an empirical paper, that's a gap.\n\nBottom line: the architecture is coherent, and I believe the reported numbers are real, but the central claim of superiority over the listed baselines is not proven as stated. This deserves a serious referee, but I'd make acceptance conditional on a same-protocol retraining of the strongest baselines and a code release. For an SR reading group, it's a useful case study in how evaluation protocols can masquerade as architectural gains.","headline":"Competent incremental lightweight SR paper whose headline claim is plausible but rests on an uncontrolled comparison; worth peer review with a request for retrained baselines and code.","tokens_in":18813,"tokens_out":5094,"would_cite":false,"duration_ms":56000,"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 410K-parameter lightweight super-resolution network claims the best PSNR and SSIM scores on five standard benchmarks.","keywords":["lightweight image super-resolution","star distillation","large kernel attention","multi-scale attention","information distillation","PSNR evaluation","single image super-resolution","efficient CNN"],"falsifier":"Train each competing lightweight method (e.g., IMDN, RFDN, OSFFNet, FIWHN) with the exact SDAN training protocol and compare on a held-out set of images not used for any hyperparameter selection; alternatively, replace SDM and MM-LKA in SDAN with a matched-parameter plain residual block and check whether the 0.33 dB gap on Set5 ×4 disappears.","tokens_in":17716,"feed_emoji":"✨","tokens_out":6114,"duration_ms":63850,"temperature":0.7,"pith_summary":"This paper proposes SDAN, a lightweight single-image super-resolution network with about 410K parameters, and claims it achieves the best PSNR/SSIM among compared lightweight methods on all five standard benchmarks (Set5, Set14, BSD100, Urban100, Manga109) for ×4 upscaling, and also for most ×2 and ×3 cases. The network's building block is a Residual Star Distillation Attention Module (RSDAM) combining two new components: a Star Distillation Module (SDM) that performs information distillation in high-dimensional non-linear feature spaces using the star operation, and a Multi-shape Multi-scale Large Kernel Attention (MM-LKA) module that captures long-range dependencies through grouped 1D and 2D depth-wise convolutions of different kernel sizes at low computational cost. The paper reports, for example, 32.54 dB on Set5 and 31.13 dB on Manga109 at ×4, with fewer parameters than the compared distilled-attention baselines. The central claim, if correct, is that lightweight SR can match or exceed the reconstruction quality of models with an order of magnitude more parameters.","feed_headline":"410K-parameter super-resolution net tops five benchmarks","feed_subtitle":"Star distillation plus multi-shape large-kernel attention reaches top 4× PSNR, if results reproduce.","key_machinery":"The key machinery is the Residual Star Distillation Attention Module (RSDAM), which stacks a Star Distillation Module (SDM) and a Multi-shape Multi-scale Large Kernel Attention (MM-LKA) module inside a residual connection. SDM uses the star operation (element-wise multiplication of two branches after depth-wise and 1×1 convolutions, as in StarNet) to map features into high-dimensional non-linear spaces, interleaved with channel splitting and 1×1 convolutions to distill features hierarchically. MM-LKA divides input features into groups: some receive 1D (strip) depth-wise convolutions and 1D dilated depth-wise convolutions, others receive 2D square depth-wise convolutions and 2D dilated depth-wise convolutions; the groups are concatenated and fused by a 1×1 convolution to produce attention weights that reweight the input. This design is intended to learn multi-shape multi-scale long-range dependencies while keeping computational and memory footprints far below standard LKA.","core_discovery":"The central discovery claimed is that a lightweight SR architecture can obtain state-of-the-art reconstruction quality by performing information distillation in high-dimensional non-linear feature spaces (using the star operation) instead of in low-dimensional spaces, and by capturing multi-shape multi-scale long-range dependencies with a group-wise 1D/2D depth-wise large-kernel attention module that avoids the quadratic complexity of standard large kernel attention. The paper reports that SDAN with 410K parameters attains the best PSNR/SSIM among compared lightweight methods on all five benchmarks for ×4 SR, e.g., 32.54 dB on Set5 and 31.13 dB on Manga109, and also matches or exceeds heavyweight networks such as EDSR, RDN, and RCAN while using at least 10 times fewer parameters. The authors attribute this to the discriminative representation learning enabled by SDM and the representative long-range dependencies captured by MM-LKA.","pith_inferences":["The paper's ablations tune kernel sizes on Manga109 and Set5, the same benchmarks used for final reporting; a fairer estimate of generalizable gain would come from a held-out validation set.","Because the training protocol (Adan, constant learning rate 5e-3, 1e6 iterations, EMA 0.999) differs from that of prior baselines, part of the reported margin may be training-recipe rather than architecture; re-running baselines under the same protocol would isolate the module contribution.","The same grouped multi-shape large-kernel attention idea could transfer to high-level vision tasks, where LKA has been used for classification and detection, and the parameter savings might be analogous.","A testable extension is to increase the number of RSDAM blocks at fixed parameter budget to see whether the star-distillation design continues to gain, or saturates sooner than standard residual blocks."],"forward_implications":["If the reported results reproduce, lightweight super-resolution can reach the quality of models with 10 to 100 times more parameters, lowering the barrier for on-device SR.","The star operation may serve as a cheap nonlinearity for channel distillation in other low-level vision tasks such as denoising and deblurring.","MM-LKA's group-wise combination of strip and square depth-wise kernels offers a template for reducing large-kernel attention cost in other CNN architectures.","The ablations suggest MM-LKA contributes more than SDM to the gain (0.22 dB vs 0.12 dB on Set5 ×4), pointing future work toward attention design rather than distillation alone."],"supporting_citations":[{"why":"Introduces the feature distillation strategy for lightweight SR that SDM extends.","marker":"[11]"},{"why":"IMDN is a primary distillation-based baseline whose low-dimensional mapping SDM aims to improve.","marker":"[12]"},{"why":"RFDN is a key lightweight baseline and provides the overall network skeleton SDAN adapts.","marker":"[13]"},{"why":"Defines the large kernel attention (LKA) decomposition that MM-LKA generalizes.","marker":"[40]"},{"why":"Provides the separable large-kernel attention idea used to reduce LKA's quadratic cost.","marker":"[25]"},{"why":"Shows the star operation maps inputs into high-dimensional non-linear feature spaces, the property SDM exploits.","marker":"[59]"},{"why":"OSFFNet is a recent lightweight baseline and the source of the training protocol (DF2K, L1 loss) SDAN follows.","marker":"[18]"}],"fun_headline_variants":["410K-param star-distillation SR tops five benchmarks","Star distillation + multi-shape LKA: tiny SR net wins on quality","Lightweight SR net uses star distillation to beat larger models","SDAN: 410K-parameter SR with star distillation leads benchmarks","Star Distillation Attention Network: high quality at 410K params"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the reported PSNR/SSIM improvements are caused by the proposed SDM and MM-LKA modules rather than by the particular training recipe or by kernel-size choices tuned on the test benchmarks.","fun_headline_variants_meta":{"raw":{"variants":["410K-param star-distillation SR tops five benchmarks","Star distillation + multi-shape LKA: tiny SR net wins on quality","Lightweight SR net uses star distillation to beat larger models","SDAN: 410K-parameter SR with star distillation leads benchmarks","Star Distillation Attention Network: high quality at 410K params"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000239,"raw_usage":{"total_tokens":1542,"prompt_tokens":999,"completion_tokens":543,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":615,"completion_tokens_details":{"reasoning_tokens":453}},"tokens_in":615,"tokens_out":543,"duration_ms":6829,"temperature":1.0,"reasoning_tokens":453,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T00:48:14.035498+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train each competing lightweight method (e.g., IMDN, RFDN, OSFFNet, FIWHN) with the exact SDAN training protocol and compare on a held-out set of images not used for any hyperparameter selection; alternatively, replace SDM and MM-LKA in SDAN with a matched-parameter plain residual block and check whether the 0.33 dB gap on Set5 ×4 disappears.","supporting_citations":[{"cited_title":"Fast and accurate single image super-resolution via information distillation network,","cited_arxiv_id":null,"evidence_quote":"Introduces the feature distillation strategy for lightweight SR that SDM extends."},{"cited_title":"Lightweight im- age super-resolution with information multi-distillation network,","cited_arxiv_id":null,"evidence_quote":"IMDN is a primary distillation-based baseline whose low-dimensional mapping SDM aims to improve."},{"cited_title":"Residual feature distilla- tion network for lightweight image super-resolution,","cited_arxiv_id":null,"evidence_quote":"RFDN is a key lightweight baseline and provides the overall network skeleton SDAN adapts."},{"cited_title":"Visual attention network,","cited_arxiv_id":null,"evidence_quote":"Defines the large kernel attention (LKA) decomposition that MM-LKA generalizes."},{"cited_title":"Large separable kernel attention: Rethinking the large kernel attention design in cnn,","cited_arxiv_id":null,"evidence_quote":"Provides the separable large-kernel attention idea used to reduce LKA's quadratic cost."},{"cited_title":"Rewrite the stars,","cited_arxiv_id":null,"evidence_quote":"Shows the star operation maps inputs into high-dimensional non-linear feature spaces, the property SDM exploits."},{"cited_title":"Osffnet: Omni-stage feature fusion network for lightweight image super-resolution,","cited_arxiv_id":null,"evidence_quote":"OSFFNet is a recent lightweight baseline and the source of the training protocol (DF2K, L1 loss) SDAN follows."}],"review_version":1}