{"id":"d738625b-86a5-4a35-8328-6fe592b292d5","arxiv_id":"2412.19535","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"StyleRWKV applies recurrent RWKV-style attention with deformable shifting and skip scanning to achieve fast, high-quality arbitrary style transfer.","lead":"This paper introduces StyleRWKV, an image style transfer model that uses an RWKV-like sequence architecture to combine a content photo with an artist's style in a single forward pass. It claims to match or beat prior style transfer methods while running faster and using less memory, because its attention mechanism scales linearly rather than quadratically with image size.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Evaluation protocol is the weakest load-bearing link: q and p are selected on what appears to be the same 5,000-pair test set used for final numbers, and baseline timings show signs of inconsistent measurement; until this is checked, 'outperforms SOTA' is not independently verifiable.","rationale":"The reader's verdict is CONDITIONAL, and I agree that the paper's empirical claim is not fully established. My stress-test converges on the same weakest assumption: evaluation hygiene. I would keep the conditional stance rather than escalate to REJECT or UNVERDICTED because the architectural proposal is coherent and the core efficiency argument is plausible on its face: the bidirectional WKV formula in Eq. 5 is separable into prefix and suffix running sums, which gives O(T) complexity for each of the q=2 sweeps; Deform-Shifting is a local deformable convolution; and S-Scanning splits the token grid into four strided groups, so total work remains linear. The paper also reports concrete numbers, user-study responses, and ablations that are internally consistent in direction. What is missing is a demonstration that the comparisons are unbiased: no code, no error bars, no validation split, and suspicious baseline timing behavior. The proposed recomputation with a true validation set and uniform timing protocol is the single check that would settle whether the headline superiority in quality and efficiency is real. I mark agreement as partial rather than full because the reader emphasizes test-set tuning; I additionally flag the AesPA-Net flat timings and the lack of a measurement protocol as independent evidence that the efficiency comparison needs the same verification.","tokens_in":12145,"tokens_out":7321,"duration_ms":69550,"concrete_test":"Release the code and rerun the full protocol with an explicit split: partition the 70,097 COCO-WikiArt pairs into disjoint train, validation, and test subsets (for example 60k/5,097/5,000), select the recurrence q and skip step p on the validation subset only, and then recompute Table 1's ArtFID/FID/LPIPS and inference timings on the untouched test set with identical GPU, batch size, and input sizes for every baseline. If the new ArtFID for q=2,p=2 stays below the best re-measured baseline and the StyleRWKV timings remain fastest among linear-complexity methods, the central claim survives; otherwise the paper should report a smaller, properly validated advantage.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim in the abstract ('outperforms state-of-the-art methods in terms of stylization quality, model complexity, and inference efficiency') rests on Table 1. That table uses the configuration q=2 and p=2, but Section 4.3 selects q=2 and p=2 from Tables 2 and 4 using ArtFID/FID/LPIPS, and no separate validation set is described in Section 4.1. Since the only evaluation set named in the paper is the 5,000-pair COCO-WikiArt test set, the ablations may have been used to select hyperparameters on the test set; this would make the reported ArtFID 26.370/FID 16.362/LPIPS 0.451 optimistically biased. Notably, Table 2 shows q=3 improves all three quality metrics over q=2, so the choice of q=2 is a selection among test-set numbers, not a pre-registered configuration. The efficiency comparison also needs scrutiny: AesPA-Net times are nearly flat from 0.668 s at 128 to 0.695 s at 1024, while StyTR2 grows to 43,319 s at 1024; no hardware, batch size, or measurement protocol is provided. These issues do not imply the architecture is wrong; the Re-WKV bidirectional exponential-decay form can be implemented with prefix/suffix scans in O(T), and Deform-Shifting/S-Scanning are plausibly linear. The concern is that the evidence presented does not yet establish the comparative claim without an independent, protocol-fixed evaluation.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes StyleRWKV, a hierarchical encoder-decoder architecture for arbitrary neural style transfer that replaces quadratic-complexity attention with RWKV-like linear-complexity components. Three components are introduced: a Recurrent WKV (Re-WKV) attention with bidirectional exponential decay applied recurrently, a Deformable Shifting (Deform-Shifting) layer using deformable convolution for local token interaction, and a Skip Scanning (S-Scanning) mechanism that groups tokens with a step size to build global dependencies. The model is trained with VGG-based content, style, and identity losses, and evaluated on 5,000 COCO-WikiArt pairs. The paper reports the best ArtFID, FID, and LPIPS among the compared methods, lower parameter counts than transformer and diffusion baselines, and faster inference than Mamba-based and transformer-based competitors at long sequence lengths, with ablations over recurrence count, shifting mechanism, and scanning method.","tokens_in":12532,"tokens_out":4434,"duration_ms":39920,"significance":"If the empirical claims hold, StyleRWKV is a meaningful step toward efficient arbitrary style transfer: it combines the linear-scaling property of RWKV with a hierarchical, multi-scale design, and the ablation study suggests that each proposed module contributes positively. The architecture itself is plausible and the complexity analysis is credible. The paper also responds to a real gap, since most existing style transfer models are either quadratic in complexity or require iterative diffusion sampling. However, the comparative claims in the abstract and Table 1 rest on an evaluation protocol that is currently not fully verifiable: hyperparameters appear to be selected on the same test set used for final numbers, baseline timings lack a measurement protocol, and the user-preference row is ambiguous. The contribution is of interest to the style-transfer and efficient-sequence-modeling communities, but the evidence needs to be tightened before the central claim can be accepted.","major_comments":[{"comment":"The hyperparameters q=2 and p=2 are selected using the same metrics (ArtFID, FID, LPIPS) that are later reported as the final results in Table 1, but Section 4.1 describes only a single 5,000-pair COCO-WikiArt test set with no separate validation split. This creates a risk of test-set tuning that would bias the reported ArtFID 26.370, FID 16.362, and LPIPS 0.451 optimistically. Notably, Table 2 shows q=3 improves all three quality metrics over q=2, so the choice of q=2 is explicitly a trade-off between quality and efficiency, not a quality-maximizing selection. The paper should either introduce a held-out validation set for hyperparameter selection or clearly state that the ablation tables are exploratory comparisons and that the chosen configuration is a speed-quality compromise.","section":"Section 4.3, Tables 2 and 4"},{"comment":"No measurement protocol is reported for the inference times: no GPU model, batch size, framework version, warm-up procedure, or number of repeated runs. Some numbers are implausible or at least unexplained: AesPA-Net goes from 0.668 s at 128 to 0.695 s at 1024 (an almost flat curve for a 4x increase in sequence length), while StyTR2 grows to 43,319 s at 1024 (over 12 hours) despite having only 35.39M parameters. Without a standardized protocol, the efficiency claim in the abstract and the comparison in Table 1 are not independently verifiable. The authors should specify the hardware and timing methodology and provide error bars or at least explain the anomalously flat AesPA-Net scaling.","section":"Table 1, inference time rows"},{"comment":"The preference row is ambiguous: the 'Ours' column contains '0.7*' while every other method has a pair such as '0.66/0.34'. The text states that most participants favored the stylized results, but the colon-separated pairs are not defined. If the pair is 'ours vs. competitor', the table should say so explicitly; if the values are the competitor's win rate, the claim is contradicted. In addition, no statistical significance test or inter-subject agreement is reported for the 2,250 responses, so the reader cannot tell whether the preference differences are meaningful.","section":"Table 1, user preference row"},{"comment":"The definition of Re-WKV is internally inconsistent. The recursion states wkv^{(j)} = Bi-WKV^{(j)}(K, wkv^{(j-1)}), which uses the previous output as the value input, but the closed-form Bi-WKV(K, V) immediately below uses V as the value input. The paper should clarify whether the first iteration uses V = V_{ms}^{n} and subsequent iterations use V = wkv^{(j-1)}, or whether Eq. (5) contains a typo. This is load-bearing for reproducibility because the recurrent mechanism is the paper's central architectural novelty. Additionally, 'According to Eq. 4' in the same section appears to reference the wrong equation; Eq. (5) or the Bi-WKV definition is meant.","section":"Section 3.3, Eq. (5)"},{"comment":"The S-Scanning operation is not defined precisely enough to be reproduced. The slicing notation K[..., a::p, b::p] with (a,b) given by Eq. (8) is not a standard slicing expression, and the arrow-based notation for the split and merge steps is garbled in the text. The authors should state exactly how the four groups O_i are formed, how the scanning order is fed into Re-WKV, and how the groups are merged back into the full feature map. Without this, the S-Scanning mechanism cannot be implemented from the paper.","section":"Section 3.4, Eq. (7)"}],"minor_comments":[{"comment":"The note '128×128, 256×256, etc. refer to sequence lengths rather than resolution' is confusing because the headers read 128×128, 256×256. If the sequence length is 128 (rather than 128×128 patches), the headers should be '128', '256', etc., or the text should specify the patch size used to derive the sequence length.","section":"Table 1 note"},{"comment":"The sentence 'According to Eq. 4, Bi-WKV is partially determined by the relative position bias' references Eq. (4), which is the channel-mix formula; the relevant formula is Eq. (5) or the definition of Bi-WKV.","section":"Section 3.3"},{"comment":"There are typographical issues throughout the paper, e.g., 'DMs-based methods ... are hindered by their time- consuming generation process' in the introduction, and reference [27] contains an apparent page-number artifact ('arXiv:2303.15403, 3, 2023'). A thorough proofread is needed.","section":"General"},{"comment":"The 'Time/s' columns in the ablation tables do not state the input sequence length at which the timings were measured, making it hard to compare them with the timing rows of Table 1.","section":"Tables 2-4"},{"comment":"The paper does not report random seeds, the number of runs per evaluation, or any error bars for the main metrics. At minimum, the authors should state that metrics are single-run and report seeds for reproducibility.","section":"Section 4.1"}],"recommendation":"major_revision","confidential_remarks":"The main blocker is the evaluation protocol: hyperparameter selection on what appears to be the test set, unexplained baseline timings, and an ambiguous user-preference table. These are fixable with a validation split, a timing protocol, and clearer table notation, so I do not recommend rejection. I would also ask the editor to verify the novelty claim against concurrent RWKV-based style transfer work, since the related-work section surveys Mamba-based and transformer-based methods but no RWKV style transfer prior art; the claim of being 'the first' should be carefully worded."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The first thing to know: this is the first RWKV-style model for style transfer, and the three mechanisms it adds—bidirectional recurrent WKV, deformable convolution for token shifting, and atrous-style skip scanning—are plausible and each ablates positively. The architecture is described clearly enough to reimplement, and the linear-complexity argument for Re-WKV is sound: it's a bidirectional exponential-decay attention that can be computed with prefix/suffix scans.\n\nThe paper does not, however, support its headline claim that it 'outperforms state-of-the-art methods in stylization quality, model complexity, and inference efficiency.' The evaluation protocol has two problems. First, hyperparameters q and p are chosen in Section 4.3 using ArtFID/FID/LPIPS on what appears to be the same 5,000-pair COCO-WikiArt set named as the test set in Section 4.1; no separate validation set is described. Table 2 shows q=3 gives better quality than q=2, so the choice is an efficiency trade-off, but it's being made on test numbers. Table 4 does the same for p. That biases the reported metrics. Second, the timing comparison is hard to trust: AesPA-Net times are almost flat from 128 to 1024 resolution (0.668 to 0.695 s), which is implausible for a feed-forward CNN, and StyTR2 at 43,319 s is a red flag. No hardware or measurement protocol is given.\n\nThere are also smaller issues: no code, no error bars or multiple seeds, and the loss functions use VGG features while LPIPS—an evaluation metric—also uses VGG, which creates mild circularity. That last one is common in style transfer and not disqualifying.\n\nWhat the paper gets right: the combination is genuinely new, the ablations are methodical, and the qualitative results look competitive. If the efficiency and quality numbers hold up under a clean protocol, this is a useful linear-complexity alternative to Mamba/Transformer backbones for style transfer. But as submitted, the comparative claims are not independently verifiable.\n\nWho this is for: researchers working on efficient image-to-image translation or applying RWKV/SSM architectures to vision. It deserves a serious referee—the novelty and clarity justify referee time—but the referee should require code, a validation set, and a fixed timing protocol before the efficiency claim can be accepted.","headline":"A plausible first RWKV-based style transfer model with well-ablated components, but the evaluation protocol needs fixing before the 'outperforms SOTA' claim is credible.","tokens_in":13017,"tokens_out":4062,"would_cite":false,"duration_ms":32623,"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":"The paper claims that an RWKV-style backbone, StyleRWKV, delivers state-of-the-art arbitrary style transfer while keeping inference linear in sequence length.","keywords":["style transfer","RWKV","linear complexity","attention mechanism","deformable convolution","efficient inference","encoder-decoder architecture"],"falsifier":"A re-run that selects q and p on a separate validation split and then evaluates on the untouched 5,000-pair test set—plus an independent timing of StyTR2 at 512×512 on the same hardware—would determine whether the reported quality and efficiency margins hold.","tokens_in":11975,"feed_emoji":"🎨","tokens_out":7770,"duration_ms":64880,"temperature":0.7,"pith_summary":"StyleRWKV aims to show that an RWKV-style sequence model—previously used for long-context language modeling—can match or beat specialized Transformer-, diffusion-, and state-space-based style transfer models on both quality and speed. The paper introduces three components: Recurrent WKV (Re-WKV) attention for global context with linear complexity, Deformable Shifting (Deform-Shifting) for adaptive local token fusion, and Skip Scanning (S-Scanning) for long-range dependencies. On a 5,000-pair COCO-WikiArt test set it reports the best ArtFID, FID, and LPIPS among all compared methods, while keeping inference time at 1.53 seconds at 512×512 resolution and scaling linearly with resolution. If correct, it establishes RWKV-like architectures as a practical backbone for efficient high-quality arbitrary style transfer, especially at high resolutions.","feed_headline":"RWKV architecture wins style transfer with linear-time speed","feed_subtitle":"New encoder-decoder with recurrent WKV attention improves ArtFID, FID, and LPIPS while cutting inference time at high resolution.","key_machinery":"The load-bearing mechanism is the Recurrent WKV (Re-WKV) attention: it takes the bidirectional WKV attention from Vision-RWKV and applies it recurrently q times along the S-Scanning paths, with a channel-wise spatial decay vector w and a current-token bonus u, to compute a global attention output in linear time. This is paired with two complementary modules: Deform-Shifting, which learns convolution offsets to move tokens from regions of interest and enhance local dependencies, and S-Scanning, which slices the feature map with a step size p to create sparse scanning directions that connect distant patches. Together they give the ST-RWKV block both global and local receptive fields while avoiding quadratic cost.","core_discovery":"The paper claims that StyleRWKV, a 4-level U-shaped encoder-decoder built from ST-RWKV blocks, outperforms state-of-the-art arbitrary style transfer methods in stylization quality, model complexity, and inference efficiency. The core of the model is the Re-WKV attention, which runs Bi-WKV attention recurrently q=2 times along S-Scanning directions, giving a global receptive field at linear cost; local details are handled by Deform-Shifting, which uses deformable convolution to shift tokens from regions of interest; and S-Scanning samples tokens with a stride p=2 to build long-range context. In experiments the model achieves ArtFID 26.370, FID 16.362, LPIPS 0.451, and inference times of 0.266s at 128 and 3.670s at 1024 sequence lengths, better than all compared Transformer-, diffusion-, and Mamba-based baselines.","pith_inferences":["Because the hyperparameters q=2 and p=2 were selected using test-set metrics, re-running the ablation with a separate validation split would verify whether the claimed margins hold on an untouched test set.","The same recurrent bidirectional WKV + skip-scanning design could transfer to other high-resolution dense prediction tasks, such as super-resolution or semantic segmentation, where global context and local detail both matter.","Deform-Shifting's learned offsets could be visualized to reveal whether the model implicitly discovers object boundaries, which would explain its strong preservation of local details like the duck's beak in the qualitative results."],"forward_implications":["Transformer-based style transfer methods incur quadratic cost as resolution grows, so a linear-complexity alternative like StyleRWKV can scale to high-resolution images without that penalty.","The reported inference times—0.266s at 128, 1.530s at 512, and 3.670s at 1024 sequence lengths—make the method fast enough for practical use on standard hardware.","With 28.8M parameters, the model is far smaller than diffusion-based style transfer models, which reduces memory usage and training overhead.","The user study reports that 70% of responses favored StyleRWKV over ten competing methods, suggesting the quality gains are perceptually visible, not just metric-based."],"supporting_citations":[{"why":"Supplies the original WKV attention and fixed-direction token shifting that StyleRWKV adapts into Re-WKV and Deform-Shifting.","marker":"[37]"},{"why":"Provides the bidirectional WKV (Bi-WKV) attention that Re-WKV runs recurrently, as well as the Quad-Shifting baseline used in ablations.","marker":"[15]"},{"why":"Supplies the deformable convolution whose learned offsets form the basis of the Deform-Shifting layer.","marker":"[8]"},{"why":"The Mamba-based baseline that StyleRWKV compares against for long-sequence efficiency and quality.","marker":"[2]"},{"why":"The Transformer-based style transfer baseline (StyTR2) whose quadratic complexity motivates the linear-complexity design.","marker":"[13]"},{"why":"Introduces AdaIN, which the StyleRWKV hierarchy uses to align content and style feature statistics and is also a compared baseline.","marker":"[26]"},{"why":"Defines the ArtFID metric that serves as the paper's primary evaluation measure.","marker":"[48]"},{"why":"Provides the MS-COCO content images used to train and test the model.","marker":"[4]"},{"why":"Provides the WikiArt style images used to train and test the model.","marker":"[39]"}],"fun_headline_variants":["RWKV linear-time style transfer beats SOTA","StyleRWKV: fast, high-quality art with linear cost","Linear-time RWKV attention for style transfer","RWKV outpaces SOTA style transfer models","Recurrent WKV accelerates style transfer linearly"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The central claim depends on the reported comparisons being fair—that the 5,000-pair test set was not used to select hyperparameters (q=2, p=2) and that the baseline inference times were measured under comparable conditions.","fun_headline_variants_meta":{"raw":{"variants":["RWKV linear-time style transfer beats SOTA","StyleRWKV: fast, high-quality art with linear cost","Linear-time RWKV attention for style transfer","RWKV outpaces SOTA style transfer models","Recurrent WKV accelerates style transfer linearly"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000286,"raw_usage":{"total_tokens":1686,"prompt_tokens":956,"completion_tokens":730,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":572,"completion_tokens_details":{"reasoning_tokens":658}},"tokens_in":572,"tokens_out":730,"duration_ms":7718,"temperature":1.0,"reasoning_tokens":658,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T00:13:47.268217+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"A re-run that selects q and p on a separate validation split and then evaluates on the untouched 5,000-pair test set—plus an independent timing of StyTR2 at 512×512 on the same hardware—would determine whether the reported quality and efficiency margins hold.","supporting_citations":[{"cited_title":"Deformable convolutional networks","cited_arxiv_id":null,"evidence_quote":"Supplies the deformable convolution whose learned offsets form the basis of the Deform-Shifting layer."},{"cited_title":"Stytr2: Image style transfer with transformers","cited_arxiv_id":null,"evidence_quote":"The Transformer-based style transfer baseline (StyTR2) whose quadratic complexity motivates the linear-complexity design."},{"cited_title":"Arbitrary style transfer in real-time with adaptive instance normalization","cited_arxiv_id":null,"evidence_quote":"Introduces AdaIN, which the StyleRWKV hierarchy uses to align content and style feature statistics and is also a compared baseline."},{"cited_title":"Artfid: Quantitative evaluation of neural style transfer","cited_arxiv_id":null,"evidence_quote":"Defines the ArtFID metric that serves as the paper's primary evaluation measure."},{"cited_title":"Coco- stuff: Thing and stuff classes in context","cited_arxiv_id":null,"evidence_quote":"Provides the MS-COCO content images used to train and test the model."}],"review_version":1}