{"id":"2e826a68-120e-4e95-b129-bb73e0ce1c37","arxiv_id":"2506.05488","paper_version":1,"verdict":"REJECT","confidence":"HIGH","novelty_score":4.0,"correctness_risk":"high","formal_verification":"none","parameter_count":5,"one_line_summary":"VR-INR claims a single video-restoration network handles arbitrary super-resolution scales and zero-shot denoising, but the reported PSNR values and per-video fine-tuning undermine the claim.","lead":"A video restoration network called VR-INR claims to learn on one upscale factor and then work at any scale and remove noise it never saw during training. The reported accuracy is far higher than standard benchmarks allow, and the paper's own conclusion says the model is fine-tuned per video, which contradicts the zero-shot claim.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Admitted per-video fine-tuning in Sec. 5 invalidates the zero-shot generalization claims; reported PSNRs likely stem from test-time adaptation rather than the once-trained ×4 model.","rationale":"The reader's weakest_assumption is precisely the load-bearing weakness: the paper's own conclusion admits per-video fine-tuning, which converts the claimed zero-shot and arbitrary-scale results into per-sequence fits. I agree with the REJECT verdict because the inconsistency is structural, not cosmetic. The abstract and Sec. 3.5 promise a single network trained only on clean ×4 pairs with no additional training at inference, while Sec. 5 explicitly contradicts this. The reported PSNR values are implausibly high for genuine generalization, and the absence of code, error bars, and a description of the fine-tuning protocol means the contradiction cannot be resolved in the authors' favor. The concern is not that the numbers are outside current consensus; it is that the paper's own text makes the stated evaluation protocol incompatible with the central claim. A concrete and simple reproduction test -- evaluating the once-trained checkpoint with no per-video adaptation -- would settle the matter, but until that test is run and reported, the central claim is unverified and internally inconsistent.","tokens_in":12594,"tokens_out":2315,"duration_ms":26970,"concrete_test":"Obtain the authors' code (the project page is listed but no code is released) and run the decisive check: train VR-INR once on the clean ×4 training set as described in Sec. 3.4, then evaluate on Vid4, REDS4, GOPRO, and DAVIS without any per-video fine-tuning. If the ×4 PSNR drops from the reported 44.21 dB to near or below the baseline range (≤ 31 dB) or if the training logs reveal per-video gradient updates before every reported entry, the central generalization and zero-shot claims collapse. A complementary check within the same protocol: for Table 2, verify that no noisy test frame or its clean HR counterpart participates in any gradient update; otherwise the term 'zero-shot' is false.","verdict_should_be":"REJECT","load_bearing_attack":"The central claim requires a single network trained once on clean ×4 LR–HR pairs that generalizes at test time to unseen scales and zero-shot denoising. However, the conclusion (Sec. 5) states: 'By fine-tuning per video using only clean LR–HR pairs, VR-INR adapts to each sequence's unique content and noise characteristics...' If per-video fine-tuning is applied to each test video, the evaluation is no longer measuring generalization of a once-trained model; every reported number becomes a per-sequence fit. The numbers are consistent with this reading: 44.21 dB PSNR on Vid4 ×4 and 31.50 dB on zero-shot denoising far exceed published generalization results (SA-VSR gives 30.95 dB ×2 and 24.50 dB ×4; typical Vid4 ×4 SOTA is near 28 dB). The experimental section (Sec. 4) does not describe the fine-tuning protocol, its computational cost, or how it is separated from the reported test numbers, and the abstract and Sec. 3.5 claim 'without additional training' and 'zero-shot', which is internally inconsistent with Sec. 5's fine-tuning statement. The paper therefore fails to establish the load-bearing premise that arbitrary-scale SR and zero-shot denoising are measured without per-video test-time adaptation.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes VR-INR, an implicit neural representation framework for video restoration, claiming that a single network trained only on clean LR-HR pairs at a fixed ×4 upscaling factor generalizes to arbitrary unseen super-resolution scales and performs zero-shot denoising. The method combines hierarchical spatial-temporal-texture encoding, multi-resolution hash embedding, a top-down attention mechanism, and a proposed pixel-error amplified loss. Experiments on Vid4, REDS4, GOPRO, and DAVIS report large improvements over prior video super-resolution and denoising baselines. However, the conclusion states that VR-INR is fine-tuned per video using clean LR-HR pairs, which directly contradicts the abstract and inference sections' claims of no additional training and zero-shot generalization.","tokens_in":12995,"tokens_out":3388,"duration_ms":31928,"significance":"If the central claim were valid—a single model trained only on clean ×4 pairs that generalizes to arbitrary scales and unseen noise—the contribution would be significant for video restoration. The proposed architecture and loss may contain useful ideas. However, the admitted per-video fine-tuning in Sec. 5 undermines the central generalization claim: the reported evaluations may measure per-sequence fitting rather than the behavior of a once-trained model. The reported PSNR values (e.g., 44.21 dB on Vid4 ×4, 46.93 dB in the ablation) are implausibly high relative to published SOTA, and the experimental section does not describe a protocol that would make them reproducible. The paper does not ship code, checkpoints, or detailed experimental settings, and the internal inconsistency between the abstract and conclusion is load-bearing rather than cosmetic.","major_comments":[{"comment":"The conclusion states: 'By fine-tuning per video using only clean LR–HR pairs, VR-INR adapts to each sequence's unique content and noise characteristics.' This directly contradicts the abstract and Sec. 3.5, which claim operation 'without additional training' and 'zero-shot denoising.' If the model is fine-tuned on each test video, the reported results are fitted values for each sequence, not generalization of a model trained once on clean ×4 pairs. The paper must specify precisely which parameters are trained once and which are optimized per video, and must re-run the evaluation without any test-time adaptation to support the stated central claim.","section":"Section 5 (Conclusion)"},{"comment":"The reported PSNR values are far outside the range of published video super-resolution results. For example, the paper reports 44.21 dB on Vid4 ×4 and 42.00 dB on DAVIS ×4, whereas typical SOTA on Vid4 ×4 is around 28 dB and SA-VSR (a strong arbitrary-scale method) is reported at 30.95 dB ×2 and 24.50 dB ×4 in the paper's own table. No error bars, number of test clips, or train/test separation are provided. The experimental section does not describe whether the reported numbers come from the once-trained model or from per-video fine-tuning. Without this information, the comparison with baselines is uninterpretable.","section":"Section 4.2, Table 1"},{"comment":"The ablation study reports 46.93 dB PSNR on DAVIS ×4, which is 4.93 dB higher than the 42.00 dB reported for the same dataset and scale in Table 1. If the higher number arises from per-video fine-tuning in the ablation while Table 1 uses a single fixed model, the paper must clarify this; if not, the inconsistency suggests that the reported metrics are not reproducible from the described experimental settings. The paper should report the evaluation protocol for each table explicitly.","section":"Section 4.3, Table 4"},{"comment":"The baselines (VideoINR, VRT, IART, SA-VSR) are evaluated without any fine-tuning at unseen scales or noise levels, while VR-INR is fine-tuned per video according to Sec. 5. This creates an unfair comparison: the numbers attributed to VR-INR may reflect test-time adaptation that the baselines are not allowed. The paper must either evaluate all methods under the same protocol (i.e., no per-video fine-tuning) or separate results into 'zero-shot' and 'fine-tuned' conditions.","section":"Section 4.2, 'Compared with SOTAs'"}],"minor_comments":[{"comment":"Equation (17) contains malformed notation with nested and garbled parentheses that make the expression unreadable; it should be rewritten with clear definitions of all subterms.","section":"Section 3.5, Eq. (17)"},{"comment":"The dataset name 'DA VIS' should be 'DAVIS' (as in the DAVIS challenge reference), and the method name 'NERV' in Table 3 and related text should be 'NeRV'.","section":"Throughout"},{"comment":"The sentence 'For DA VIS and GOPRO, we first resized the video frames to 256 × 256 pixels, which served as the HR ground truth' is unclear about the generation of LR inputs; specify the exact downsampling procedure used to produce LR frames from the HR ground truth.","section":"Section 4.2, Dataset paragraph"},{"comment":"The text claims that the PEA-loss hyperparameters τ, ε, and δ have little impact within small ranges, but no quantitative ablation table is provided for these values. The selected values (τ=0.01, ε=0.005, δ=0.001, α=5) should be supported by a table or at least a figure.","section":"Section 4.3, PEA-loss hyperparameters"},{"comment":"The conclusion claims the method is 'computationally efficient' and 'simplifies the restoration pipeline,' but no runtime, parameter count, or FLOPs comparisons are given anywhere in the paper.","section":"Section 5, Conclusion"}],"recommendation":"reject","confidential_remarks":"The self-admitted per-video fine-tuning in Sec. 5 is a decisive contradiction of the paper's central claim. Combined with the implausibly high PSNR values and the absence of any described evaluation protocol, the reported results cannot be considered evidence for zero-shot generalization. The paper would need a complete rework of the experimental design and a major clarification of the training/evaluation split before it could be considered for publication."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Hi — the short version: the abstract claims a single network trained once on clean ×4 pairs that generalizes to arbitrary scales and zero-shot denoising. The conclusion (Section 5) says the opposite: \"By fine-tuning per video using only clean LR–HR pairs, VR-INR adapts to each sequence's unique content and noise characteristics.\" That is the whole game. Every number in Tables 1–3 is a per-sequence fit, not a prediction. And the numbers are implausible — 44.21 dB on Vid4 ×4, 46.93 dB in the ablation, while SA-VSR and similar methods sit near 28–31 dB. There are no error bars, no train/test separation, no code, and the feature-code length is F=3 in the method but F=6 in Table 4. I don't see how the central claim survives.\n\nTo give credit where it is due: the architecture is a genuine combination of LIIF-style local implicit functions, multi-resolution hash encoding from Instant-NGP, and temporal coordinates as in VideoINR, plus a top-down attention mechanism and a pixel-error amplified loss. That specific mashup appears to be new, and the attention design is not obviously broken. The related-work coverage is adequate, and the paper correctly positions itself relative to SA-VSR and VideoINR.\n\nThe soft spots are structural. The abstract and Section 3.5 say \"zero-shot\" and \"without additional training\"; Section 5 admits per-video fine-tuning. If fine-tuning is part of the test, then \"generalization to unseen scales\" is really per-video optimization, and the high PSNR is expected from fitting. The paper never states when fine-tuning happens, how many steps, or how the reported test numbers are separated from it. The zero-shot denoising claim is even weaker: there is no mechanism explaining why a network trained on clean pairs would suppress noise, other than the unstated per-video adaptation.\n\nThis paper is for someone tracking INR-based video SR architectures who might mine the design, but no one should trust the numbers. It is not ready for peer review. It needs a corrected protocol, a clear statement of whether the model is fine-tuned at test time (and if so, a proper per-video fitting comparison), code, seeds, and error bars. A referee would spend their time rediscovering what the conclusion admits.\n\nMy recommendation: desk reject in its current form. If the authors resubmit with an honest protocol and realistic numbers, the architecture might merit a closer look.","headline":"Admitted per-video fine-tuning in the conclusion guts the zero-shot and arbitrary-scale claims; the 44 dB numbers are fits, not predictions.","tokens_in":13413,"tokens_out":4194,"would_cite":false,"duration_ms":40474,"reading_group":"no","serious_thinker":"no","would_accept_peer_review":false},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"A network trained only on clean ×4 video pairs claims to upscale at any scale and denoise zero-shot.","keywords":["implicit neural representation","video super-resolution","zero-shot denoising","arbitrary-scale video restoration","multi-resolution hash encoding","coordinate-based networks","pixel-error amplified loss"],"falsifier":"Run the published evaluation protocol exactly as described in Section 4.2 but disable any per-video fine-tuning before testing; if the ×2 and ×8 PSNR values on Vid4 exceed the baselines by the reported margins, the zero-shot claim stands, and if they fall to baseline levels, the claim is falsified.","tokens_in":12347,"feed_emoji":"🎥","tokens_out":7716,"duration_ms":82799,"temperature":0.7,"pith_summary":"VR-INR claims that one implicit neural network, trained only on clean low- and high-resolution video pairs at a single upscaling factor (×4), can at test time produce high-resolution frames at arbitrary unseen scales (from ×2 to ×32) and suppress Gaussian and Poisson noise it has never seen. The method replaces explicit optical flow with a hierarchical spatial–temporal–texture encoding and multi-resolution hash interpolation, decoding each output pixel from a continuous coordinate representation. If the claim holds, video restoration no longer needs a separate model per scale or per degradation, and clean data alone may suffice to handle noisy inputs. The paper reports consistently higher PSNR/SSIM than fixed-scale baselines across Vid4, REDS4, GOPRO, and DAVIS at ×2, ×4, and ×8.","feed_headline":"One ×4-trained model upscales video at any scale, denoises zero-shot","feed_subtitle":"A single implicit neural network trained on clean ×4 pairs claims arbitrary scaling and zero-shot noise removal.","key_machinery":"The load-bearing object is a multi-resolution implicit hash encoding over a 6-dimensional coordinate-texture space. For each target high-resolution coordinate, the network retrieves 64 neighboring hash-table features, combines them with learned adaptive weights, refines features from coarse to fine with a top-down attention mechanism, and decodes the result to RGB. Because the map from coordinates to RGB is continuous, the same weights can be queried at any upscaling factor; a pixel-error amplified loss additionally up-weights low-error pixels so fine detail is not smoothed away by MSE.","core_discovery":"The paper's central claim is that VR-INR is a single network that performs both arbitrary-scale video super-resolution and zero-shot denoising after being trained only on clean ×4 LR-HR pairs. The authors argue that representing video as a continuous function over a 6-dimensional spatial-temporal-texture space, with features read from a multi-resolution hash table and decoded by an MLP, is what lets the same model answer queries at ×2 through ×32 and remove unseen noise without retraining or motion estimation. They support this with quantitative comparisons showing large PSNR/SSIM advantages over VideoINR, VRT, IART, and SAVSR at in-distribution ×4 and out-of-distribution ×2/×8, and with zero-shot denoising results on DAVIS (e.g., around 31.5 dB at σ=30 versus roughly 19.9 dB for the strongest baseline).","pith_inferences":["The conclusion mentions per-video fine-tuning with clean LR-HR pairs, so the public numbers should be reported both with and without that adaptation to separate learned generalization from test-time fitting; this distinction is an open question the paper leaves implicit.","If the scale generalization is genuine, it likely comes from the multiresolution hash grids acting as a scale-aware dictionary; training the same architecture at multiple scales would be a natural stress test, since multi-scale training should improve the arbitrary-scale curve if the single-scale prior is already saturated.","Zero-shot denoising may be an emergent bias of the coordinate decoder toward smooth, low-frequency outputs; measuring output high-frequency energy on noisy inputs, or adding a small noise-removal head, would clarify the mechanism."],"forward_implications":["Video super-resolution systems could serve arbitrary user-selected zoom factors from one trained model, eliminating per-scale retraining and model zoo maintenance.","Denoising could be obtained as a byproduct of clean-data super-resolution training, reducing the need for paired noisy and clean video data.","The flow-free design avoids optical-flow failures on occlusions and large motion, simplifying the restoration pipeline.","The pixel-error amplified loss is a transferable training objective for any coordinate-based reconstruction task where standard MSE over-smooths edges and textures."],"supporting_citations":[{"why":"Supplies the prior INR per-video denoising approach that VR-INR extends to zero-shot settings.","marker":"[1]"},{"why":"Establishes continuous coordinate-based image upsampling, the paradigm VR-INR generalizes to video.","marker":"[4]"},{"why":"VideoINR is the main video-INR baseline for arbitrary space-time super-resolution.","marker":"[5]"},{"why":"NeRV is the neural video representation baseline used in reconstruction and denoising comparisons.","marker":"[6]"},{"why":"SAVSR is the learned scale-adaptive arbitrary-scale video SR baseline compared in the experiments.","marker":"[17]"},{"why":"VRT is the transformer-based video restoration baseline compared at ×4 and ×8.","marker":"[18]"},{"why":"Instant-NGP's multiresolution hash encoding is the mechanism VR-INR adapts with learned interpolation.","marker":"[23]"},{"why":"IART is the implicit-resampling super-resolution baseline compared in the experiments.","marker":"[37]"},{"why":"Vid4 provides the main evaluation set for arbitrary-scale SR and reconstruction comparisons.","marker":"[20]"},{"why":"DAVIS provides the evaluation set for zero-shot denoising at multiple noise levels.","marker":"[27]"}],"fun_headline_variants":["Trained on ×4 only, this model upscales at any scale and denoises","A single INR does arbitrary-scale video SR plus zero-shot denoising","One implicit model: any-scale super-resolution, zero-shot denoising","VR-INR: one ×4-trained net for any upscale and unseen noise","Arbitrary-scale video restoration and zero-shot denoising in one INR"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the reported generalization comes from the single ×4-trained network; the conclusion says VR-INR is fine-tuned per video with clean LR-HR pairs, and if that per-video adaptation is required to reach the reported PSNR values, the zero-shot and arbitrary-scale claims would instead describe test-time fitting.","fun_headline_variants_meta":{"raw":{"variants":["Trained on ×4 only, this model upscales at any scale and denoises","A single INR does arbitrary-scale video SR plus zero-shot denoising","One implicit model: any-scale super-resolution, zero-shot denoising","VR-INR: one ×4-trained net for any upscale and unseen noise","Arbitrary-scale video restoration and zero-shot denoising in one INR"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000182,"raw_usage":{"total_tokens":1300,"prompt_tokens":921,"completion_tokens":379,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":537,"completion_tokens_details":{"reasoning_tokens":276}},"tokens_in":537,"tokens_out":379,"duration_ms":4093,"temperature":1.0,"reasoning_tokens":276,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T10:19:48.990205+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the published evaluation protocol exactly as described in Section 4.2 but disable any per-video fine-tuning before testing; if the ×2 and ×8 PSNR values on Vid4 exceed the baselines by the reported margins, the zero-shot claim stands, and if they fall to baseline levels, the claim is falsified.","supporting_citations":[{"cited_title":"Unsupervised coordinate-based video denoising","cited_arxiv_id":null,"evidence_quote":"Supplies the prior INR per-video denoising approach that VR-INR extends to zero-shot settings."},{"cited_title":"Learning continuous image representation with local implicit image function","cited_arxiv_id":null,"evidence_quote":"Establishes continuous coordinate-based image upsampling, the paradigm VR-INR generalizes to video."},{"cited_title":"Videoinr: Learning video implicit neural representation for continuous space-time super- resolution","cited_arxiv_id":null,"evidence_quote":"VideoINR is the main video-INR baseline for arbitrary space-time super-resolution."},{"cited_title":"Nerv: Neural representations for videos","cited_arxiv_id":null,"evidence_quote":"NeRV is the neural video representation baseline used in reconstruction and denoising comparisons."},{"cited_title":"Savsr: arbitrary-scale video super-resolution via a learned scale-adaptive network","cited_arxiv_id":null,"evidence_quote":"SAVSR is the learned scale-adaptive arbitrary-scale video SR baseline compared in the experiments."},{"cited_title":"Vrt: A video restoration transformer","cited_arxiv_id":null,"evidence_quote":"VRT is the transformer-based video restoration baseline compared at ×4 and ×8."},{"cited_title":"Instant neural graphics primitives with a multiresolution hash encoding","cited_arxiv_id":null,"evidence_quote":"Instant-NGP's multiresolution hash encoding is the mechanism VR-INR adapts with learned interpolation."},{"cited_title":"Enhancing video super-resolution via implicit resampling-based alignment","cited_arxiv_id":null,"evidence_quote":"IART is the implicit-resampling super-resolution baseline compared in the experiments."},{"cited_title":"A bayesian approach to adaptive video super resolution","cited_arxiv_id":null,"evidence_quote":"Vid4 provides the main evaluation set for arbitrary-scale SR and reconstruction comparisons."}],"review_version":1}