{"id":"73a756f7-81f4-4de3-ab71-58a4624124ff","arxiv_id":"2501.01681","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"SNeRV decomposes frames with wavelet transforms, embeds only low-frequency content, and regenerates high-frequency details, outperforming prior NeRV models on reconstruction and interpolation.","lead":"This paper proposes SNeRV, a neural network that stores a video by compressing only the low-frequency part of each frame and reconstructing fine details in a decoder. It reports sharper reconstructions and better video compression and interpolation than previous NeRV-style models.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Missing control: with DWT+IDWT an invertible linear map, the reported gains may come from the added modules/loss/compute rather than frequency decomposition.","rationale":"The reader's verdict is already CONDITIONAL, and its strongest_claim explicitly conditions the claim on the gains being caused by the frequency decomposition rather than extra supervision or hyperparameter selection. My concern sharpens that condition: because DWT+IDWT is a linear invertible transformation, SNeRV's decoder is mathematically equivalent to a direct RGB decoder with a reparameterized output space; the claimed 'spectra-preserving' mechanism is therefore not established unless the wavelet-specific components are isolated. The reader's weakest_assumption focuses on the predictability of HF from LF, which is a related but distinct issue: even if HF are predictable, the paper still needs a control showing that the wavelet parameterization, MFU/HFR modules, wavelet loss, or extra compute are what produce the gains. My concrete test (replace DWT/IDWT with an ordinary downsampling/upsampling path, keep the same backbone and loss except the wavelet term) directly settles this attribution question. If the control matches SNeRV, the central narrative collapses into generic decoder engineering; if it does not, the frequency-domain contribution is validated. The verdict should remain CONDITIONAL pending this control and code reproducibility, so I recommend UNCHANGED with respect to the reader's recommendation.","tokens_in":20738,"tokens_out":14083,"duration_ms":147393,"concrete_test":"Implement a control with the same SNeRV encoder and decoder backbone, but replace the HFR/IDWT output path with a single transposed-convolution block that upsamples the MFU output (at half resolution) directly to full-resolution RGB; feed the encoder the same 2×2 average-pooled image (identical to Haar CLL) so that the input information is unchanged. Train with the same L1+SSIM loss, but omit the wavelet-coefficient term L(C_t, Ĉ_t); keep the parameter count and epoch schedule identical. Evaluate on the UVG and DAVIS regression protocols of Table 1. If this control is within ~0.3 dB of Ours(B), the DWT/HFR/wavelet-loss specific components are not the cause of the gains; if it is more than ~1 dB below, the frequency-domain formulation is doing real work.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim attributes SNeRV's PSNR gains to its wavelet-based analysis–synthesis (Sec. 3.2). However, 2D Haar DWT followed by IDWT is a fixed, invertible linear transform: the decoder's output—CLL plus three HF subbands at half resolution—contains exactly the same number of values (12 × H/2 × W/2 = 3 × H × W) as a direct full-resolution RGB map. Any RGB-output decoder can be re-expressed as a wavelet-coefficient decoder via this linear map, and vice versa. Consequently, the wavelet decomposition itself cannot create new information or 'preserve spectra' in a representational sense; it only reparameterizes the output and adds the wavelet-domain loss L(C_t, Ĉ_t). The paper never ablates the DWT while keeping the rest constant: the comparisons to HNeRV/DNeRV differ simultaneously in (i) the DWT input, (ii) the MFU/HFR modules, (iii) the wavelet loss, and (iv) the compute budget (Ours(B) uses 90.49G MACs vs 60.89G for HNeRV and 48.39G for DNeRV, Table 2b). A matched-capacity decoder that outputs RGB directly (or uses a 2× average-pooled input and a simple learned upsampler) is the missing control. Absent that control, the reported improvements cannot be attributed to frequency decomposition rather than to the added modules, the extra loss term, or the higher compute.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes SNeRV, a neural representation for video that applies a 2D discrete wavelet transform to each frame, embeds only the low-frequency (LL) subband, and reconstructs the three high-frequency subbands with a decoder containing newly introduced multi-resolution fusion and high-frequency restoration modules. The loss in Eq. (3) combines a reconstruction loss on the output frame with a loss on the wavelet coefficients. A temporal extension applies an additional 1D DWT along time and uses temporally extended upsampling blocks. Experiments on Bunny, UVG, and DAVIS report improved PSNR/MS-SSIM over NeRV, E-NeRV, HNeRV, and DNeRV for video regression, interpolation, compression, and inpainting, with ablation studies for the proposed modules and loss terms.","tokens_in":21110,"tokens_out":5162,"duration_ms":51786,"significance":"If the reported gains are reproducible and correctly attributed, SNeRV is a practical and compact NeRV backbone that improves reconstruction while providing a clear frequency-domain design rationale. The paper is strong in empirical breadth: it includes results across multiple datasets, resolutions, model sizes, training epochs, and tasks, and it releases code and detailed supplementary tables. The module-level ablations (MFU/HFR, TUB, loss terms) are useful. The main weakness is that the central attribution of the gains to the wavelet decomposition is not tested with a matched control, and several quantitative claims would benefit from variance reporting and a more careful separation of design choices from evaluation-set tuning.","major_comments":[{"comment":"The central claim that the wavelet analysis-synthesis scheme is responsible for the PSNR gains is not established because the DWT is never ablated while keeping the rest of the model constant. Since 2D Haar DWT followed by IDWT is an invertible linear map, any RGB-output decoder can be re-expressed as a wavelet-coefficient decoder; the frequency decomposition itself cannot add information. The comparisons with HNeRV and DNeRV in Table 1 differ simultaneously in the DWT input, the MFU/HFR modules, the coefficient loss L(C_t, C_hat_t) in Eq. (3), and the compute budget (Ours(B) uses 90.49G MACs versus 60.89G for HNeRV and 48.39G for DNeRV, Table 2b). Please add a controlled ablation with identical modules, loss, and compute that outputs RGB directly (or a 2x-average-pooled input with a learned upsampler), and separately ablate the wavelet coefficient loss. Without such a control, the reported improvements cannot be attributed to the frequency decomposition rather than to the added modules, the extra loss term, or the higher compute.","section":"Sec. 3.2, Eq. (3), Table 2b"},{"comment":"Several hyperparameters that affect the headline results are selected on the same datasets used for evaluation. The number of residual blocks per MFB is set to NRB=6 on Bunny (Supplementary Table 3a), the TUB expansion rate is set to x2 on Bunny (Supplementary Table 3b), and the loss weight alpha=0.7 and the L1+SSIM combination are tuned on Yacht (Supplementary Table 3c); these datasets also appear in the main comparisons in Tables 1-3. This selection procedure creates a risk of overfitting to the evaluation set and makes the reported margins harder to interpret. Please either report performance for the selected hyperparameters on held-out sequences or show that the conclusions are insensitive to these choices.","section":"Sec. 4.4 and Supplementary Tables 3a-3c"},{"comment":"All quantitative results are reported as single numbers without error bars or multiple seeds. Some of the claimed improvements are small relative to the typical run-to-run variation in NeRV training (e.g., Tab. 1 UVG 960, Beauty 34.43 vs 34.12; Tab. 3 UVG, Bosph 36.63 vs 35.57), so it is unclear which differences are statistically significant. Please provide at least three seeds with standard deviations for the main tables, or otherwise justify that the training procedure has negligible variance.","section":"Tables 1-3"}],"minor_comments":[{"comment":"There are several typographical issues: \"T able 1\", \"T able 2\", and \"T able 3\" in the main text should be \"Table 1\", \"Table 2\", and \"Table 3\", and \"MS-SIM\" in Supplementary Table 3b should be \"MS-SSIM\".","section":"Throughout"},{"comment":"The term \"In-painting\" should be \"inpainting\" to match standard terminology and the spelling used in most of the paper.","section":"Sec. 4.2, Table 4"},{"comment":"The sentence \"Ours(B) has 31.3 decoding fps, which is relatively slower than 48.4 fps of HNeRV but is significantly faster than 1.75 fps of DCVC\" would be clearer with the measurement conditions (GPU, sequence, resolution) stated in the same sentence.","section":"Sec. 4.2"},{"comment":"Figure 7 would benefit from axis labels and a legend; the text refers to PSNR of LF and HF components, but the figure does not define how those components are scored.","section":"Fig. 7"}],"recommendation":"major_revision","confidential_remarks":"The missing control identified in the stress-test note is the primary barrier to acceptance; the requested control is feasible within the manuscript's scope. The paper is otherwise clearly written and the supplementary material is extensive. I would also ask the authors to address the inpainting results in Supplementary Table 10, where Ours(B) is worse than HNeRV and DNeRV on some sequences (e.g., dog-agility and drift-chicane), even though the main text emphasizes average gains. This does not affect my recommendation but should be clarified in revision."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Read SNeRV. The short version: it's a solid empirical methods paper for neural video representation, and the gains are probably real, but the headline claim—that the frequency decomposition itself is what does the work—is not actually tested. The ablation table shows MFU and HFR help, but there's no control that swaps DWT for a plain downsampled input while keeping everything else fixed. So the 'spectra-preserving' framing is a bit of an overreach.\n\nWhat's genuinely new: applying 2D DWT to NeRV, embedding only the LL band, and having the decoder synthesize the three HF bands via the HFR, with the MFU fusing multi-scale features. That combination isn't in NeRV, E-NeRV, HNeRV, or DNeRV. The temporal extension with 1D DWT over frame pairs and TUBs is a clever addition, and it gives a clean regression-vs-interpolation trade-off: the backbone wins at regression, the temporal version wins at interpolation. That's an interesting and reportable finding.\n\nThe experiments are extensive: UVG, DAVIS, Bunny, multiple model sizes, multiple training epochs, plus compression and inpainting. They retrained the baselines with the original code and kept total model size roughly matched. The PSNR improvements over HNeRV/DNeRV are consistent, around 1.3–2.4 dB in regression, and the interpolation gains for Ours(T) over DNeRV are meaningful. For an empirical paper, that's a real contribution.\n\nThe soft spots: the stress-test about the missing control is fair. DWT+IDWT is an invertible linear map, so any RGB-output decoder can be re-expressed in wavelet-coefficient space. The paper never isolates the DWT input from the new modules, the wavelet-domain loss, or the higher compute (Ours(B) uses 90.5G MACs vs 60.9G for HNeRV and 48.4G for DNeRV). So the claim that frequency decomposition alleviates spectral bias is supported only circumstantially. The NTK discussion is motivation, not proof; Fig. 7 shows training curves, not an NTK eigenvalue analysis. Also no error bars or repeated seeds, and hyperparameters like NRB, α, and embedding sizes look tuned on the same datasets used for the headline numbers. These are fixable in revision.\n\nBottom line: the paper deserves a serious referee. I'd send it out, but the reviewers should ask for a matched-capacity control without DWT, plus a few seeds and a more careful attribution of the gains. It's a useful backbone for NeRV-based compression and interpolation, and the architecture is worth building on, but the 'spectra-preserving' story needs to be tempered or backed with real evidence.","headline":"Solid empirical NeRV paper with a genuine architectural novelty, but the frequency-decomposition attribution is not isolated; send to review and demand a matched-capacity control.","tokens_in":21605,"tokens_out":3583,"would_cite":true,"duration_ms":33970,"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":"SNeRV keeps only each frame's low-frequency band and regenerates the fine detail, beating previous video networks by up to 2.35 dB in reconstruction.","keywords":["implicit neural representation","neural representation for video","discrete wavelet transform","spectral bias","high-frequency restoration","video regression","video interpolation","video compression"],"falsifier":"Take a video of randomly flickering high-frequency noise overlaid on a uniform background: each frame's wavelet decomposition has near-zero low-frequency content but large high-frequency detail coefficients. If SNeRV's PSNR on such a sequence collapses while a method that embeds all four wavelet bands stays high, that shows the low-frequency-prior assumption is load-bearing; if SNeRV still recovers the noise, the restorer is doing more than the paper's sparsity argument claims.","tokens_in":20546,"feed_emoji":"🎞️","tokens_out":10412,"duration_ms":87927,"temperature":0.7,"pith_summary":"This paper tries to show that the main obstacle to high-quality neural video representation—a family of methods (NeRV) that stores a video inside a network's weights—is spectral bias, the tendency of networks to learn low-frequency content much faster than high-frequency edges and textures, and that the obstacle can be bypassed by decomposing each frame with a wavelet transform. SNeRV splits a frame into one low-frequency band and three high-frequency bands, embeds only the low-frequency band, and trains a decoder to regenerate the high-frequency bands from that low-frequency prior. Because natural-video high-frequency bands are sparse, the decoder can synthesize them with few parameters, so the model stays compact while capturing fine spatial detail and motion. On the UVG and DAVIS benchmarks the authors report average PSNR gains of about 1.3 to 2.4 dB over HNeRV and DNeRV at matched model size, and the temporal extension improves video interpolation by about 1.1 to 1.6 dB over DNeRV. If the gains come from the frequency split itself rather than extra supervision or tuning, this gives a general architecture-level cure for spectral bias in implicit video representations.","feed_headline":"Storing only the blur beats prior video networks by up to 2.35 dB","feed_subtitle":"A wavelet split lets a small network regenerate edges and textures that other NeRV models lose to spectral bias.","key_machinery":"The carrying mechanism is an analysis-and-synthesis wavelet pipeline. Each frame goes through a 2D discrete wavelet transform (Haar filters) that produces $C_{LL}$ (low-frequency approximation) and $C_{LH}$, $C_{HL}$, $C_{HH}$ (horizontal, vertical, and diagonal detail); only $C_{LL}$ is embedded, which is the point where compactness is bought. In the decoder, the multi-resolution fusion unit (MFU) progressively refines the low-frequency features by fusing outputs of several up-sampling blocks through transposed convolutions and residual blocks, and the high-frequency restorer (HFR)—three small two-layer convolution blocks—generates the three high-frequency subbands from the fused low-frequency representation. Adding a 1D DWT along time yields temporal embeddings $e_{\\Delta t}^{b}$ and $e_{\\Delta t}^{f}$, processed by temporally extended up-sampling blocks (TUBs) that merge neighboring-frame information before the MFU and HFR finish the frame. The load-bearing property is sparsity: because wavelet detail coefficients are near zero in flat regions, a low-capacity restorer can plausibly synthesize them from the low-frequency prior, which is what lets the network avoid spectral bias without growing the model.","core_discovery":"The paper's central claim is that a neural representation for video becomes both more accurate and more compact when the frequency bands are handled separately rather than learned together. A 2D Haar discrete wavelet transform decomposes each frame into the low-frequency approximation $C_{LL}$ and three high-frequency detail bands $C_{LH}$, $C_{HL}$, $C_{HH}$. The encoder embeds only $C_{LL}$, because low-frequency content is what an implicit network already learns readily; the decoder then reconstructs the high-frequency bands from that low-frequency prior using a multi-resolution fusion unit (MFU) that refines features across scales and a high-frequency restorer (HFR) that outputs the three detail bands, after which an inverse wavelet transform recombines all four bands into the frame. The authors argue that the high-frequency bands of natural video are sparse, significant only near edges, so synthesizing them from the low-frequency prior costs fewer parameters than learning them directly, and the freed capacity can be spent on representing the low-frequency structure. The temporal extension adds a 1D wavelet transform along the time axis, embeds spatio-temporal low-frequency features with temporally extended up-sampling blocks (TUBs), and treats motion as high-frequency temporal detail to be generated rather than stored. The evidence offered is consistent gains in PSNR on UVG and DAVIS for the backbone and in interpolation for the temporal extension, with the learning-curve analysis in Fig. 7 showing that the high-frequency components are exactly where previous methods stagnate and SNeRV improves.","pith_inferences":["The same analysis-and-synthesis split should transfer to other implicit neural representations that suffer spectral bias, such as coordinate-MLP scene reconstruction or audio wavefields, wherever the high-frequency residual is sparse with respect to a low-frequency backbone; this extension is not tested in the paper.","The documented trade-off between time duration and spectral band suggests the embedding-size ratio is a continuous control: a deployment could choose a different point on the trade-off curve for compression versus interpolation, but the paper does not propose how to select that point automatically.","If the sparsity of wavelet detail coefficients is the true source of the gain, then on videos where high-frequency content is dense and decorrelated from low-frequency content—random sensor noise, static, or fine repeated micro-texture—the HFR should degrade sharply; this is a testable boundary condition for the method.","The comparison pipeline uses the same pruning and quantization as previous work, but the paper's own supplementary shows NNCodec compression preserves much of the advantage; this implies the benefit is not an artifact of a particular coding pipeline, an inference the authors state only implicitly."],"forward_implications":["At matched total model size, the SNeRV backbone reports higher regression PSNR than NeRV, E-NeRV, HNeRV, and DNeRV on UVG and DAVIS, including gains of 1.32–1.46 dB at 960p and 2.17–2.35 dB on the DAVIS subsets, so the frequency split is claimed as a general regression improvement.","The temporal extension, Ours(T), reports the best interpolation results among the tested NeRV variants, beating DNeRV by roughly 1.14 dB on UVG and 1.62 dB on DAVIS, because the embedded spatio-temporal low-frequency features support continuous time modeling.","The decoded high-frequency components improve with more training where prior methods saturate, which the paper reads as direct evidence that spectral bias has been mitigated rather than merely re-parameterized.","The same backbone integrates into HiNeRV's coding pipeline and improves BD-PSNR by about 0.3 dB, and NNCodec compression of the decoder yields an average 14.10% BD-rate saving on UVG, indicating the scheme is compatible with existing neural-video compression tools.","The backbone and its temporal extension trade regression performance against interpolation performance, so the architecture offers a tunable balance between spatial spectral fidelity and temporal consistency."],"supporting_citations":[{"why":"Supplies the hybrid content-adaptive embedding plus learned decoder framework that SNeRV's encoder and decoder are built on, and serves as the primary regression baseline.","marker":"[8]"},{"why":"Defines the original frame-wise NeRV architecture and the up-sampling NeRV blocks with positional encoding that SNeRV reuses for its up-sampling blocks.","marker":"[9]"},{"why":"Provides the difference-based dynamic modeling and the temporal-extension idea that SNeRV's 1D DWT temporal branch builds on; also the main interpolation and compression comparison.","marker":"[50]"},{"why":"The disentangled spatial-temporal E-NeRV baseline that SNeRV compares against in regression and interpolation.","marker":"[25]"},{"why":"Gives the neural tangent kernel analysis of spectral bias that motivates separating low- and high-frequency learning.","marker":"[7]"},{"why":"Provides the result that neural networks converge at different rates for functions of different frequencies, which the paper cites as the mechanism behind spectral bias.","marker":"[38]"},{"why":"Supplies the wavelet transform theory and Haar filter definitions used for the 2D and 1D decompositions.","marker":"[29]"},{"why":"Provides the PixelShuffle operation used inside the up-sampling blocks that the decoder inherits.","marker":"[40]"}],"fun_headline_variants":["Wavelet split lets neural video store only blur, regenerate edges","Neural video codec beats spectral bias with wavelet frequency split","Wavelet decomposition halves storage, boosts video reconstruction by 2.35 dB","Trick: store low-frequency, generate texture — neural video gains 2.35 dB","Frequency split in NeRV: store blur, synthesize edges, beat prior models"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"Everything depends on the assumption that the fine detail in a video can be recreated from its blurred, low-frequency version: if a scene contains textures whose high-frequency content carries information that the low-frequency band does not contain, encoding only the low-frequency band discards it permanently.","fun_headline_variants_meta":{"raw":{"variants":["Wavelet split lets neural video store only blur, regenerate edges","Neural video codec beats spectral bias with wavelet frequency split","Wavelet decomposition halves storage, boosts video reconstruction by 2.35 dB","Trick: store low-frequency, generate texture — neural video gains 2.35 dB","Frequency split in NeRV: store blur, synthesize edges, beat prior models"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000756,"raw_usage":{"total_tokens":3458,"prompt_tokens":1143,"completion_tokens":2315,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":759,"completion_tokens_details":{"reasoning_tokens":2217}},"tokens_in":759,"tokens_out":2315,"duration_ms":16137,"temperature":1.0,"reasoning_tokens":2217,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T22:22:12.048885+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a video of randomly flickering high-frequency noise overlaid on a uniform background: each frame's wavelet decomposition has near-zero low-frequency content but large high-frequency detail coefficients. If SNeRV's PSNR on such a sequence collapses while a method that embeds all four wavelet bands stays high, that shows the low-frequency-prior assumption is load-bearing; if SNeRV still recovers the noise, the restorer is doing more than the paper's sparsity argument claims.","supporting_citations":[{"cited_title":"In: IEEE Conf","cited_arxiv_id":null,"evidence_quote":"Supplies the hybrid content-adaptive embedding plus learned decoder framework that SNeRV's encoder and decoder are built on, and serves as the primary regression baseline."},{"cited_title":"Advances in Neural Information Processing Systems 34, 21557–21568 (2021)","cited_arxiv_id":null,"evidence_quote":"Defines the original frame-wise NeRV architecture and the up-sampling NeRV blocks with positional encoding that SNeRV reuses for its up-sampling blocks."},{"cited_title":"In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition","cited_arxiv_id":null,"evidence_quote":"Provides the difference-based dynamic modeling and the temporal-extension idea that SNeRV's 1D DWT temporal branch builds on; also the main interpolation and compression comparison."},{"cited_title":"In: European Confer- ence on Computer Vision","cited_arxiv_id":null,"evidence_quote":"The disentangled spatial-temporal E-NeRV baseline that SNeRV compares against in regression and interpolation."},{"cited_title":"Advances in Neural Infor- mation Processing Systems32 (2019)","cited_arxiv_id":null,"evidence_quote":"Provides the result that neural networks converge at different rates for functions of different frequencies, which the paper cites as the mechanism behind spectral bias."},{"cited_title":"Elsevier (1999)","cited_arxiv_id":null,"evidence_quote":"Supplies the wavelet transform theory and Haar filter definitions used for the 2D and 1D decompositions."},{"cited_title":"In: Proceedings of the IEEE/CVF conference on com- puter vision and pattern recognition","cited_arxiv_id":null,"evidence_quote":"Provides the PixelShuffle operation used inside the up-sampling blocks that the decoder inherits."}],"review_version":1}