{"id":"10ea209b-03aa-4d86-8ad0-95360ed471f7","arxiv_id":"2501.09277","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"Sharing an INR's weights across frames and modulating only per-frame biases via a time-conditioned hypernetwork gives a compact continuous video representation that outperforms prior video INRs on interpolation, super-resolution, denoising, and inpainting.","lead":"ActINR represents video with a neural network that keeps the same weights for every frame and changes only its bias values, which are produced by a second small network for each time instant. This simple design yields smooth continuous video representations and shows strong results on slow-motion, super-resolution, denoising, and inpainting benchmarks.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Block-locality assumption directly contradicts the paper's claim to model large motions; the central 'continuous video representation' result is therefore bounded by object displacement relative to patch size.","rationale":"The Reader's weakest_assumption correctly identifies the block-size assumption. My stress-test sharpens it: the assumption is not merely a limitation but directly contradicts the Introduction's 'small and large motions alike' claim, and the main quantitative experiments do not include the overlapping-window remedy that would mitigate it. I do not think this invalidates the empirical comparisons on the tested benchmarks; a conditional verdict remains appropriate, with the claim of universal large-motion capability removed or the method evaluated with overlapping windows. Additional secondary concerns (qualitative-only 10x interpolation, single-baseline denoising, missing error bars) reinforce conditionality but are less load-bearing than the architectural locality bound.","tokens_in":15484,"tokens_out":9874,"duration_ms":92837,"concrete_test":"Build a synthetic sequence of a single Gaussian blob translating horizontally at constant speed, with total displacement per 10-frame GOP ranging from 0 to 2 patch widths (96x96 patches). Train ActINR with the paper's default settings and report interpolation PSNR on held-out frames as a function of displacement. If PSNR degrades sharply once the blob crosses a patch boundary, the Section 5 limitation is confirmed as the operative bound on the large-motion claim; if PSNR stays flat, the concern is refuted.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The Introduction claims ActINR 'can model small and large motions alike,' but the method's patchwise formulation and its own limitation statement bound it to local motion. Section 5 states: 'Our approach assumes that motion remains confined within a designated block size.' Eq. (2) assigns per-patch bias vectors to a shared-weight INR, and the implementation independently fits 96x96x10 blocks; the supplementary failure case (Fig. 13) shows that when an object crosses a patch boundary, reconstruction degrades and patch transitions exhibit artifacts. The bias-INR only modulates biases within each patch; it has no mechanism to transfer an object's identity from one patch's function to the next. The overlapping-window remedy (Fig. 14) is not used in the main quantitative comparisons. Since the headline capability is a continuous video representation for interpolation and super-resolution, and those tasks involve large displacements, the central claim is conditional on displacement staying inside the block size.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes ActINR, a video implicit neural representation in which the weights of a frame INR are shared across frames while per-frame and per-layer bias vectors are predicted by a separate bias-INR conditioned on a continuous time index. Videos are processed in independent 96x96x10 blocks with WIRE activations, and the bias-INR is shared across patches through a per-patch latent code. The manuscript reports quantitative gains over prior video INRs on 2x interpolation, inpainting, denoising, and space-time super-resolution on UVG and DAVIS, plus a qualitative 10x slow-motion result and ablations on activation choice and on the bias-INR versus linear interpolation of biases. The paper also states an explicit limitation that motion must remain within the block size and includes a supplementary failure case showing artifacts when objects cross patch boundaries.","tokens_in":15687,"tokens_out":6472,"duration_ms":65510,"significance":"ActINR is a conceptually clean and potentially impactful reformulation of video INRs: instead of injecting time through an extra coordinate, residual weights, or optical flow, it represents motion as bias modulation of compact basis functions and predicts the biases with a small hypernetwork. The held-out evaluation is legitimate: frames used for interpolation are unseen during training, and the oracle/linear-interpolation ablation in Table 4 directly supports the choice of a learned continuous bias predictor. If the reported gains survive repeated runs and code release, ActINR would be a strong baseline for continuous video representation and for inverse problems such as denoising and inpainting. The main caveats are the acknowledged block-locality assumption, the absence of error bars and code, and the qualitative-only treatment of the advertised 10x slow-motion and part of the space-time super-resolution experiments.","major_comments":[{"comment":"The limitation statement in Section 5 ('Our approach assumes that motion remains confined within a designated block size') directly contradicts the Introduction's claim that ActINR 'can model small and large motions alike.' Because the method fits independent 96x96x10 blocks and Eq. (2)'s per-patch biases cannot transfer an object's basis functions across patch boundaries, the central interpolation and space-time super-resolution claims are conditional on all motion between training frames staying within roughly 96 pixels. The paper's own supplementary failure case (Fig. 13) demonstrates artifacts when an object crosses a patch boundary, and the overlapping-window remedy (Fig. 14) is not used in any main quantitative comparison. Please add a quantitative characterization of per-sequence motion magnitudes relative to the patch and GOP sizes used in Tables 1 and 3, and either restrict the large-motion claims accordingly or incorporate the overlapping-window strategy into the main evaluation.","section":"Section 5 (Limitations), Section 3, Eq. (2), and supplementary Fig. 13"},{"comment":"All quantitative comparisons report a single run with no variance estimates, and no code is released. This is load-bearing for the SOTA claims, particularly in Table 1 on UVG: ActINR's average PSNR gain over H-NeRV Boost is only 0.4 dB, and ActINR is worse on Honey (37.6 vs 38.8) and Ready (25.9 vs 26.6). With a single run, these differences may be within run-to-run noise. Please release code and report mean and standard deviation over at least three seeds for the main comparisons, or otherwise provide per-sequence repeated-run numbers.","section":"Tables 1-3 and 5"},{"comment":"The headline capabilities of '10x video slow motion' and space-time super-resolution lack quantitative tables. Figure 5 is qualitative only, and Figure 8 reports PSNR numbers in the caption for a single scene (Bosphorus) without stating how the 10x or 8x targets are generated, how baselines are trained for those factors, or how the GOP is defined when every tenth frame is skipped. Since the abstract and introduction advertise these capabilities, they need formal evaluation with tables, including the exact temporal sampling protocol and baseline configurations.","section":"Section 4.1, Figure 5, and Figure 8"},{"comment":"The denoising setup is internally inconsistent: the text says input PSNR is between 16 and 18 dB, the Implementation Details say the noise settings produce approximately 20 dB, and the Figure 6 captions state 18.6 dB and 16.6 dB in different places. Moreover, Table 3 compares only a single baseline. Please correct the noise-level reporting, state the exact photon-count and readout-noise parameters used for each sequence, and justify the single-baseline comparison for a task that is presented as one of ActINR's main contributions.","section":"Section 4.1 (Denoising) and Implementation Details"}],"minor_comments":[{"comment":"The first line of Eq. (3) reads y_t^(0) = [x, y]^T, but the bias-INR is supposed to be conditioned on time, not spatial coordinates; this appears to be a typo and should read y_t^(0) = t (or the RFF of t).","section":"Section 3, Eq. (3)"},{"comment":"The latent code z_l in Eq. (3) is written with the subscript l, which clashes with the layer index l used in Eqs. (1) and (2); please rename it, for example z_i for patch i, to avoid confusion.","section":"Section 3, Eq. (3)"},{"comment":"The baseline name is inconsistent across the paper: H-NeRV, HNeRV, H-NeRV Boost, and HBoostNeRV all refer to the same method; please unify the notation.","section":"Throughout"},{"comment":"References [18] and [19] are identical (both are 'Motion-adjustable neural implicit video representation'); the Phase-INR citation needs a distinct entry.","section":"References"},{"comment":"The dataset is referred to as 'DA VIS' throughout; the standard name is DAVIS, and the spacing should be corrected.","section":"Dataset names"}],"recommendation":"major_revision","confidential_remarks":"The paper's core idea is plausible and the held-out evaluation is legitimate; I do not see a circularity problem. The main risks are that the SOTA comparisons rest on single runs and that the block-locality assumption is acknowledged but not quantified against the motion content of the test sequences. Both are fixable within the manuscript's scope, so I recommend major revision rather than rejection. Code release and repeated-seed results would substantially increase confidence in the reported gains."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"You should know this paper has a genuinely simple and plausible idea: in a video INR with compact activations, keep the weights fixed and let per-frame biases carry the motion. They implement it with a small hypernetwork that maps time to bias vectors, share the frame-INR weights across all frames, and fit the whole thing patch-wise. On the evidence in front of me, the idea works: the interpolation, space-time super-resolution, inpainting, and denoising tables all show consistent gains over the listed baselines, often by several dB, and the ablations support the bias-motion story rather than just asserting it. That is real credit.\n\nThe novelty is a combination, not a new paradigm. Mod-SIREN, Phase-INR, and ResFields already modulated weights, phase, or residual matrices; ActINR's contribution is restricting the modulation to biases and generating them from a time-conditioned INR. That is a reasonable and well-executed step, though the underlying \"biases control basis location\" observation is familiar from the dictionary-perspective literature they cite. The citation pattern looks fair and the math is straightforward.\n\nThe soft spots are mostly about support, not the core mechanism. No code release, no error bars or multiple seeds, the 10x slow-motion and extreme-interpolation results are qualitative only, and the denoising comparison has a single baseline (HNeRV-Boost). The larger tension is the one the paper itself flags: motion is assumed to stay inside a 96x96x10 block. The introduction says the method \"can model small and large motions alike,\" but the block-locality assumption is load-bearing. The supplementary failure case (Fig. 13) shows exactly the expected artifact when an object crosses a patch boundary, and the overlapping-window remedy is not used in the main quantitative comparisons. So the claim of large-motion modeling is true only relative to block size. That is an overstatement in the intro, not a fatal flaw: the method does handle large interpolation factors on the tested sequences, and the authors are honest in the limitations section.\n\nWho this is for: anyone working on video INRs or implicit video representations. The paper gives a clean recipe (modulate biases, share weights, use WIRE) and a clearly stated failure mode. It deserves a serious referee. For the review, I'd ask for code, multi-seed results, and a quantitative extreme-interpolation table before accepting, but the central idea holds up. Send it out.","headline":"Solid, well-motivated video INR paper with a genuine bias-modulation idea, let down mainly by missing code/error bars and an intro that overclaims large-motion handling beyond its block-local design.","tokens_in":16213,"tokens_out":2535,"would_cite":true,"duration_ms":26223,"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":"Video motion lives in the bias term, claims ActINR","keywords":["implicit neural representation","video representation","bias modulation","temporal interpolation","space-time super-resolution","video denoising","video inpainting","wavelet activation"],"falsifier":"Train ActINR on a synthetic video of a bright dot translating at constant speed across a 96x96 patch boundary, holding out intermediate frames for interpolation; if the interpolated dot trajectory breaks or ghosts at the seam, or if PSNR drops sharply as the crossing fraction increases, the block-confinement and bias-location assumptions would be violated.","tokens_in":15282,"feed_emoji":"🎥","tokens_out":3821,"duration_ms":42740,"temperature":0.7,"pith_summary":"The paper proposes ActINR, a video implicit neural representation that keeps one set of network weights for all frames and encodes each frame's motion only in the network's bias vectors. It argues that an INR with compactly supported activations behaves like a dictionary of localized basis functions, where weights set the shapes and biases set the locations, so translating bases through biases is enough to represent moving scene content. A second INR, conditioned on continuous time, outputs the biases, making the representation smooth in time and queryable at arbitrary timestamps. The authors claim this design outperforms prior video INRs on temporal interpolation, space-time super-resolution, denoising, and inpainting, with reported gains of several dB.","feed_headline":"Video motion lives in the bias term, claims ActINR","feed_subtitle":"A shared-weight INR with per-frame biases yields 10x slow motion and gains up to 8.5 dB.","key_machinery":"The load-bearing mechanism is bias modulation in a shared-weight INR. Each patch-wise frame INR computes $y^{(l)}_i = \\sigma(W^{(l)} y^{(l-1)}_i + b^{(l)}_i)$, with weights $W^{(l)}$ shared across all frames and patches, while a bias INR $\\psi$ maps $(\\gamma(t), z_i)$ to the per-layer bias vectors $b^{(l)}_i$, where $\\gamma(t)$ is a random Fourier feature encoding of time and $z_i$ is a patch-specific latent code. The activation $\\sigma$ is WIRE, a wavelet activation with compact spatial support, which is what makes a basis-location interpretation approximately valid; the paper's ablation shows that non-compact activations like SIREN spread motion artifacts into static regions. The representation is trained block-wise, with each 96x96 spatial patch over a short frame group handled by its own small INR, following a divide-and-conquer strategy that keeps the motion local and the parameter count manageable.","core_discovery":"ActINR's central claim is that the motion in a video can be carried entirely by bias modulation: with a compactly supported activation such as WIRE, an INR is a learned dictionary whose basis shapes are fixed by the weights and whose positions are set by the biases, so shifting a basis function is equivalent to shifting a moving object's location. The paper's design shares all frame-INR weights across frames and assigns each frame a unique bias vector produced by a separate bias-INR hypernetwork that takes a continuous time index, random Fourier features, and a patch latent code as input. Training the frame INR and the bias INR together with a pixel-wise MSE loss yields a continuous video representation that supports arbitrary temporal queries and spatial queries. The authors report that this representation outperforms existing video INRs on 2x and 10x temporal interpolation, on combined spatial and temporal super-resolution (up to 8x spatial with 2x temporal), on extreme denoising, and on video inpainting, with improvements often exceeding 6 dB.","pith_inferences":["Beyond the paper, the per-frame bias sequence itself could serve as a compact motion descriptor, potentially supporting motion editing, motion transfer, or a lightweight optical-flow substitute by tracking basis centers over time.","The block-confinement failure suggests a testable extension: hierarchical or adaptive patch assignment, or allowing bases to be re-assigned across blocks, could extend the method to large global motions without the reported patch-boundary artifacts.","The bias-INR-as-regulator idea is a general trick: any INR-based video or dynamic-scene model could impose temporal smoothness by modulating biases through a time-conditioned hypernetwork, rather than by introducing explicit flow or residual weights.","One could test the basis-location hypothesis directly by extracting the bias time series from a video with known rigid motion, e.g., a translating dot, and checking whether the inferred basis centers track the true trajectory linearly."],"forward_implications":["Temporal interpolation becomes a query of the bias INR at unseen time indices, producing 10x slow-motion video with reported PSNR gains of up to 5 dB over baselines.","Arbitrary spatial queries remain available because the frame INR maps continuous coordinates, so the same model performs 4x spatial upsampling combined with 2x temporal upsampling, with reported gains up to 8.5 dB.","The smoothness enforced by the bias INR acts as an implicit temporal regularizer, which the paper shows yields strong denoising (3-5 dB over a baseline) and improved inpainting (about 1.6 dB average) without explicit noise or motion models.","The representation remains competitive on standard video regression and is faster at decoding than some image-wise baselines, while the authors note it trails specialized methods on compression.","Because biases are the only per-frame parameters, encoding cost is dominated by the shared frame INR, and the authors report per-video encoding and decoding times comparable to prior video INRs."],"supporting_citations":[{"why":"Supplies the WIRE wavelet activation with compact support, which makes the bias-as-location interpretation valid and is used in all experiments.","marker":"[37]"},{"why":"Frames an INR as a basis-function expansion, grounding the claim that weights control basis shape.","marker":"[35]"},{"why":"Also frames INRs through a dictionary perspective, supporting the weight-shape and bias-location decomposition.","marker":"[48]"},{"why":"Provides the divide-and-conquer patch-wise INR strategy that ActINR adapts for local motion modeling and tractability.","marker":"[34]"},{"why":"Introduces NeRV, the baseline video INR paradigm that ActINR extends by sharing weights while varying biases.","marker":"[3]"},{"why":"FF-NeRV is the optical-flow-based interpolation baseline that ActINR compares against on interpolation tasks.","marker":"[14]"},{"why":"Boosting NeRV is the main image-wise baseline for interpolation, denoising, inpainting, and compression comparisons.","marker":"[49]"},{"why":"Phase-INR is the pixel-wise baseline for space-time super-resolution, which ActINR outperforms across upscaling factors.","marker":"[19]"},{"why":"ResFields is the pixel-wise residual-field baseline compared on 2x interpolation.","marker":"[24]"},{"why":"DS-NeRV is the baseline for video inpainting comparisons.","marker":"[47]"}],"fun_headline_variants":["Motion lives in biases: ActINR's video trick","Shifting biases, not weights, makes video INRs move","ActINR: Video motion encoded entirely in bias vectors","10x slow motion from bias-only video INRs"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The paper's approach assumes that a video's motion can be represented by translating compactly supported basis functions through bias changes alone, with the network weights fixed; the authors state the practical version of this assumption as motion remaining confined within a designated block size, so objects that cross patch boundaries degrade the representation.","fun_headline_variants_meta":{"raw":{"variants":["Motion lives in biases: ActINR's video trick","Shifting biases, not weights, makes video INRs move","ActINR: Video motion encoded entirely in bias vectors","10x slow motion from bias-only video INRs"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000476,"raw_usage":{"total_tokens":2363,"prompt_tokens":950,"completion_tokens":1413,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":566,"completion_tokens_details":{"reasoning_tokens":1355}},"tokens_in":566,"tokens_out":1413,"duration_ms":10673,"temperature":1.0,"reasoning_tokens":1355,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T20:05:45.973082+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train ActINR on a synthetic video of a bright dot translating at constant speed across a 96x96 patch boundary, holding out intermediate frames for interpolation; if the interpolated dot trajectory breaks or ghosts at the seam, or if PSNR drops sharply as the crossing fraction increases, the block-confinement and bias-location assumptions would be violated.","supporting_citations":[{"cited_title":"Wire: Wavelet implicit neural representations","cited_arxiv_id":null,"evidence_quote":"Supplies the WIRE wavelet activation with compact support, which makes the bias-as-location interpretation valid and is used in all experiments."},{"cited_title":"Implicit Neural Representations and the Algebra of Complex Wavelets","cited_arxiv_id":"2310.00545","evidence_quote":"Frames an INR as a basis-function expansion, grounding the claim that weights control basis shape."},{"cited_title":"A structured dictionary perspective on implicit neural representations","cited_arxiv_id":null,"evidence_quote":"Also frames INRs through a dictionary perspective, supporting the weight-shape and bias-location decomposition."},{"cited_title":"Kilonerf: Speeding up neural radiance fields with thousands of tiny mlps","cited_arxiv_id":null,"evidence_quote":"Provides the divide-and-conquer patch-wise INR strategy that ActINR adapts for local motion modeling and tractability."},{"cited_title":"Nerv: Neural representations for videos","cited_arxiv_id":null,"evidence_quote":"Introduces NeRV, the baseline video INR paradigm that ActINR extends by sharing weights while varying biases."},{"cited_title":"Ffnerv: Flow-guided frame-wise neural representa- tions for videos","cited_arxiv_id":null,"evidence_quote":"FF-NeRV is the optical-flow-based interpolation baseline that ActINR compares against on interpolation tasks."},{"cited_title":"Boosting neu- ral representations for videos with a conditional decoder","cited_arxiv_id":null,"evidence_quote":"Boosting NeRV is the main image-wise baseline for interpolation, denoising, inpainting, and compression comparisons."},{"cited_title":"Motion-adjustable neural implicit video representation","cited_arxiv_id":null,"evidence_quote":"Phase-INR is the pixel-wise baseline for space-time super-resolution, which ActINR outperforms across upscaling factors."},{"cited_title":"Ds-nerv: Implicit neural video representation with decomposed static and dynamic codes","cited_arxiv_id":null,"evidence_quote":"DS-NeRV is the baseline for video inpainting comparisons."}],"review_version":1}