{"id":"5f8ed7cc-3c86-4010-a2f9-a125a690e581","arxiv_id":"2607.15298","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"A single SIREN MLP can store a short video plus audio and be compressed to 2.33 MiB, but it remains far less efficient than H.264/HEVC/MP3 on every tested clip.","lead":"Researchers trained one neural network to hold a video and its audio in its weights, then shrank those weights with quantization and lossless compression. The compressed network is about one-third the size of one test video, but H.264, HEVC, and MP3 still beat it by one to two orders of magnitude in compression ratio.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"No matched-capacity separate-network baseline leaves the unified-representation claim unmeasured; a two-network control would settle whether shared layers hurt per-modality quality.","rationale":"Agree with the reader: the missing matched-capacity baseline is the most load-bearing gap. The paper is honest and internally consistent in its numbers; I verified Table I totals and the teacher/student tables. But the central novelty is the shared-stack architecture, and the paper does not measure whether sharing helps or hurts. The distillation anomaly (Section V-D) is not an isolated bug—it is the observable consequence of the unmeasured interference. A two-network control would directly test the claim and would also clarify whether the audio collapse is caused by the student's smaller branch or by sharing. I considered other concerns (no error bars on the bit-width sweep, cherry-picked compression ratio, fixed size floor) but these are secondary: the first is a robustness issue on one video, the second is acknowledged in the limitations, the third is an engineering trade-off. The internal contradiction between the conclusion's 'essentially unchanged' and the 38 dB audio drops is real, but it is a presentation flaw, not the load-bearing issue; the core claim about the unified representation is what needs the baseline. Therefore the reader's CONDITIONAL verdict remains appropriate.","tokens_in":9625,"tokens_out":7909,"duration_ms":81435,"concrete_test":"Train two separate SIRENs—one audio-only, one video-only—with the same total parameters as the unified teacher (e.g., split the shared hidden stack accordingly), and run the exact same KD/16-bit/LZMA2 pipeline at matched file sizes. Compare per-modality PSNR/SSIM/LSD against the unified student for Videos 1 and 3. If the separate networks match or beat the unified student at matched size, the shared layers are not an efficient joint representation; if they are worse, the unified claim is supported.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central contribution is that a single sample-wise MLP with shared hidden layers can jointly represent video and audio (Section III-B), and the abstract reports the compressed student's quality. But the experiments never compare against the natural control: two modality-specific SIRENs with the same total parameter count, same distillation, and same quantization. Without that control, the shared hidden stack might contribute little; the actual fitting could be done in the separate initialization layers and branches. Table I shows the shared stack is ~66% of teacher parameters, but the paper provides no evidence that this sharing is beneficial or neutral. The paper's own distillation results (Tables V and VII) show a massive audio collapse—62.5 to 24.2 dB on Video 3 while video PSNR rises—consistent with capacity being reallocated from audio to video during KD. This is exactly the kind of interference the unified-representation claim must quantify. The introduction asks 'what it costs' (Section I), but the cost is never isolated. The conclusion's statement that quality is 'essentially unchanged' is also contradicted by these audio drops, further obscuring the pipeline's effect.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a video codec that stores an entire video-plus-audio clip as the weights of a single sinusoidal MLP (SIREN). The network has separate audio and video initialization layers, a shared hidden stack, one video branch, and two Siamese audio branches used for noise estimation. The overfitted teacher is compressed by knowledge distillation into a smaller student, followed by 16-bit symmetric quantization and LZMA2 encoding. The authors report per-modality quality metrics, a bit-width study, comparisons against H.264/HEVC and MP3, and a browser-based prototype. The central claim is that this is the first sample-wise, MLP-only representation in which one set of shared hidden layers serves both pixel and audio functions, and that the pipeline compresses the representation from 9.05 MiB to 2.33 MiB at an overall ratio of 2.61.","tokens_in":9923,"tokens_out":6813,"duration_ms":70948,"significance":"If the shared-hidden-layer architecture were properly validated, the paper would be a compelling proof-of-concept for resolution- and length-independent joint storage of audio and video in one implicit neural representation. The authors are unusually candid: they report negative comparisons with conventional codecs, disclose the fixed model-size floor, and explicitly acknowledge that distillation trades audio quality for video quality (Sec. VI). The parameter table (Table I), the distillation/quantization tables, and the bit-width sweep provide useful reproducibility anchors. However, the central contribution is currently under-supported because the experiments never isolate the effect of the shared stack from the per-modality components, and the headline compression ratio is mislabeled. These issues are fixable within the scope of a revision.","major_comments":[{"comment":"The paper's central novelty is that a single shared hidden stack can jointly represent audio and video in one plain MLP. The Introduction asks 'what it costs' (Sec. I), but no experiment compares this architecture against two modality-specific SIRENs with matched total parameter count, the same distillation, and the same quantization. Without such a control, the shared stack's contribution is not isolated; the separate initialization layers and output branches could be doing most of the work. Table VII strengthens this worry: after distillation Video 3 audio PSNR drops from 62.5 dB to 24.2 dB while video PSNR rises, and Sec. VI admits this trade is unresolved. Please add a matched-capacity two-network baseline (or an ablation that reallocates shared layers) and report per-modality PSNR, SSIM, and LSD.","section":"§III-B, §V-D, §VI"},{"comment":"The abstract states that the pipeline 'shrinks the representation from 9.05 MiB to 2.33 MiB, an overall compression ratio of 2.61.' But 9.05/2.33 is approximately 3.88; the value 2.61 is 6.08/2.33, i.e., the original AVI/WAV container size divided by the stored model size. The phrase 'overall compression ratio' conflates two different bases. This is a central quantitative claim and should be corrected by explicitly labeling the numerator (original container vs. teacher representation) and reporting both ratios if both are intended.","section":"Abstract; §V-F, Table VIII"},{"comment":"The 16-bit operating point is chosen from a bit-width sweep on a single video (Video 1) using five reference frames, with no error bars or repeated runs. Since 16-bit quantization is used for the whole pipeline (Sec. III-E) and the abstract claims quality 'saturates at 16 bits,' this is insufficient support for a global claim. Please provide sweeps for at least two other videos, ideally with substantial motion (Videos 3 and 5), or report the 16-bit quantization cost for all five videos and quantify run-to-run variance.","section":"§V-E, Fig. 9"},{"comment":"The conclusion says that distillation, quantization, and LZMA2 together leave reconstruction quality 'essentially unchanged.' The authors' own Table VII shows Video 3 audio PSNR dropping from 62.5 dB (teacher) to 24.2 dB (student) during distillation, a 38 dB loss, while video PSNR improves from 21.9 dB to 28.9 dB. The statement is accurate only for the quantization and encoding stages, not for the full pipeline. Rephrase to say that quantization and LZMA2 are nearly lossless, and explicitly attribute the audio/video trade to distillation.","section":"§VII vs. §V-D, Table VII"}],"minor_comments":[{"comment":"The 'Compression ratio' row is computed with different numerators across tables (source video size vs. teacher model size). Define the convention in each caption or once in Sec. III-C, and distinguish 'container compression ratio' from 'representation compression ratio.'","section":"Tables IV, VI, VII, VIII"},{"comment":"The Siamese difference is treated as the noise estimate and used to drive spectral denoising, but there is no ablation comparing the denoised audio against either a single raw branch or an oracle-noise version. Since the denoiser is part of the decode pipeline, a one-line ablation would substantiate this design choice.","section":"§III-C.3"},{"comment":"Training continues past loss flattening because perceptual metrics keep improving, yet the checkpoint kept is 'the checkpoint with the lowest loss.' Clarify whether the reported metrics come from the lowest-loss checkpoint or a checkpoint selected by perceptual metrics; the two criteria may select different epochs.","section":"§IV-B"},{"comment":"For direct comparison, include a teacher row in Table VII (or in its caption) showing teacher audio PSNR/LSD for Video 3, since the abstract's audio figure (24.18 dB) is for the student and the teacher's 62.5 dB appears only in Table V.","section":"Table VII"},{"comment":"The description of NeRV-A is brief; specify the architectural difference (frame-wise vs. sample-wise, convolutional vs. MLP-only) more concretely so the claimed gap is clear.","section":"§II, Related Work"}],"recommendation":"major_revision","confidential_remarks":"The paper is honest and unusually transparent about its limitations, which is a real strength. The main technical gap is the absence of a matched-capacity separate-network baseline; this is directly in scope for a revision and should be added before the unified-representation claim can be accepted. The compression-ratio labeling issue is a simple but important correction. No concerns about citation practice or journal fit."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: this is a solid engineering study of a genuinely new architecture combination — one plain MLP maps coordinates to both video pixels and audio samples, with separate initialization layers and shared hidden layers. It is not a compression breakthrough; conventional codecs beat it by 10–100x and there's a ~2.4 MiB floor. But it is the first sample-wise, MLP-only joint audio-video INR I know of, and the paper reports its limitations honestly.\n\nThe useful parts: the architecture is clearly described, the parameter counts line up, the quantization sweep from 1 to 32 bits gives a clean knee at 16 bits, and the comparison to H.264/HEVC/MP3 is direct and unfavorable in the right places. The Siamese audio branch for noise estimation is a nice trick, even if the denoiser threshold is blunt. The fixed-size floor is stated as a feature for long content, which is fair.\n\nThe soft spots are real. The central claim — that one shared stack can carry both modalities — is never tested against the natural control: two modality-specific SIRENs with the same total capacity, distillation, and quantization. Without that, the shared layers might just be underutilized. The paper's own distillation results show audio PSNR collapsing from 62.5 dB to 24.2 dB on Video 3 while video PSNR rises; the authors note it but then conclude quality is 'essentially unchanged,' which is not supported. That needs fixing. The bit-width sweep is on one video and five frames with no error bars — minor, but worth a line. No code or data is provided, which makes replication harder. The denoiser treats the Siamese difference as noise by definition; it works, but an external validation would be stronger.\n\nOverall, this is a demonstration, not a leap. It deserves a serious referee: the missing baseline is addressable, and the overclaim is easy to correct. I'd send it to review with a request to add the two-network control and to soften the conclusion. If you work on INR codecs, it's worth a read; if not, you can skip it.","headline":"Genuinely new INR architecture combination, but the shared-stack claim is under-tested because there is no matched-capacity separate-network baseline.","tokens_in":10433,"tokens_out":2276,"would_cite":true,"duration_ms":22572,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"This paper claims that a video and its audio track can be stored as the weights of a single sine-activated MLP, and that the resulting representation can be compressed from 9.05 MiB to 2.33 MiB with only modest quality loss.","keywords":["implicit neural representation","SIREN","video compression","audio compression","knowledge distillation","weight quantization","joint audio-video coding","coordinate MLP"],"falsifier":"Train two separate SIRENs — one audio-only, one video-only — whose combined parameter count equals the unified student's 1.30M, with the same training data, epochs, and loss; if their per-modality PSNR on Video 3 surpasses the unified student's audio PSNR of 24.18 dB by more than a small margin, the shared-representation claim is unsupported; if they do not, the unified design is justified.","tokens_in":9535,"feed_emoji":"🎬","tokens_out":6399,"duration_ms":54544,"temperature":0.7,"pith_summary":"This paper tries to show that a video and its audio track can be stored not as compressed pixel samples but as the trained weights of one plain multilayer perceptron with sine activations, mapping space-time coordinates to RGB values and audio amplitudes. On a 6.08 MiB test clip, the pipeline that overfits a teacher, distills it into a smaller student, quantizes weights to 16 bits, and losslessly encodes the result shrinks the stored representation from 9.05 MiB to 2.33 MiB (2.61x) while holding video PSNR at 28.72 dB and SSIM at 0.75. The paper further claims this is the first sample-wise, MLP-only joint audio-video representation, that quality saturates at 16-bit weights, and that the stored size is constant in the video's length and resolution — a property no conventional codec has. A sympathetic reader would care because, if true, it makes video codecs a form of model compression and opens the door to resolution-independent, differentiable representations of both modalities; the paper also reports honestly that it trails H.264/HEVC/MP3 except at their lowest-quality settings.","feed_headline":"One MLP stores video and audio, 2.61x smaller","feed_subtitle":"Stored size stops depending on clip length; quality saturates at 16-bit weights.","key_machinery":"The central object is a unified SIREN (sinusoidal representation network): a coordinate MLP with sine activations, separate initialization layers for audio and video, five shared hidden layers in the teacher (three in the student), one linear video output branch, and two linear Siamese audio output branches. Its load-bearing role is to let a single set of weights serve both modalities; the difference between the two audio branches gives the noise estimate that drives the spectral denoiser, while the separate first-layer weight ranges let the network fit audio's high frequencies without corrupting video colors.","core_discovery":"The central discovery is that one shared-hidden-layer MLP can jointly fit pixels and audio amplitudes from coordinates, provided audio and video get separate first-layer weight ranges (audio (-25,25), video (-2/3,2/3)), and that two Siamese audio branches whose outputs differ can estimate reconstruction noise for spectral denoising at decode time. The representation is then treated as a code: response-based knowledge distillation cuts the teacher from 2.37M to a 1.30M-parameter student, 16-bit symmetric quantization halves size again, and LZMA2 provides a small further lossless gain. On the paper's headline test video, the quantized student reaches 28.72 dB video PSNR / 0.75 SSIM and 24.18 d","pith_inferences":["Editorial: if the constant-size property holds beyond short clips, then a single model could encode arbitrarily long footage at fixed storage, turning codec design from bit-rate allocation into model-capacity allocation; this is testable by training on hour-long video and checking whether stored size and quality stay flat.","Editorial: the distillation results on Video 3 — audio PSNR falling from 62.5 dB to 24.2 dB while video PSNR rises — suggest a capacity-allocation trade-off between modalities inside the shared MLP; a per-modality distillation temperature or a loss schedule that protects audio could recover audio quality without hurting video, which is a direct, testable extension.","Editorial: the Siamese-branch noise-estimate trick is not specific to codecs; it could serve as a self-supervised denoiser in other INR-based generation tasks, since the branch disagreement is obtained without paired noisy/clean data.","Editorial: because the model maps continuous coordinates to values, the same compressed weights could support frame interpolation or resolution upsampling by evaluating at unobserved coordinates; the paper measures only reconstruction at training coordinates, so testing at interpolated positions would probe this."],"forward_implications":["The stored model size is independent of video length and resolution, so the codec becomes relatively more efficient as content grows; conventional bitstreams scale linearly with content.","At the aggressive end of conventional codec settings (CRF 51), the quantized student can beat H.264/HEVC on video PSNR (28.72 dB versus 22.17–24.16 dB on Video 3), while its audio lands near MP3 at 64 kbps; at higher quality settings conventional codecs win by one to two orders of magnitude in compression ratio.","Reconstruction quality saturates at 16-bit quantization: below 10 bits quality collapses and above 16 bits additional precision buys nothing visible, fixing the operating point for this pipeline.","Distillation and quantization are nearly lossless at 16 bits (SQNR about 39 dB), so model compression, not signal coding, is what determines the bitstream.","Encoding currently means training a network per video for thousands of epochs, which rules out real-time use; lowering the fixed size floor and handling motion are named as the next steps."],"fun_headline_variants":["One MLP encodes video+audio as weights, 2.61x smaller","SIREN codec: video+audio in network weights, 2.61x compression","MLP codec compresses 2.61x, but falls short of H.264","Video stored as MLP weights: 16-bit quantization saturates quality"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The load-bearing premise is that sharing hidden layers between audio and video does not degrade either modality relative to separate networks of the same total capacity — a premise the paper never isolates, since no two-network baseline is run.","fun_headline_variants_meta":{"raw":{"variants":["One MLP encodes video+audio as weights, 2.61x smaller","SIREN codec: video+audio in network weights, 2.61x compression","MLP codec compresses 2.61x, but falls short of H.264","Video stored as MLP weights: 16-bit quantization saturates quality"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001276,"raw_usage":{"total_tokens":5086,"prompt_tokens":809,"completion_tokens":4277,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":553,"completion_tokens_details":{"reasoning_tokens":4185}},"tokens_in":553,"tokens_out":4277,"duration_ms":27015,"temperature":1.0,"reasoning_tokens":4185,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-02T08:57:34.060184+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train two separate SIRENs — one audio-only, one video-only — whose combined parameter count equals the unified student's 1.30M, with the same training data, epochs, and loss; if their per-modality PSNR on Video 3 surpasses the unified student's audio PSNR of 24.18 dB by more than a small margin, the shared-representation claim is unsupported; if they do not, the unified design is justified.","supporting_citations":[],"review_version":1}