{"id":"85a4be3b-fa9f-4f75-ba1d-0334ebec4300","arxiv_id":"2607.14305","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"DCVC-MB, a neural B-frame video codec using Mamba state-space fusion, reports BD-rate savings up to 8.98% over prior neural codecs and up to 30.45% over VTM-19.0-LDP.","lead":"This paper describes a machine-learning video compressor that looks at both past and future frames to shrink video more efficiently. It reports large bitrate savings over existing neural codecs and a standard video codec, with gains up to about 30% on some test sets.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Claimed BD-rate gains are inconsistent with Table II: 8.21% over DCVC-FM (Intra=32) matches no reported variant, and the 1.81% over VTM-RA is from the open-GoP variant, not the base model.","rationale":"The most load-bearing condition for the central SOTA claim is that the reported quantitative results are accurate and consistently attributed. The paper's own Section IV and abstract contain numbers that do not match Table II when recomputed. This is an internal inconsistency, not a matter of interpretation, and it directly threatens the evidence for the headline improvement. Cross-resolution generalization is a valid concern, but the paper provides an ablation showing canonical embeddings work at 1080p; the inconsistency is a more immediate correctness risk. The reader noted this as a red flag but focused on cross-resolution as the weakest assumption, hence partial agreement. The concern can be settled by recalculating the averages; if the claimed percentages do not match any variant, the paper must be revised or its SOTA claim weakened. The reader's conditional verdict remains appropriate, so no verdict change is recommended.","tokens_in":12376,"tokens_out":17339,"duration_ms":156793,"concrete_test":"Recompute from Table II the average BD-rate differences for each DCVC-MB variant (base, og, bdc-og) against DCVC-DC and DCVC-FM for both Intra-period settings, and compare with the text's claimed 3.43%, 21.86%, 8.21%, and 8.98%. Specifically, identify which variant yields 8.21% over DCVC-FM at Intra=32; if none does, the reported quantitative claims are internally inconsistent, and the SOTA claim is not supported by the data.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central SOTA claim rests on average BD-rate reductions stated in Section IV: 3.43%/21.86% over DCVC-DC and 8.21%/8.98% over DCVC-FM for Intra-period 32/-1, plus a 1.81% gain over VTM-19.0-RA(Inter-GoP=16). Recomputing from Table II shows these numbers are not consistently tied to one model variant. The 3.43% (Intra=32) is the bdc-og variant vs DCVC-DC (-28.63 vs -25.20); 21.86% (Intra=-1) is the bdc variant (-30.45 vs -8.59); 8.98% (Intra=-1) is bdc vs DCVC-FM (-30.45 vs -21.47). However, the claimed 8.21% over DCVC-FM at Intra=32 matches no variant: base gives 7.95, og gives 9.18, bdc-og gives 9.19. The introduction instead reports 9.21% for this cell, compounding the inconsistency. Similarly, the abstract's '1.81% over VTM-RA' corresponds to DCVC-MB-og (-28.62 vs -26.82), while the base model's margin is only 0.57%. The text 'DCVC-MB is shown to be the only NVC that surpasses the stronger VTM configuration and does so by −1.81%' misattributes the open-GoP result to the base architecture. Since the base model is the architectural contribution and the inference-time add-ons are separate, failing to attribute the headline numbers to a specific variant makes the quantitative evidence for SOTA unreliable.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes DCVC-MB, a B-frame neural video codec built on the DCVC-DC framework. The main technical contributions are (i) a bidirectional Mamba-based fusion module that combines past and future reference features with linear complexity, (ii) a canonical position embedding designed to generalize across resolutions, (iii) an adaptive latent-skipping mechanism for faster entropy coding, and (iv) two inference-time strategies (open GoP and bidirectional coding). The authors report BD-rate savings over VTM-19.0 and prior neural codecs on HEVC, UVG, and MCL-JCV, and claim state-of-the-art performance. The paper includes ablation studies showing the benefit of Mamba fusion, canonical position embeddings, and the inference-time add-ons, plus runtime and memory analyses.","tokens_in":12816,"tokens_out":7848,"duration_ms":75903,"significance":"If the reported numbers are correct, the work is a relevant empirical contribution to neural B-frame coding: the Mamba-based fusion offers a practical way to use bidirectional references at high resolutions, the canonical position embedding is a simple and potentially generalizable fix for the train/test resolution gap, and the latent-skipping mechanism gives substantial entropy-coding speedups. The ablation study (Table IV) and memory-scaling experiment (Appendix Fig. 1) are valuable. However, the paper does not release code or checkpoints, and the headline quantitative claims are not consistently tied to a single model variant. The numerical inconsistencies described below must be resolved before the SOTA claim can be accepted.","major_comments":[{"comment":"The claimed average BD-rate gains are not consistently supported by Table II. The text reports '8.21% and 8.98% over DCVC-FM' for Intra-periods 32 and -1, but at Intra-period=32 no row in Table II gives 8.21%: the base model gives 7.95% (-27.39 vs -19.44), DCVC-MB-og gives 9.18%, and DCVC-MB-bdc-og gives 9.19%. The Introduction's '9.21%' is also unmatched in the table. Please correct these numbers and state precisely which configuration (base, og, bdc, bdc-og) supports each average improvement.","section":"Section IV, Table II"},{"comment":"The headline SOTA claim is misattributed to the base architecture when it actually comes from inference-time variants. The abstract's '1.81% over VTM-19.0-RA' corresponds to DCVC-MB-og (-28.62 vs -26.82); the base model's margin is only 0.57% (-27.39 vs -26.82). Similarly, the '3.43% over DCVC-DC at Intra-period=32' is from DCVC-MB-og (-28.62 vs -25.20), and the '21.86% at Intra-period=-1' is from DCVC-MB-bdc (-30.45 vs -8.59). Since the base model is the architectural contribution and the open-GoP/bidirectional-coding add-ons are separate inference-time choices, every comparison should be labeled with its exact variant.","section":"Section IV.B, Abstract; Section III.G"},{"comment":"The canonical position embedding is a load-bearing component for resolution generalization, since training is at 256×256 and testing includes 720p/1080p/4K. The ablation in Appendix Table II covers only HEVC Classes B/C/D (up to 1080p). The main evaluation also includes UVG (4K) and MCL-JCV, but no ablation is reported at those resolutions. To validate the resolution-generalization contribution at the resolutions where the headline results are claimed, please provide an ablation or analysis on UVG (or a 4K subset), or explicitly state that this mechanism's benefit at 4K is not directly measured.","section":"Appendix V, Table II; Section IV.A"}],"minor_comments":[{"comment":"The notation in Eq. (2) is unclear: the reader must infer that μ_i and σ_i are the predicted mean and scale of the latent distribution, and that 'mean(σ)' is computed over the stage being coded. Please define these quantities and specify the threshold computation more precisely.","section":"Section III.F, Eq. (2)"},{"comment":"The text says the B-frame model is 'approximately 2×' the latency of the P-frame, but the decoder latency in Table I is 492 ms vs 141 ms (≈3.5×), while the encoder is 626 vs 323 (≈1.9×). Please clarify whether the 2× figure refers to encoder only or to a different aggregate.","section":"Appendix II, Table I; Section III.F"},{"comment":"The header 'DCVC-DC p_s nop_s p_ce' is confusing. Please label the columns explicitly (e.g., 'DCVC-DC', 'no pos. emb.', 'standard 2D pos. emb.', 'canonical pos. emb.').","section":"Appendix V, Table II"},{"comment":"The percentage improvements differ between the abstract ('up to 8.98%'), the Introduction ('8.98%/9.21%'), and Section IV ('8.21% and 8.98%'). Unify these numbers after correcting the variant attribution.","section":"Abstract, Introduction, Section IV"},{"comment":"No code or checkpoints are released. Given the centrality of the SOTA claim and the difficulty of reproducing learned codecs, releasing the trained models and evaluation scripts would substantially increase the credibility of the reported numbers.","section":"General"}],"recommendation":"major_revision","confidential_remarks":"The numerical inconsistencies are fixable but must be corrected before the paper can be accepted. The technical ideas are plausible and the ablation study is informative; the main risk is that the reported SOTA margins shrink substantially when attributed to the base model alone. I would also encourage the editor to consider whether code release is required for this type of empirical SOTA claim."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"You should know two things before spending time on arXiv:2607.14305. First, the core technical claim — Mamba-based bidirectional fusion for B-frames, the canonical position embedding, and the first bidirectional GoP coding in an NVC — is genuinely new relative to the cited work and is backed by a reasonable ablation. Second, the headline BD-rate numbers in the abstract are not consistently tied to any single model variant, and one of them (8.21% over DCVC-FM at Intra=32) matches no variant in Table II at all. The stress-test note's arithmetic checks out.\n\nWhat the paper does well: the memory-scaling analysis is concrete — a transformer-based fusion model exceeds 40GB VRAM at 1080p while the Mamba pipeline fits in 13.8GB, which is the right kind of practical argument. The canonical PE ablation is convincing: standard 2D positional embeddings collapse at higher resolutions (+55.12 BD-rate on HEVC Class B), while the canonical version generalizes across Classes B–D. The adaptive latent-skipping timings (roughly 9x for P-frames, 5x for B-frames) are useful engineering data. The paper also honestly self-reports its weak spots — MCL-JCV underperformance, the animated-content domain gap, and the naive YUV conversion. That is more candid than most of this literature.\n\nThe soft spot is proportional to how central it is: the SOTA claim rests on numbers that don't reconcile. The abstract's 1.81% over VTM-RA (Inter-GoP=16) comes from the open-GoP variant; the base model's margin is 0.57%. The 8.98% over DCVC-FM comes from the bidirectional-coding variant at Intra=-1. Those are legitimate results for the full system, but they are not the base architecture's results, and Section IV's wording that DCVC-MB surpasses VTM-RA \"by -1.81%\" is wrong for the base model. The introduction's 20.19% over VTM-RA also lacks a clear anchor in the tables. The reader's worry about canonical PE generalizing to unseen resolutions is mostly answered by the appendix; the actual weak spot is dataset-level variance, which the paper admits — mixed UVG results and clearly negative MCL-JCV numbers at Intra=32. No code, no checkpoints, no error bars are par for this field, but they matter more when thin margins are the selling point.\n\nWho this is for: researchers working on learned video compression, specifically B-frame coding and SSM-based architectures. It does not reshape the subfield, but it is a solid architectural data point. I would send it to peer review and require a reconciled table naming the variant behind every headline figure, plus ideally a code release, before the SOTA claim is taken at face value.","headline":"The Mamba B-frame fusion and canonical position embedding are real, well-ablated contributions, but the abstract's headline BD-rate numbers mix base and inference-time variants and one figure matches no variant in Table II, so the SOTA claim needs a reconciled table before it stands.","tokens_in":13346,"tokens_out":10719,"would_cite":true,"duration_ms":94628,"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":"A neural video codec that codes B-frames with a Mamba state-space fusion model, beating VTM-19.0 by up to 30% BD-rate.","keywords":["neural video compression","B-frame coding","state-space models","Mamba","bidirectional fusion","entropy-aware latent skipping","canonical position embedding"],"falsifier":"Train or fine-tune the model on Vimeo90k at 256×256, then evaluate on HEVC Class B with standard pixel-coordinate positional embeddings; the paper reports a +55.12 BD-rate degradation versus DCVC-DC, while the canonical embedding gives −7.34. A direct falsification test is to evaluate DCVC-MB on sequences with aspect ratios far from training (e.g., vertical 9:16 or ultrawide 21:9) at high resolution; if BD-rate is no better than DCVC-DC, then the canonical embedding does not generalize as claimed. Alternatively, inspect the positional harmonic frequencies at test resolution; if the sampled har","tokens_in":12294,"feed_emoji":"🎥","tokens_out":5991,"duration_ms":57113,"temperature":0.7,"pith_summary":"This paper introduces DCVC-MB, a neural video compression framework that adopts an IBP B-frame structure and uses a bidirectional state-space (Mamba) fusion model to combine information from a past and a future reference frame. The central claim is that this design yields BD-rate savings of up to 30.45% over the VTM-19.0-LDP traditional codec, and is the only neural codec in the comparison to surpass the stronger VTM-19.0-RA configuration (by 1.81%). The authors attribute much of the gain to a canonical position embedding that normalizes coordinates to [0,1], enabling the model to generalize from 256×256 training patches to 720p/1080p/4K, and to an adaptive latent skipping mechanism that reduces entropy coding time by roughly 5–9× without hurting rate-distortion. A sympathetic reader would see this as evidence that B-frame coding, which traditional codecs use for substantial gains, can finally be made competitive in neural codecs at high resolutions.","feed_headline":"Neural B-frame codec beats VVC reference by 30%","feed_subtitle":"Mamba-based bidirectional fusion makes B-frames practical, cutting bitrate by up to 30% over VTM-19.0-LDP.","key_machinery":"The Bidirectional Mamba Block (BMB) is a selective state-space model with two parameter sets — one for the forward row-major scan and one for the backward scan — whose outputs are summed. Each scan discretizes the continuous ODE h'=Ah+Bx with zero-order hold, producing recurrent updates hm = Ā hm−1 + B̄ xm and output ym = C hm. The BMB fuses the feature pyramids of the past and future reference frames to condition the encoder, decoder, and entropy model. The other essential component is the canonical position embedding, which maps token positions to normalized x,y∈[0,1] before applying sinusoidal basis functions, so that the positional harmonics are resolution-independent; this prevents the","core_discovery":"DCVC-MB is a neural B-frame codec built on the DCVC-DC backbone but with two decoded picture buffers (past and future) instead of one. The core discovery is that a bidirectional Mamba-based fusion block, running forward and backward row-major scans over concatenated reference features, can effectively exploit temporal redundancy in both directions at linear O(N) complexity. The paper reports BD-rate reductions of up to 8.98% over prior neural codecs (DCVC-FM) and up to 30.45% over VTM-19.0-LDP, and it is the only neural codec in its comparison to exceed VTM-19.0-RA (Inter-GoP=16) by 1.81%. Two inference-time strategies, open GoP and bidirectional coding, provide additive savings, and the ent","pith_inferences":["If the canonical position embedding is truly resolution-agnostic, the same model should handle extreme aspect ratios (e.g., 9:16 vertical or 21:9) without retraining; this is a direct testable consequence not explored in the paper.","The adaptive latent skipping, with its per-frame and per-stage relative threshold, is a generic speedup module that could be dropped into other entropy-coded neural codecs (e.g., image compressors) to reduce coding time.","The bidirectional Mamba fusion is a generic two-stream context fusion mechanism; it might benefit other tasks requiring past/future context, such as video interpolation or super-resolution, though the paper does not examine these.","The paper's YUV420 comparison uses a naive RGB-to-YUV conversion; a model trained directly for YUV420 might close the gap to VTM-RA in that color space, but the paper's claim is limited to RGB BT.709."],"forward_implications":["If the claims hold, neural codecs can now enter the B-frame regime that traditional codecs have exploited for decades, yielding large bitrate savings at comparable quality.","The Mamba fusion's linear complexity means B-frame coding fits in 13.8GB VRAM at 1080p, a configuration where transformer-based fusion would exceed 40GB, making high-resolution neural B-frame coding practical on current GPUs.","Adaptive latent skipping reduces entropy coding time by ~9× for P-frames and ~5× for B-frames with minor RD cost, which could enable real-time or near-real-time neural coding.","The canonical position embedding allows a single checkpoint to be used across resolutions from 256×256 to 4K, eliminating the need for resolution-specific retraining (paper's results show −7.34 BD-rate on HEVC Class B vs DCVC-DC).","The two inference-time strategies, open GoP and bidirectional coding, provide additive 0.9%–1% BD-rate gains and can be toggled based on complexity needs; bidirectional coding doubles encoding time but is recommended for storage-critical scenarios."],"fun_headline_variants":["B-frame codec using Mamba cuts bitrate by 30% over VVC","DCVC-MB: Mamba-based B-frame codec beats VTM by 30%","State-space fusion makes B-frame video coding 30% cheaper than VVC","Bidirectional Mamba codec: B-frames at 30% less bitrate than VVC","Mamba-based B-frame codec outperforms VVC by up to 30%"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The canonical position embedding must preserve the positional harmonics across unseen resolutions and aspect ratios; if the [0,1] normalization does not keep the harmonics aligned at 4K or nonstandard aspect ratios, the reported BD-rate gains over VTM-19.0, which rely on cross-resolution generalization, would shrink substantially.","fun_headline_variants_meta":{"raw":{"variants":["B-frame codec using Mamba cuts bitrate by 30% over VVC","DCVC-MB: Mamba-based B-frame codec beats VTM by 30%","State-space fusion makes B-frame video coding 30% cheaper than VVC","Bidirectional Mamba codec: B-frames at 30% less bitrate than VVC","Mamba-based B-frame codec outperforms VVC by up to 30%"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000737,"raw_usage":{"total_tokens":3127,"prompt_tokens":740,"completion_tokens":2387,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":484,"completion_tokens_details":{"reasoning_tokens":2273}},"tokens_in":484,"tokens_out":2387,"duration_ms":15810,"temperature":1.0,"reasoning_tokens":2273,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-02T02:29:05.895195+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train or fine-tune the model on Vimeo90k at 256×256, then evaluate on HEVC Class B with standard pixel-coordinate positional embeddings; the paper reports a +55.12 BD-rate degradation versus DCVC-DC, while the canonical embedding gives −7.34. A direct falsification test is to evaluate DCVC-MB on sequences with aspect ratios far from training (e.g., vertical 9:16 or ultrawide 21:9) at high resolution; if BD-rate is no better than DCVC-DC, then the canonical embedding does not generalize as claimed. Alternatively, inspect the positional harmonic frequencies at test resolution; if the sampled har","supporting_citations":[],"review_version":1}