{"id":"baa5b2ef-bbd3-4722-94e0-a742b698c16d","arxiv_id":"2505.14541","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":1,"one_line_summary":"DCMVC modulates the propagated temporal context with an additional oriented context from the reference frame, reporting 10.1 percent bitrate savings over DCVC-FM and 22.7 percent over VVC on standard test sets.","lead":"This paper presents DCMVC, a neural video codec that adds a second temporal context derived from the reference frame and uses it to refine the standard propagated context. On standard benchmarks it reports about 10 percent bitrate savings over the previous best neural codec DCVC-FM and about 23 percent over the traditional VVC codec.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Ablation baseline Ma is internally inconsistent with Table 1, implying the reproduced DCVC-DC is ~5.5 pp weaker than published, so module gains and the 10.1% SOTA margin need recalibration.","rationale":"The reader's weakest assumption identified the reproduced DCVC-DC baseline Ma as the key risk. My read sharpens this into a quantitative internal inconsistency: Table 3's 10.3% gain over Ma cannot be reconciled with Table 1's 4.8 percentage-point margin between DCMVC and published DCVC-DC on the same datasets and setting unless Ma is roughly 5.5 pp worse than the original codec. This is the most load-bearing concern because the ablation table is the primary evidence for attributing the reported gains to flow orientation, context compensation, and decoupling loss; a weak baseline would inflate every component's contribution. It does not directly refute the central claim that the final model beats VVC and DCVC-FM, since that claim rests on Table 2, but it does mean the scientific contribution is conditional on baseline fidelity. The reader's CONDITIONAL verdict already captures this risk, so no verdict change is needed. The proposed concrete test—evaluating official DCVC-DC weights under the same protocol—would settle the issue by exposing the absolute RD of the baseline and by checking whether Table 3 and Table 1 become consistent.","tokens_in":14268,"tokens_out":10300,"duration_ms":90769,"concrete_test":"Obtain the official DCVC-DC pretrained weights (or, if unavailable, retrain with the authors' exact recipe) and evaluate them under the paper's protocol on HEVC B/C/D/E with intra-period 32, RGB-PSNR, 96 frames, VTM-13.2 anchor. Compute the official model's BD-rate vs VTM and compare with the implied Ma score of about -11.9. If the official model lands near -17.4 as in Table 1, then Ma is a weak proxy, Table 3's 10.3% gain is not attributable to the proposed modules, and the paper must report absolute RD values and re-run ablations from the stronger baseline.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's own tables imply that the ablation baseline Ma is not a faithful DCVC-DC proxy. Section 4.3 and Table 3 report that the full model Mg gains 10.3% BD-rate over Ma on HEVC datasets with intra-period 32. Table 1, on the same datasets and setting, shows DCMVC beats the published DCVC-DC by only 4.8 percentage points on average (HEVC B/C/D/E: DCMVC -14.5, -14.4, -31.6, -28.1 vs DCVC-DC -13.9, -8.8, -27.7, -19.1). If Ma were equivalent to DCVC-DC, DCMVC should be about 10.3 pp better than Table 1's DCVC-DC; instead the implied Ma score is roughly -11.9, about 5.5 pp worse than the published -17.4. This large gap means each incremental gain in Table 3 is inflated relative to a strong baseline, and the headline 10.1% margin over DCVC-FM may partly reflect training-strategy differences rather than the proposed context modulation. Because the absolute RD of Ma is never reported, this cannot be checked from the paper. This does not disprove the final-model comparison to VVC/DCVC-FM, but it makes the contribution claim conditional on recalibration.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes DCMVC, a conditional-coding neural video codec that augments the DCVC-DC framework with two components: flow orientation, which uses SpyNet to extract an oriented flow between the reference frame and the prediction frame and generates an additional oriented temporal context; and context compensation, which modulates the propagated temporal context with the oriented context via global-local extractors, fusion networks, and a decoupling loss. Experiments on UVG, MCL-JCV, HEVC Classes B/C/D/E, and USTC-TD claim an average 22.7% bitrate reduction over VTM-13.2 and a 10.1% bitrate saving over DCVC-FM under intra-period -1, with a 19.4% average saving over VTM under intra-period 32. Ablations on HEVC report a cumulative 10.3% BD-rate improvement over a reproduced DCVC-DC baseline Ma, broken down into contributions from the two architectural components and two training-strategy components.","tokens_in":14593,"tokens_out":6629,"duration_ms":61285,"significance":"If the reported results are reproducible, DCMVC would be a new state of the art among conditional coding NVCs on the tested datasets, with a generous margin over the previous SOTA and a competitive trade-off in complexity. The paper ships a code link, compares on multiple standard datasets under two intra-period settings, includes complexity analysis, and provides visualizations of flows, contexts, and global/local features. The main weakness is the uncalibrated ablation baseline: the reproduced DCVC-DC (Ma) appears substantially weaker than the published DCVC-DC on the same datasets, which inflates the per-component gains and obscures how much of the final margin over DCVC-FM comes from the proposed architecture versus the 32-frame training strategy. The final-model comparison against external codecs is not directly invalidated, but the paper's attribution of gains to the proposed modules is conditional on resolving this inconsistency.","major_comments":[{"comment":"The ablation baseline Ma is stated in Section 4.3 to be a reproduced DCVC-DC, but the numbers in Table 1 imply otherwise. On HEVC Classes B/C/D/E with intra-period 32, DCMVC beats the published DCVC-DC by only 4.8 percentage points on average (Table 1: -14.5 vs -13.9, -14.4 vs -8.8, -31.6 vs -27.7, -28.1 vs -19.1; average difference 4.8 pp), whereas Table 3 reports that the full model Mg gains 10.3% BD-rate over Ma on the same datasets and setting. If Ma were equivalent to the published DCVC-DC, DCMVC should be about 10.3 pp better than the published numbers, not 4.8 pp. The implied Ma is roughly 5.5 pp weaker than the published DCVC-DC. Since the absolute RD performance of Ma is never reported, the per-component gains (1.9, 3.5, 4.4, 1.0, 4.3) cannot be calibrated or trusted as contributions of the proposed modules. Please report Ma's absolute BD-rate versus VTM on the HEVC datasets, reconcile the ablation with Table 1, and re-state the module contributions accordingly.","section":"§4.3, Table 3 vs Table 1"},{"comment":"The paper attributes the headline improvement to context modulation, but Table 3 shows that long-sequence training alone (Mf) accounts for 4.3 of the total 10.3 percentage point gain over Ma, i.e., about 42% of the total. Combined with the uncalibrated baseline discussed above, the reader cannot determine how much of the 10.1% margin over DCVC-FM is due to the proposed flow orientation and context compensation versus the 32-frame cascaded training procedure. The paper should separate these factors more explicitly, for example by evaluating Mf (trained with long sequences but without the proposed modules) directly against the published DCVC-FM and DCVC-DC numbers, and by reporting the absolute RD curves for Ma and Mf. This would also clarify whether the gain attributed to flow orientation and context compensation is robust when the baseline is at the published DCVC-DC level.","section":"§4.3, Table 3"}],"minor_comments":[{"comment":"In 'the prediction frame ˇxt is obtained by warping the reference frame ˆxt', the subscript should be t-1 for the reference frame; the current text is a typo.","section":"§3.2, first paragraph"},{"comment":"The text says 'Ldecouple consists of the cosine similarity of both global features and local features', but Eq. (1) defines a ratio of squared local correlation to squared global correlation. Please state explicitly that the objective minimizes this ratio and that the global correlation appears in the denominator.","section":"Eq. (1) and §3.3"},{"comment":"The notation for the compensated context is inconsistent between the figure caption (C with a bar) and the main text; please unify the symbol used for C̅_t^0 throughout.","section":"§3.3, Fig. 3"},{"comment":"The description of the 32-frame cascaded training set as 'a subset of 9000 sequences' from raw Vimeo videos should specify the selection criterion or random seed, since the composition of this subset can affect the long-sequence training results.","section":"§4.1, training data"},{"comment":"It is not stated whether the DCVC-DC and DCVC-FM results in Tables 1 and 2 are taken from the original papers or re-evaluated with the authors' own test pipeline. The note in Table 1 about quality indexes suggests some adjustment; please describe the exact protocol (e.g., which models, which weight points, whether re-encoding was performed) so the comparison is reproducible.","section":"§4.2, Tables 1-2"},{"comment":"The complexity table reports MACs and timings for 1080p inputs, but the RD evaluations use 96 frames with intra-period 32 or -1; please clarify whether the encoding/decoding times are per-frame averages and whether the reported MACs refer to the entire model including the proposed context modulation at the largest resolution.","section":"§4.4, Table 4"}],"recommendation":"major_revision","confidential_remarks":"The central empirical comparison to VTM and DCVC-FM is potentially strong, and the code release is a positive sign. The main risk is the ablation baseline: the paper's own tables imply Ma is substantially weaker than published DCVC-DC, so the contribution claim and the decomposition of gains are not yet trustworthy. The authors should be asked to provide the absolute RD performance of Ma and to reconcile the Table 3 ablation with Table 1 before the paper can be accepted; the final results may still stand after recalibration, but the current presentation is misleading. I would not reject outright, as the issue is fixable within the manuscript's scope by re-reporting the ablation and rerunning the key comparisons."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Here's the short version: this is a competent, incremental step in learned video compression, and the headline gains over VTM and DCVC-FM are probably real. But the ablation study has a hole in it: their reproduced DCVC-DC baseline (Ma) is weaker than the published model, and they never report its absolute RD numbers. That means the contribution of each proposed module is overstated.\n\nThe new pieces are flow orientation and context compensation. Flow orientation runs a SpyNet between the reference frame and the prediction frame to produce an 'oriented flow' that captures correlation the decoded motion vector misses. Context compensation then uses that oriented context to modulate the propagated context, with a decoupling loss that pushes global features together and local features apart. These are genuine modifications to the DCVC line, and they are not in DCVC-DC or DCVC-FM. The paper also ships code, which is welcome.\n\nThe main experimental comparison is direct: Table 2 shows DCMVC beating published DCVC-FM numbers on four datasets, and the margin is substantial. I see no reason to doubt the central claim that DCMVC is better than DCVC-FM under these test conditions. What is questionable is the attribution. Section 4.3 states Ma is a reproduced baseline because the DCVC-DC training code is not public. That is fine, but Ma's absolute performance is never given. If you plug the numbers from Table 1 and Table 3 together, Ma comes out roughly 5.5 percentage points worse than the published DCVC-DC on the HEVC B/C/D/E average. That is a large gap. It means the 10.3% total gain over Ma is inflated, and the split between long-sequence training and the architecture changes cannot be trusted as reported. The final comparison to DCVC-FM is less affected, because that is a direct head-to-head against published or reproduced numbers, but the contribution claim depends on knowing how much of the gain is just better training.\n\nThe rest is standard for this subfield: no error bars over seeds, no runtime analysis beyond one GPU, no ablation of the decoupling loss weight. Those are minor.\n\nWho is this for? Anyone working on conditional learned video codecs. It is a useful reference point and a solid submission. It deserves a serious referee, but the authors should be asked to report Ma's absolute RD curve and to redo the ablation against a strong, calibrated baseline before acceptance.","headline":"Solid incremental SOTA codec; ablation baseline is underpowered and distorts the module attribution.","tokens_in":15064,"tokens_out":2893,"would_cite":true,"duration_ms":26177,"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":"This paper claims that modulating propagated temporal context with a bit-free oriented context mined from the reference frame makes a neural codec, DCMVC, cut bitrate by 22.7% average versus H.266/VVC and 10.1% versus the previous best…","keywords":["neural video compression","context modulation","temporal context","flow orientation","context compensation","decoupling loss","conditional coding","low-delay video coding"],"falsifier":"Obtain the original DCVC-DC model and rerun the Section 4.3 ablations and the Table 1/2 BD-rate comparisons on HEVC Class D with intra-period 32. If the rate-distortion gap between DCMVC and the original DCVC-DC is far below the reported 10.3% ablation gain, the context-modulation improvements are partly artifacts of the baseline reproduction.","tokens_in":14086,"feed_emoji":"🎥","tokens_out":5258,"duration_ms":46014,"temperature":0.7,"pith_summary":"This paper argues that the temporal context used in conditional-coding neural video codecs is under-exploited, and proposes a two-step modulation to fix it. First, a flow-orientation module uses a learned inter-frame correlation to produce an extra 'oriented' context from the reference frame, without spending extra bits. Second, a context-compensation network merges that oriented context with the propagated context, and a decoupling loss encourages the two to be complementary rather than redundant. The resulting codec, DCMVC, is claimed to cut bitrate by 22.7% on average versus H.266/VVC and by 10.1% versus the previous leading neural codec DCVC-FM under low-delay settings. If accepted, it would establish that side-information-free context modulation is an effective route to better neural video compression.","feed_headline":"Neural video codec beats VVC by 22.7% bitrate","feed_subtitle":"Context modulation mines extra temporal correlation from the reference frame and cuts bits 10.1% below the prior neural codec.","key_machinery":"The machinery is a pair of modules inserted into the DCVC-DC conditional coding pipeline. Flow orientation: the decoded frame and reference frame are fed to a pyramid inter-frame correlation extractor (SpyNet) to obtain an 'oriented flow'; warping the reference frame with this flow yields the oriented temporal context $\\check{C}_t^0$ with no extra bit cost. Context compensation: global and local extractors (local ones using invertible affine coupling layers) pull features from the oriented and propagated contexts; the global features are added and fused, the local features are added and fused, and a shared mutual fusion produces the compensated context $\\bar{C}_t^0$. The decoupling loss $L_{\\text{decouple}} = \\frac{(\\text{Cor}(\\check{L}_t^0, L_t^0))^2}{(\\text{Cor}(\\check{G}_t^0, G_t^0))^2 + \\delta}$ supervises the extractors so that global structure correlates across the two contexts while local texture stays decorrelated, which the authors say removes irrelevant propagated information and alleviates error accumulation.","core_discovery":"The central discovery is that the propagated reference feature in conditional coding carries irrelevant accumulated information, while the adjacent reconstructed reference frame is cleaner; a codec can exploit this by generating a second temporal context from the reference frame via an oriented flow field, then using that oriented context to modulate the propagated context. The paper shows experimentally that this context modulation, together with a decoupling loss that decorrelates local details while correlating global structure, improves rate-distortion performance on six test sets under both 32-frame and infinite intra-periods, with the best model beating H.266/VVC by an average 22.7% bitrate and the previous best neural codec by 10.1%. The ablation attributes the gain to both the oriented-context path and the compensation/synergy mechanism, and per-frame plots indicate the modulation also slows quality decay in long prediction chains.","pith_inferences":["An implication the authors do not draw: the oriented-flow idea is a learned analogue of multi-hypothesis prediction in classical codecs, recovering correlation that motion-compensated prediction loses without transmitting a second motion field; a natural extension is to generalize it to multiple reference frames or higher-order temporal contexts.","The global-local decoupling loss is a generic synergy regularizer that could be transplanted to other multi-source fusion tasks in compression or image fusion, where one source is reliable in structure and another in texture.","A testable extension is to apply context compensation at all three temporal-context scales rather than only the largest scale $C_t^0$, since the paper applies it at one scale and the complexity analysis shows the gains come at substantial MAC cost (4131G); whether the method is practical depends on deployment constraints."],"forward_implications":["If the 22.7% average bitrate saving over H.266/VVC holds, DCMVC is the strongest reported conditional-coding neural codec on UVG, MCL-JCV, HEVC B–E, and USTC-TD under the tested low-delay settings.","Each proposed component is claimed to contribute: flow orientation alone saves 1.9% BD-rate, context compensation adds up to 4.4%, decoupling loss another 1.0%, and 32-frame long-sequence training 4.3%, for a combined 10.3% over the reproduced baseline.","The per-frame plots show the modulation keeps quality flatter across the prediction chain while spending fewer bits than DCVC-FM, so the method is claimed to mitigate error propagation, not just improve average rate-distortion.","Because the oriented context costs no additional side information, the bitrate savings come entirely from better conditioning of the contextual coder, making the architecture a drop-in change to existing conditional coding pipelines."],"supporting_citations":[{"why":"Provides the baseline DCVC-DC framework that DCMVC extends, including the multi-scale temporal context mining and training recipe.","marker":"[28]"},{"why":"The previous state-of-the-art DCVC-FM that DCMVC compares against and whose feature-modulation error-propagation analysis motivates the approach.","marker":"[29]"},{"why":"Establishes the conditional coding paradigm for neural video compression on which the whole framework relies.","marker":"[26]"},{"why":"Introduces the reference feature propagation structure and temporal context mining that context modulation builds upon.","marker":"[48]"},{"why":"SpyNet is used directly as the inter-frame correlation extractor inside the flow orientation module.","marker":"[47]"}],"fun_headline_variants":["Neural codec beats VVC by 22.7% bitrate","Context modulation neural codec beats VVC by 22.7%","Neural codec beats VVC by 22.7% and prior NVC by 10.1%"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The ablation results assume the authors' reproduced DCVC-DC baseline ('Ma') is a faithful stand-in for the original DCVC-DC, whose training code is not public, so the reported per-module gains could be inflated if their reproduction is weaker than the original.","fun_headline_variants_meta":{"raw":{"variants":["Neural codec beats VVC by 22.7% bitrate","Context modulation neural codec beats VVC by 22.7%","Neural codec beats VVC by 22.7% and prior NVC by 10.1%"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001256,"raw_usage":{"total_tokens":5142,"prompt_tokens":934,"completion_tokens":4208,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":550,"completion_tokens_details":{"reasoning_tokens":4144}},"tokens_in":550,"tokens_out":4208,"duration_ms":24392,"temperature":1.0,"reasoning_tokens":4144,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T15:31:43.711997+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Obtain the original DCVC-DC model and rerun the Section 4.3 ablations and the Table 1/2 BD-rate comparisons on HEVC Class D with intra-period 32. If the rate-distortion gap between DCMVC and the original DCVC-DC is far below the reported 10.3% ablation gain, the context-modulation improvements are partly artifacts of the baseline reproduction.","supporting_citations":[{"cited_title":"Neural video compression with diverse contexts","cited_arxiv_id":null,"evidence_quote":"Provides the baseline DCVC-DC framework that DCMVC extends, including the multi-scale temporal context mining and training recipe."},{"cited_title":"Neural video compression with feature modulation","cited_arxiv_id":null,"evidence_quote":"The previous state-of-the-art DCVC-FM that DCMVC compares against and whose feature-modulation error-propagation analysis motivates the approach."},{"cited_title":"Deep contextual video com- pression","cited_arxiv_id":null,"evidence_quote":"Establishes the conditional coding paradigm for neural video compression on which the whole framework relies."},{"cited_title":"Temporal context mining for learned video compression","cited_arxiv_id":null,"evidence_quote":"Introduces the reference feature propagation structure and temporal context mining that context modulation builds upon."},{"cited_title":"Optical flow estima- tion using a spatial pyramid network","cited_arxiv_id":null,"evidence_quote":"SpyNet is used directly as the inter-frame correlation extractor inside the flow orientation module."}],"review_version":1}