{"id":"3bffa84a-1044-4ae1-8ce9-5a570ce1fbd1","arxiv_id":"2508.04491","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"OpenDCVCs provides a unified, training-ready PyTorch implementation of DCVC, DCVC-TCM, DCVC-HEM, and DCVC-DC, with benchmarking showing OpenDCVC-DC besting the official DCVC anchor by an average BD-Rate of -59.93%.","lead":"OpenDCVCs is an open-source PyTorch library that makes four deep-learning video codecs from the DCVC series trainable and evaluable, and it benchmarks them across standard test videos. It aims to close the reproducibility gap left by the original evaluation-only code releases.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Benchmark claim is not verified against original DCVC variants; Table I anchors only to DCVC-official, and the +10.40% MCL-JCV regression for OpenDCVC is unreconciled.","rationale":"The reader's weakest assumption is essentially the same as the concern identified here: the paper does not demonstrate that its implementations faithfully match the original DCVC models and that the reported BD-Rates reflect genuine reproducibility. I find no internal logical contradiction strong enough to reject the paper outright, but the benchmarking evidence is under-specified because the only baseline is DCVC-official. A direct comparison against official checkpoints for the three variants would settle whether OpenDCVCs reproduces the methods or merely reports numbers from a different training/evaluation setup. Since the paper's value depends on the released code and documentation—which are positive but not independently verified—the appropriate verdict remains CONDITIONAL, not ACCEPT, and not REJECT.","tokens_in":6203,"tokens_out":4975,"duration_ms":65297,"concrete_test":"Obtain the official checkpoints for DCVC-TCM, DCVC-HEM, and DCVC-DC from the original repositories, evaluate them with the OpenDCVCs test script on HEVC-B, UVG, and MCL-JCV under the stated conditions (96 frames, intra period 32, low delay), and compare BD-Rate to Table I and to the original papers' tables. If the official checkpoints fail to reproduce the original published numbers under this harness, or if the OpenDCVC variants do not track those official baselines within a small tolerance, the benchmarking evidence for faithful reproduction collapses.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim requires that OpenDCVCs faithfully reproduces or improves the original DCVC series. The evidence in Table I does not establish this: every BD-Rate is relative to DCVC-official only, with no comparison against the published results of DCVC-TCM, DCVC-HEM, or DCVC-DC. If the official checkpoints for those models were evaluated in the same harness, we could tell whether OpenDCVCs is actually implementing the same methods. Section IV.B states that OpenDCVC 'slightly outperforms' official DCVC, yet Table I shows a +10.40% BD-Rate regression on MCL-JCV; the paper does not reconcile this. The training modifications in Section III—mixed quantization, scale reparameterization with a lower bound >0.1, random frame shuffling, and progressive pretraining—are plausible but are not shown to be equivalent to, or better than, the original training objectives. Without official-checkpoint baselines, the reported gains for OpenDCVC-DC (-61.56%, -65.49%, -52.74%) cannot be separated from benchmark or checkpoint-selection artifacts.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper presents OpenDCVCs, a PyTorch implementation of four learned video codecs from the DCVC series (DCVC, DCVC-TCM, DCVC-HEM, DCVC-DC), together with training modifications, a two-stage training strategy, and rate-distortion, runtime, and memory benchmarks. The central claim is that the implementation is training-ready, self-contained, and reproduces or improves the published performance of the DCVC series. The quantitative evidence is Table I, which reports BD-Rate relative to the official DCVC checkpoint. The authors report large gains for OpenDCVC-DC, moderate gains for OpenDCVC-TCM and OpenDCVC-HEM, and mixed results for OpenDCVC itself, including a +10.40% BD-Rate regression on MCL-JCV.","tokens_in":6441,"tokens_out":2118,"duration_ms":25835,"significance":"If the claims are correct, this would be a valuable community asset: it would provide the first training-ready open-source implementation of the DCVC series, with documented training recipes and unified benchmarking. The paper explicitly identifies a real gap (official releases are inference-only) and offers a plausible path to filling it. However, the significance is contingent on the implementation actually matching the original methods and on the reported numbers being reproducible. The manuscript does not yet provide sufficient evidence for either: no official-checkpoint baselines for DCVC-TCM/HEM/DC, no released model weights or hashes, and an unreconciled regression in the OpenDCVC row of Table I. The code release is a concrete contribution, but the performance claims are not yet substantiated.","major_comments":[{"comment":"The BD-Rate comparison is anchored only to DCVC-official. The central claim is faithful reproduction of the entire DCVC series, but there is no comparison against officially released checkpoints or published numerical results for DCVC-TCM, DCVC-HEM, or DCVC-DC. Because the OpenDCVC variants are trained with modified objectives and quantization, one cannot tell whether the reported gains come from reproducing the original architectures or from the specific training modifications. The authors should evaluate the official checkpoints of all four models in the same harness, or at least report the original papers' numbers, and then reconcile any differences.","section":"Table I and Section IV.B"},{"comment":"Table I reports OpenDCVC as +10.40% BD-Rate on MCL-JCV while Section IV.B states that OpenDCVC 'slightly outperforms' official DCVC. A positive BD-Rate means a bitrate increase at equal PSNR, i.e., a regression, not an improvement. This inconsistency is not discussed. Even if the improvement claim refers only to high-rate points, the aggregate BD-Rate should be acknowledged and explained. The authors should report per-dataset R-D points, clarify how the BD-Rate was computed, and discuss possible causes (training data distribution, evaluation setting, or checkpoint selection).","section":"Table I, OpenDCVC row"},{"comment":"The training modifications—mixed quantization, scale reparameterization with a lower bound of 0.1, random frame shuffling, and progressive pretraining—are plausible but are not shown to be equivalent to the original models. In particular, the scale reparameterization in Eq. (1) changes the effective entropy model and can affect rate-distortion performance. The claim that these modifications 'enable efficient, reproducible training' does not establish architectural fidelity. A needed check is to load an official checkpoint into OpenDCVC and verify that inference outputs match; without this, the reported BD-Rate gains could be artifacts of the modified objective rather than faithful reproductions.","section":"Section III.A and III.C"},{"comment":"The paper states that code is publicly available, but it does not release trained checkpoints, training logs, or hashes of the exact models used for Table I. For a reproducibility-focused paper, this is a major omission: without the exact checkpoints, an independent group cannot verify the reported BD-Rate numbers or determine whether the training procedure produces them. The authors should provide model weights for all four codecs and the official DCVC anchor, plus configuration files and random seeds.","section":"Availability and Reproducibility"}],"minor_comments":[{"comment":"The caption says 'Please zoom in for more details' and the R-D curves are hard to read. Consider splitting into subfigures or using vector graphics.","section":"Figure 1 caption"},{"comment":"The caption and table header would benefit from specifying that all BD-Rate values are relative to DCVC-official, not to each algorithm's own original release. Also clarify the inference-time measurement (e.g., per frame or per sequence).","section":"Table I caption"},{"comment":"The claim that 'previous public code releases have been limited to evaluation codes' is too strong. OpenDMC, mentioned later, includes training pipelines. The statement should be qualified to 'official repositories of the DCVC series'.","section":"Introduction"},{"comment":"The learning rate and epoch numbers are given only as prose. A table listing the four pretraining stages, their losses, durations, learning rates, and the fine-tuning stage would improve clarity and reproducibility.","section":"Section III.C"},{"comment":"There are minor grammatical issues, e.g., 'an learning rate' in Section III.C.2 and 'parameters' in the abstract. A proofread pass is recommended.","section":"General"}],"recommendation":"major_revision","confidential_remarks":"The manuscript's core weakness is the absence of a direct comparison with the official DCVC variants beyond DCVC. Since the paper is explicitly about reproducing and evaluating the DCVC series, the authors should be required to evaluate official checkpoints for all four methods and release their own trained checkpoints. The +10.40% MCL-JCV regression should be addressed head-on. This is fixable within the manuscript's scope, so I recommend major revision rather than reject."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"You should know: this is a software deliverable, not a research discovery, and as software it is genuinely useful. They ship trainable PyTorch implementations of four DCVC variants, with a described training recipe and benchmark numbers. If the code works, it lowers the barrier for a useful subfield.\n\nThe key novelty is making the DCVC series trainable; the original repos are evaluation-only, and prior open libraries cover only DVC/SSF. The training modifications in Section III—mixed quantization, softplus scale reparameterization with a lower bound, progressive pretraining with MV warm-up, multi-frame finetuning—are concrete and plausible. The benchmark table includes RD curves, runtime, memory. The code is public. That is real infrastructure credit.\n\nThe reader's concern is right. Every BD-Rate in Table I is relative to DCVC-official; there is no comparison against the published numbers or official checkpoints of DCVC-TCM, DCVC-HEM, or DCVC-DC. So the central claim that OpenDCVCs 'reproduces' those models is not actually verified. And the +10.40% MCL-JCV regression for OpenDCVC sits in the same table as the text claims 'slightly outperforms' official DCVC—unreconciled. It may be a benchmark artifact, but the paper should explain it. Also, no trained checkpoints are released with the manuscript, so a reader cannot spot-check the numbers without re-training, which is expensive. Those are fixable in a revision.\n\nProportionately: the central contribution is code, not the specific BD-Rate deltas. I would not treat the -60% numbers as established until official checkpoints of the other variants are run in the same harness. But the implementation and the described training pipeline are the deliverable, and they look honest.\n\nFor whom: graduate students and researchers who want to train DCVC-family models without reimplementing them. A serious referee can check whether the repo actually matches the described modifications and whether the comparisons are fair. I would send it to review, with the expectation of a major revision that adds official-checkpoint baselines, reconciles the MCL-JCV result, and ideally releases checkpoints.","headline":"Useful training-ready code for the DCVC family, but the reproducibility claim is only anchored to the official DCVC checkpoint; the MCL-JCV regression needs reconciling before the numbers can be trusted.","tokens_in":6930,"tokens_out":2211,"would_cite":true,"duration_ms":24591,"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 presents OpenDCVCs, an open-source PyTorch implementation that makes the DCVC series of learned video codecs — DCVC, DCVC-TCM, DCVC-HEM, and DCVC-DC — trainable end-to-end, and reports that the resulting models match or improve o","keywords":["learned video compression","open-source implementation","DCVC","rate-distortion optimization","PyTorch","progressive training","conditional coding"],"falsifier":"Run the released training code on Vimeo-90k with the stated settings and evaluate on HEVC-B, UVG, and MCL-JCV; if the resulting BD-Rate numbers do not fall within a small margin of Table I, or if training stalls without the official pretrained components, the central claim is contradicted. Alternatively, compare OpenDCVC-TCM, OpenDCVC-HEM, and OpenDCVC-DC outputs against the original authors' published rate-distortion points.","tokens_in":6060,"feed_emoji":"🎞️","tokens_out":6111,"duration_ms":58608,"temperature":0.7,"pith_summary":"OpenDCVCs is an open-source PyTorch library that reimplements four deep learned video codecs from the DCVC family — base DCVC, DCVC-TCM, DCVC-HEM, and DCVC-DC — and, crucially, provides the training code that the original releases lacked. The paper claims that from this single codebase a researcher can train all four models end-to-end and obtain rate-distortion performance comparable to or better than the original published checkpoints. Its quantitative anchor is a set of BD-Rate measurements on three standard test sets, where the top model in the family, OpenDCVC-DC, is reported to save roughly 60% of bitrate versus the official base DCVC anchor. The value, if the claim holds, is that the DCVC line of codecs becomes reproducible and extensible rather than a set of black-box evaluation scripts.","feed_headline":"OpenDCVCs trains the entire DCVC codec family from one codebase","feed_subtitle":"Training-ready PyTorch code for DCVC, DCVC-TCM, DCVC-HEM, and DCVC-DC, benchmarked against official models.","key_machinery":"The enabling machinery is the training pipeline itself: a progressive pretraining schedule that activates motion estimation, reconstruction, contextual coding, and full rate-distortion optimization in sequence, followed by multi-frame finetuning; differentiable quantization via uniform noise plus a straight-through estimator; and a softplus reparameterization that keeps entropy scale parameters bounded below. These components turn inference-only reference code into models that can be optimized end-to-end.","core_discovery":"The central claim is that a complete, self-contained training pipeline for the DCVC series is achievable and works. To make the models trainable, the authors replace hard quantization with a mixed quantization strategy, reparameterize entropy scale parameters for stable gradients, add data augmentation, and adopt a two-stage training strategy of progressive pretraining followed by multi-frame finetuning. Their benchmark results show that all four implementations operate on the same R-D frontier, with OpenDCVC-DC reporting BD-Rate reductions of 61.56% on HEVC Class B, 65.49% on UVG, and 52.74% on MCL-JCV relative to the official DCVC anchor.","pith_inferences":["The apparent regression on MCL-JCV for base OpenDCVC (+10.40% BD-Rate relative to the official anchor) suggests that training details or checkpoint selection, not architecture, drive part of the reported gains; this would be worth investigating.","If the training recipe generalizes, the same progressive-pretraining-plus-finetuning schedule could accelerate development of newer conditional-coding video codecs beyond the DCVC family.","The release lowers the barrier for comparing learned video codecs on equal footing, potentially making BD-Rate tables in future papers more trustworthy.","The large gains of OpenDCVC-DC over the base model indicate that architectural advances within the family, rather than the training recipe alone, are the main driver of compression improvement."],"forward_implications":["Researchers can train DCVC-series codecs from scratch on their own data, enabling domain-specific adaptation without waiting for official training releases.","The unified codebase makes comparisons among DCVC variants fair because training and evaluation protocols are consistent.","New modules trained on this codebase can be plugged into the family, accelerating architectural experimentation.","The reported BD-Rate numbers provide independent checkpoints against which future learned codecs can be measured.","The repository supplies infrastructure for benchmarking inference time and GPU memory use alongside rate-distortion performance."],"supporting_citations":[{"why":"Supplies the original DCVC architecture and conditional-coding framework being reimplemented.","marker":"[2]"},{"why":"Describes DCVC-TCM, the temporal context mining variant implemented here.","marker":"[3]"},{"why":"Describes DCVC-HEM, the hybrid spatial-temporal entropy model variant.","marker":"[4]"},{"why":"Describes DCVC-DC, the diverse contexts variant that anchors the strongest reported gains.","marker":"[5]"},{"why":"Provides the progressive pretraining concept that the two-stage training strategy follows.","marker":"[12]"},{"why":"Supplies the mixed quantization strategy used to keep training differentiable.","marker":"[13]"},{"why":"The softplus-based scale reparameterization that stabilizes entropy modeling.","marker":"[14]"},{"why":"The multi-frame finetuning approach used to improve temporal error propagation.","marker":"[15]"},{"why":"Vimeo-90k, the training dataset used for all models.","marker":"[16]"}],"fun_headline_variants":["OpenDCVCs: train all DCVC codecs from one PyTorch codebase","Trainable DCVC series code now open in PyTorch","OpenDCVCs: full training pipeline for DCVC family, benchmarked","DCVC codecs made trainable: OpenDCVCs shares PyTorch code"],"cache_read_input_tokens":2816,"weakest_assumption_plain":"The claim stands on the assumption that the authors' reimplementations faithfully match the architecture and training objectives of the original DCVC models, so that the reported BD-Rate differences measure the models rather than the authors' training choices.","fun_headline_variants_meta":{"raw":{"variants":["OpenDCVCs: train all DCVC codecs from one PyTorch codebase","Trainable DCVC series code now open in PyTorch","OpenDCVCs: full training pipeline for DCVC family, benchmarked","DCVC codecs made trainable: OpenDCVCs shares PyTorch code"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000228,"raw_usage":{"total_tokens":1312,"prompt_tokens":748,"completion_tokens":564,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":492,"completion_tokens_details":{"reasoning_tokens":479}},"tokens_in":492,"tokens_out":564,"duration_ms":5817,"temperature":1.0,"reasoning_tokens":479,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-05T23:54:19.834244+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the released training code on Vimeo-90k with the stated settings and evaluate on HEVC-B, UVG, and MCL-JCV; if the resulting BD-Rate numbers do not fall within a small margin of Table I, or if training stalls without the official pretrained components, the central claim is contradicted. Alternatively, compare OpenDCVC-TCM, OpenDCVC-HEM, and OpenDCVC-DC outputs against the original authors' published rate-distortion points.","supporting_citations":[{"cited_title":"Temporal context min- ing for learned video compression,","cited_arxiv_id":null,"evidence_quote":"Describes DCVC-TCM, the temporal context mining variant implemented here."},{"cited_title":"Deep contextual video compression,","cited_arxiv_id":null,"evidence_quote":"Supplies the original DCVC architecture and conditional-coding framework being reimplemented."},{"cited_title":"Hybrid spatial-temporal entropy modelling for neural video compression,","cited_arxiv_id":null,"evidence_quote":"Describes DCVC-HEM, the hybrid spatial-temporal entropy model variant."},{"cited_title":"Neural video compression with diverse contexts,","cited_arxiv_id":null,"evidence_quote":"Describes DCVC-DC, the diverse contexts variant that anchors the strongest reported gains."},{"cited_title":"M-lvc: Multiple frames prediction for learned video compression,","cited_arxiv_id":null,"evidence_quote":"Provides the progressive pretraining concept that the two-stage training strategy follows."},{"cited_title":"Channel-wise autoregressive entropy models for learned image compression,","cited_arxiv_id":null,"evidence_quote":"Supplies the mixed quantization strategy used to keep training differentiable."},{"cited_title":"Theoretical bound-guided hierarchical vae for neural image codecs,","cited_arxiv_id":null,"evidence_quote":"The softplus-based scale reparameterization that stabilizes entropy modeling."},{"cited_title":"Content adaptive and error propagation aware deep video compression,","cited_arxiv_id":null,"evidence_quote":"The multi-frame finetuning approach used to improve temporal error propagation."},{"cited_title":"Video enhance- ment with task-oriented flow,","cited_arxiv_id":null,"evidence_quote":"Vimeo-90k, the training dataset used for all models."}],"review_version":1}