Pith. sign in

REVIEW 4 major objections 5 minor 19 references

OpenDCVCs: A PyTorch Open Source Implementation and Performance Evaluation of the DCVC series Video Codecs

T0 review · 4 major / 5 minor · reviewed 2026-08-05 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗ Review JSON

Pith's one-line read 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

desk verdict 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. read the letter →

arxiv 2508.04491 v1 pith:KRDJYWXO submitted 2025-08-06 eess.IV cs.CV

Yichi Zhang , Fengqing Zhu This is my paper · ORCID
classification eess.IVcs.CV
keywords learnedvideocompressionopen-sourceimplementationDCVCrate-distortionoptimizationPyTorchprogressivetrainingconditionalcoding
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

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.

What carries the argument

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.

What would settle it

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.

Watch

Extended reading notes

Core claim

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.

Load-bearing premise

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.

Editorial extensions

If this is right

  • 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.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • 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.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

4 major / 5 minor

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.

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 (4)
  1. [Table I and Section IV.B] 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.
  2. [Table I, OpenDCVC row] 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).
  3. [Section III.A and III.C] 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.
  4. [Availability and Reproducibility] 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.
minor comments (5)
  1. [Figure 1 caption] 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.
  2. [Table I caption] 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).
  3. [Introduction] 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'.
  4. [Section III.C] 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.
  5. [General] 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.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the paper's claims are benchmark measurements and code-release claims, with no fitted parameter renamed as a prediction and no load-bearing self-citation.

full rationale

This is an open-source implementation and benchmarking paper, not a derivation of predictions from fitted theory. All quantitative claims in Table I and Fig. 1 are measured rate-distortion results relative to the official DCVC checkpoint; they are benchmarks, not quantities constructed from fitted parameters. The only self-citation is reference [14] for the scale reparameterization trick 'scales = exp(softplus(scales + 2.3) − 2.3)', but this is an implementation detail borrowed from the authors' prior work and is not load-bearing for the central claim that OpenDCVCs implements and trains the DCVC series; replacing it with any standard softplus-style lower-bound reparameterization would not change the structure of the claim. No uniqueness theorem is invoked, no ansatz is smuggled via citation, and no known result is merely renamed. The discrepancy that OpenDCVC shows +10.40% BD-Rate on MCL-JCV while Section IV.B says it 'slightly outperforms' official DCVC is a consistency/reporting weakness, not circularity. The evaluation is anchored to an official external checkpoint and standard public datasets, so the results are self-contained as measurements. Score 0.

Assumptions & free parameters 2 free parameters · 3 assumptions · 0 invented entities

The paper introduces no new theoretical entities. The main assumptions are practical: faithful reimplementation of cited architectures, standard training data, and the chosen anchor. The free parameters are all standard training hyperparameters, which are not fitted to the test data, so their impact on the central claim is limited.

free parameters (2)
  • Reduction factor for ReduceLROnPlateau = 0.5, patience 3
    Hyperparameter choices for the learning rate scheduler, stated in Section III-C1. They are chosen by hand and affect convergence but do not materially change the central claim.
  • Learning rates and epochs for training stages = 1e-4 per stage, 30 epochs; 4e-5 for 10 epochs in finetuning
    Stated in Section III-C. Specific choices are required to reproduce the reported results.
assumptions (3)
  • domain assumption The DCVC-series architectures as described in the cited papers [2], [3], [4], [5] are correctly reproduced by OpenDCVCs.
    The entire benchmark effort assumes the reimplementation matches the original models. This is assumed in Section III and not verified against the original papers' own results except for the DCVC anchor.
  • domain assumption Training on Vimeo-90k septuplets with 256x256 random crops generalizes to the test datasets HEVC-B, UVG, MCL-JCV.
    Standard practice in learned compression, stated in Section IV-A. The paper relies on it without showing cross-dataset training robustness.
  • domain assumption BD-Rate computed relative to the authors' own DCVC checkpoint is representative.
    Used in Table I. The anchor is the authors' run, not the official released checkpoints, making the listed percentages dependent on that run's quality.

how reviews work

0 comments
Cite this review

Pith. "Pith review of OpenDCVCs: A PyTorch Open Source Implementation and Performance Evaluation of the DCVC series Video Codecs." pith.science (2026). https://pith.science/paper/KRDJYWXO

@misc{pith2026250804491,
  author       = {Pith},
  title        = {Pith review of: OpenDCVCs: A PyTorch Open Source Implementation and Performance Evaluation of the DCVC series Video Codecs},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/KRDJYWXO}},
  note         = {Machine review of arXiv:2508.04491}
}
read the original abstract

We present OpenDCVCs, an open-source PyTorch implementation designed to advance reproducible research in learned video compression. OpenDCVCs provides unified and training-ready implementations of four representative Deep Contextual Video Compression (DCVC) models--DCVC, DCVC with Temporal Context Modeling (DCVC-TCM), DCVC with Hybrid Entropy Modeling (DCVC-HEM), and DCVC with Diverse Contexts (DCVC-DC). While the DCVC series achieves substantial bitrate reductions over both classical codecs and advanced learned models, previous public code releases have been limited to evaluation codes, presenting significant barriers to reproducibility, benchmarking, and further development. OpenDCVCs bridges this gap by offering a comprehensive, self-contained framework that supports both end-to-end training and evaluation for all included algorithms. The implementation includes detailed documentation, evaluation protocols, and extensive benchmarking results across diverse datasets, providing a transparent and consistent foundation for comparison and extension. All code and experimental tools are publicly available at https://gitlab.com/viper-purdue/opendcvcs, empowering the community to accelerate research and foster collaboration.

Figures

Figures reproduced from arXiv: 2508.04491 by the authors.

Figure 1
Figure 1. R-D curves of various methods. Please zoom in for more details. TABLE I: Comparison of various methods Method BD-Rate (%) Model Parameters Inference Time GPU Memory HEVC-B UVG MCL-JCV Average Occupancy DCVC-official 0% 0% 0% 0% 7.94 M 0.2615 s 21.79 GB OpenDCVC -10.60% -6.35% 10.40% -2.18% 7.94 M 0.2620 s 21.80 GB OpenDCVC-TCM -42.35% -46.70% -27.11% -38.72% 10.70 M 0.3070 s 5.67 GB OpenDCVC-HEM -56.39% -59.75% -46.… view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

19 extracted references · 19 canonical work pages

  1. [1]

    Dvc: An end-to-end deep video compression framework,

    G. Lu, W. Ouyang, D. Xu, X. Zhang, C. Cai, and Z. Gao, “Dvc: An end-to-end deep video compression framework,” in Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , 2019, pp. 11 006–11 015

    work page 2019

  2. [2]

    Deep contextual video compression,

    J. Li, B. Li, and Y . Lu, “Deep contextual video compression,” Advances in Neural Information Processing Systems , vol. 34, pp. 18 114–18 125, 2021

    work page 2021

  3. [3]

    Temporal context min- ing for learned video compression,

    X. Sheng, J. Li, B. Li, L. Li, D. Liu, and Y . Lu, “Temporal context min- ing for learned video compression,” IEEE Transactions on Multimedia , vol. 25, pp. 7311–7322, 2022

    work page 2022

  4. [4]

    Hybrid spatial-temporal entropy modelling for neural video compression,

    J. Li, B. Li, and Y . Lu, “Hybrid spatial-temporal entropy modelling for neural video compression,” Proceedings of the ACM International Conference on Multimedia , pp. 1503–1511, 2022

    work page 2022

  5. [5]

    Neural video compression with diverse contexts,

    ——, “Neural video compression with diverse contexts,” Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pp. 22 616–22 626, 2023

    work page 2023

  6. [6]

    Neural video compression with feature modulation,

    ——, “Neural video compression with feature modulation,” Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recogni- tion, pp. 26 099–26 108, 2024

    work page 2024

  7. [7]

    Scale-space flow for end-to-end optimized video compres- sion,

    E. Agustsson, D. Minnen, N. Johnston, J. Balle, S. J. Hwang, and G. Toderici, “Scale-space flow for end-to-end optimized video compres- sion,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , 2020, pp. 8503–8512

    work page 2020

  8. [8]

    OpenDVC: An Open Source Implementation of the DVC Video Compression Method

    R. Yang, L. Van Gool, and R. Timofte, “Opendvc: An open source implementation of the dvc video compression method,” arXiv preprint arXiv:2006.15862, 2020

    work page Pith review arXiv 2006

Show all 19 references
  1. [9]

    Fvc: A new framework towards deep video compression in feature space,

    Z. Hu, G. Lu, and D. Xu, “Fvc: A new framework towards deep video compression in feature space,” Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pp. 1502–1511, 2021

    2021

  2. [10]

    Opendmc: An open-source library and performance evaluation for deep-learning- based multi-frame compression,

    W. Gao, S. Sun, H. Zheng, Y . Wu, H. Ye, and Y . Zhang, “Opendmc: An open-source library and performance evaluation for deep-learning- based multi-frame compression,” Proceedings of the ACM International Conference on Multimedia , pp. 9685–9688, 2023

    2023

  3. [11]

    Dvc-p: Deep video compression with perceptual optimizations,

    S. Zhang, M. Mrak, L. Herranz, M. G. Blanch, S. Wan, and F. Yang, “Dvc-p: Deep video compression with perceptual optimizations,” Pro- ceedings of the International Conference on Visual Communications and Image Processing, pp. 1–5, 2021

    2021

  4. [12]

    M-lvc: Multiple frames prediction for learned video compression,

    J. Lin, D. Liu, H. Li, and F. Wu, “M-lvc: Multiple frames prediction for learned video compression,” in Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , 2020, pp. 3546–3554

    2020

  5. [13]

    Channel-wise autoregressive entropy models for learned image compression,

    D. Minnen and S. Singh, “Channel-wise autoregressive entropy models for learned image compression,” Proceedings of the IEEE International Conference on Image Processing , pp. 3339–3343, 2020

    2020

  6. [14]

    Theoretical bound-guided hierarchical vae for neural image codecs,

    Y . Zhang, Z. Duan, Y . Huang, and F. Zhu, “Theoretical bound-guided hierarchical vae for neural image codecs,” Proceedings of the IEEE International Conference on Multimedia and Expo , 2024

    2024

  7. [15]

    Content adaptive and error propagation aware deep video compression,

    G. Lu, C. Cai, X. Zhang, L. Chen, W. Ouyang, D. Xu, and Z. Gao, “Content adaptive and error propagation aware deep video compression,” Proceedings of the European Conference on Computer Vision , pp. 456– 472, 2020

    2020

  8. [16]

    Video enhance- ment with task-oriented flow,

    T. Xue, B. Chen, J. Wu, D. Wei, and W. T. Freeman, “Video enhance- ment with task-oriented flow,” International Journal of Computer Vision, vol. 127, no. 8, pp. 1106–1125, 2019

    2019

  9. [17]

    JVET-J1010: JVET Common Test Conditions and Software Reference Configurations,

    J. M. Boyce, K. Suehring, X. Li, and V . Seregin, “JVET-J1010: JVET Common Test Conditions and Software Reference Configurations,” Joint Video Experts Team (JVET) of ITU-T SG 16 WP 3 and ISO/IEC JTC 1/SC 29/WG 11, San Diego, US, JVET Document JVET-J1010, July 2018, 10th Meetin...

    2018

  10. [18]

    Ultra video group test sequences,

    “Ultra video group test sequences,” http://ultravideo.cs.tut.fi, 2021, on- line; accessed 12 April 2021

    2021

  11. [19]

    Mcl-jcv: a jnd-based h. 264/avc video quality assessment dataset,

    H. Wang, W. Gan, S. Hu, J. Y . Lin, L. Jin, L. Song, P. Wang, I. Katsavounidis, A. Aaron, and C.-C. J. Kuo, “Mcl-jcv: a jnd-based h. 264/avc video quality assessment dataset,” Proceedings of the IEEE International Conference on Image Processing , pp. 1509–1513, 2016

    2016

Pith tools

Reviewed August 5, 2026 · model on record in the stance chip above.