Pith. sign in

REVIEW 3 major objections 6 minor 48 references

EHVC: Efficient Hierarchical Reference and Quality Structure for Neural Video Coding

T0 review · 3 major / 6 minor · reviewed 2026-08-05 · deepseek-v4-flash

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

Pith's one-line read EHVC aligns reference and quality structure in a neural video codec and reports average bitrate savings of about 17% over VTM-23.4 under low-delay settings.

desk verdict Solid engineering gains with a sensible reference-quality alignment idea, but the key ablation doesn't isolate the mechanism and the lookahead module has a real unexplained gap. read the letter →

arxiv 2509.04118 v1 pith:KENT6RJU submitted 2025-09-04 eess.IV cs.AI

Junqi Liao , Yaojun Wu , Chaoyi Lin , Zhipin Deng , Li Li , Dong Liu , Xiaoyan Sun This is my paper · ORCID
classification eess.IVcs.AI
keywords neuralvideocodinghierarchicalqualitystructurereferenceconditionallookaheadlayer-wisequantizationscalelow-delayrate-distortion
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

The paper argues that previous neural video codecs suffer from a reference-quality mismatch: their implicit reference structure is not aligned with their explicit hierarchical quality structure, which makes long-sequence coding unstable. EHVC fixes this by borrowing the hierarchical design of traditional codecs: each frame references both its immediate neighbor and the nearest high-quality key frame. On top of that, it adds an encoder-side lookahead from the next frame and a layer-wise quantization scale trained with random quality perturbation. In tests on standard video datasets, EHVC reports average bitrate savings of 17.14% (intra-period 32) and 17.04% (intra-period -1) over VTM-23.4, outperforming the previous state-of-the-art neural codec DCVC-FM by roughly 11 to 13 percentage points. If correct, this makes EHVC the new state of the art for low-delay neural video coding.

What carries the argument

The hierarchical multi-reference context generation module, which takes the decoder-side feature of the previous frame and the decoder-side feature of the nearest high-quality key frame, warps the key-frame feature n times (where n is the frame distance) to align it with the current frame, and fuses both branches across three scales to produce temporal contexts. Supporting it are a lookahead module that merges the next frame into the encoder-side context, and layer-wise learnable quantization scales (one per hierarchical layer) with a random quality training scheme that randomly scales the first key frame's quantization by a factor in [0.8, 1.2].

What would settle it

Evaluate EHVC under intra-period -1 on a sequence where every key frame is replaced by a no-information frame (all pixels set to 128). If the bitrate advantage over DCVC-FM largely disappears, the gain attributed to the key-frame reference would be falsified. Alternatively, measure the alignment error between an n-times-warped key-frame feature and a directly warped feature using optical flow from the key frame to the current frame for n=2,3,4; a sharp increase with n would indicate the long-range warping assumption breaks.

Watch

Extended reading notes

Core claim

EHVC claims that the reference-quality mismatch is the main cause of error propagation and instability in neural video codecs, and that explicitly aligning the reference structure with the quality structure recovers both stability and efficiency. The paper's central evidence is a hierarchical multi-reference context generation module that fuses two branches: the adjacent frame branch, which processes the previous frame's feature, and the key frame branch, which warps the nearest high-quality key frame's feature n times to the current frame. Ablation shows hierarchical multi-reference alone gives 10.47% BD-rate savings; adding lookahead gives 12.74%; layer-wise quality scale gives 13.09%; and

Load-bearing premise

The key-frame branch assumes that repeatedly warping a key frame's feature up to four times keeps enough useful information to improve the current frame's coding; if iterative warping drifts, the 10.47% bitrate saving from hierarchical multi-reference would shrink.

Editorial extensions

If this is right

  • EHVC reduces average bitrate by 17.14% over VTM-23.4 on 96 frames with intra-period 32, and by 17.04% when all frames are inter-coded (intra-period -1).
  • Compared with the previous state-of-the-art DCVC-FM, EHVC saves 10.98% more bitrate under intra-period 32 and 12.88% more under intra-period -1, measured relative to VTM-23.4.
  • The quality structure of EHVC is more stable per 32-frame refresh cycle: about 1 dB quality degradation versus about 1.5 dB for DCVC-FM, and EHVC's frame-level PSNR matches the designed hierarchical pattern more closely.
  • After a deliberately corrupted no-information frame, EHVC's reconstructed quality recovers to its pre-corruption level within a few frames, whereas DCVC-FM stays degraded, showing the explicit reference structure is more robust.
  • The computational cost is 1311 kMACs/pixel, only about 17% higher than DCVC-FM's 1125 kMACs/pixel, while delivering more than 10% bitrate savings, so the added complexity is modest relative to the gain.

Reading between the lines

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

  • The hierarchical multi-reference design could naturally extend to random-access (B-frame) coding by treating future anchor frames as key frames; the paper only evaluates low-delay configurations, so this is an untested extension.
  • The one-frame lookahead suggests a latency-quality trade-off: using more future frames could improve the quality structure further, but would increase delay; measuring this curve is a concrete next step.
  • The key-frame branch's n-times warping likely degrades with larger GOP sizes; beyond the tested 4-frame key spacing, a periodic refresh of the key frame or a distance-based weight might be needed to sustain the 10.47% gain.
  • Random quality training acts as a form of quantization-scale augmentation, which may transfer to other conditional neural codecs as a generic robustness technique even without the hierarchical reference structure.
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

3 major / 6 minor

Summary. The paper proposes EHVC, a conditional neural video codec built on DCVC-FM, with three main contributions: (1) a hierarchical multi-reference scheme that adds a key-frame reference branch to align the reference structure with a hierarchical quality structure; (2) a lookahead strategy that uses one future frame to obtain an encoder-side context; and (3) a layer-wise quantization scale with random quality training to stabilize the quality structure. Experiments report average BD-rate savings of 17.14% (intra-period 32) and 17.04% (intra-period -1) over VTM-23.4 LDB, and 10.98% and 12.88% over DCVC-FM, respectively. The paper also analyzes frame-level quality and reference stability on corrupted frames.

Significance. If the reported results are reproducible, EHVC would be a new state-of-the-art low-delay neural video codec. The performance gains are large, consistently in the double digits, and the complexity increase over DCVC-FM is modest. The paper gives a useful analysis of quality and reference structures, and provides ablations for each proposed component. However, the central ablation for the hierarchical multi-reference does not isolate the claimed 'reference-quality alignment' property, and the lookahead scheme leaves an encoder/decoder context mismatch under-specified. These issues limit the current support for the paper's conceptual claims, though the empirical SOTA claim may still hold.

major comments (3)
  1. [Section 4.3, Fig. 4] The lookahead produces an encoder-side context c^enc_t from the future frame x_{t+1}, while the decoder uses c^dec_t. The paper does not state whether the entropy model for y_t conditions on c^enc_t. If it does, the decoder cannot compute probability estimates without the future frame, breaking decodability. If it does not, then the encoder's transform is conditioned on information unavailable to the decoder, and the paper should explain how this asymmetry is trained and why it does not cause a train/test mismatch. This is load-bearing for the lookahead contribution and must be clarified.
  2. [Section 5.4, Table 6] The M_b ablation adds both a second reference branch and the key-frame selection relative to the DCVC-FM baseline M_a. The 10.47% BD-rate gain is attributed to aligning reference and quality structures, but there is no control for (a) adding a second reference branch using a non-key past frame at a similar distance, (b) adding a branch with equal capacity, or (c) using the key-frame feature without repeated warping. The gain could therefore be due to multi-reference or increased model capacity rather than the quality-alignment property. Please add a controlled ablation to support the paper's central conceptual claim.
  3. [Section 4.2, Fig. 5] The key-frame branch warps the key-frame feature n times, where n is the distance to the current frame. The paper does not analyze accumulated warp error or compare against a non-key reference at the same distance. If repeated warping degrades the key-frame feature, the improvement cannot be attributed to the quality-alignment mechanism. At minimum, report a quantitative or qualitative evaluation of the warped key-frame feature, or an ablation with a single-step warp, to demonstrate that the key-frame information remains informative.
minor comments (6)
  1. [Abstract, Section 1, Table 4] The abstract and introduction state that EHVC outperforms VTM under intra-period -1, but Table 4 shows HEVC E is +2.32% BD-rate (worse than VTM). Please qualify these claims as average or per-dataset.
  2. [Tables 3, 4, 6] No error bars or multiple-seed results are reported. Since neural codec training has stochastic variation, please state whether these are single-run results and, if feasible, report variance.
  3. [Section 5.1] The qualitative analysis in Section 3 uses VTM-20.0 while the BD-rate tables use VTM-23.4. Please clarify whether this difference affects the conclusions.
  4. [Eq. (3), Section 4.4] The random quality training scales Q_enc by omega and Q_dec by 1/omega. Please explain the rationale for the inverse scaling and specify whether omega is sampled per frame or per GOP, and whether the range 0.8–1.2 is fixed or tuned.
  5. [General] The manuscript contains broken rendering artifacts (e.g., 'uni00000013' sequences in Section 2.2.1). Please fix the text encoding.
  6. [Section 1, Footnote 3 (reproducibility)] The code is promised to be released but is not yet available. Providing code at review time would strengthen reproducibility.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: EHVC is an empirical architecture study whose claims rest on measured rate-distortion comparisons and ablations, not on fitted parameters presented as predictions or self-citation chains.

full rationale

The paper's central results are direct rate-distortion measurements on standard test sets, and its ablations compare incremental architectural additions against a DCVC-FM baseline. No equation defines an output in terms of its target; no fitted constant is renamed as a prediction; no uniqueness theorem is imported from the authors' prior work; and no ansatz is smuggled in via citation. The only self-citation ([21], EFVC) appears in a general list of modern NVCs and is not load-bearing. The ablation for the hierarchical multi-reference scheme (M_b) does not fully isolate the quality-alignment mechanism from added capacity, but that is a limitation in experimental control, not circularity: the measured 10.47% gain is not forced by construction, and the 'reference-quality mismatch' narrative is an interpretation of measured gains rather than an input to them. The design choice to define key frames by hierarchical weight is explicit and based on prior external work (DCVC-DC), not on the target result. Therefore, the derivation chain is self-contained in the relevant sense, and the appropriate circularity score is 0.

Assumptions & free parameters 3 free parameters · 5 assumptions · 0 invented entities

The paper contributes an empirical architecture; its central claim depends on the training/evaluation protocol and a few architectural assumptions. Free parameters are mostly inherited or learned; no invented entities.

free parameters (3)
  • Random quality scale range (omega) = [0.8, 1.2]
    Chosen by hand in Eq. (3) to perturb the first key frame's quantization scale during training; affects robustness of the quality structure.
  • Hierarchical lambda weights = [0.5, 1.2, 0.5, 0.9]
    Inherited from DCVC-DC and used to define three quality layers and key frames; the central claim depends on this choice.
  • Layer-wise quantization scales Q_enc_layer and Q_dec_layer = learned
    Trainable per-layer scaling of the quantizer step (Section 4.4); their values are learned on training data and affect the reported gains.
assumptions (5)
  • domain assumption The conditional coding framework and rate-distortion training objective of DCVC-FM remain valid when encoder-side and decoder-side contexts differ.
    Section 4.3 uses c_enc at the encoder and c_dec at the decoder without proof that the latent y remains decodable; the lookahead gain depends on this.
  • domain assumption Frames assigned hierarchical weight 1.2 are reconstructed at sufficiently higher quality to act as reliable key references.
    Defined in Section 4.2; if key frames do not maintain higher quality, the key-frame branch provides no benefit.
  • domain assumption Iterative n-times feature warping from the key frame to the current frame remains accurate enough over a GOP of 4 frames.
    Section 4.2 and Fig. 5; the 10.47% gain of M_b in the ablation depends on this.
  • domain assumption The no-information-frame experiment is a valid probe of reference structure stability.
    Section 3.2; the paper generalizes from this single synthetic test to the claim that implicit NVC references are fragile.
  • domain assumption Baseline DCVC-FM is reproduced faithfully for the ablation.
    Table 6 ablation uses DCVC-FM as M_a; any mismatch in implementation would bias all incremental gains.

how reviews work

0 comments
Cite this review

Pith. "Pith review of EHVC: Efficient Hierarchical Reference and Quality Structure for Neural Video Coding." pith.science (2026). https://pith.science/paper/KENT6RJU

@misc{pith2026250904118,
  author       = {Pith},
  title        = {Pith review of: EHVC: Efficient Hierarchical Reference and Quality Structure for Neural Video Coding},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/KENT6RJU}},
  note         = {Machine review of arXiv:2509.04118}
}
read the original abstract

Neural video codecs (NVCs), leveraging the power of end-to-end learning, have demonstrated remarkable coding efficiency improvements over traditional video codecs. Recent research has begun to pay attention to the quality structures in NVCs, optimizing them by introducing explicit hierarchical designs. However, less attention has been paid to the reference structure design, which fundamentally should be aligned with the hierarchical quality structure. In addition, there is still significant room for further optimization of the hierarchical quality structure. To address these challenges in NVCs, we propose EHVC, an efficient hierarchical neural video codec featuring three key innovations: (1) a hierarchical multi-reference scheme that draws on traditional video codec design to align reference and quality structures, thereby addressing the reference-quality mismatch; (2) a lookahead strategy to utilize an encoder-side context from future frames to enhance the quality structure; (3) a layer-wise quality scale with random quality training strategy to stabilize quality structures during inference. With these improvements, EHVC achieves significantly superior performance to the state-of-the-art NVCs. Code will be released in: https://github.com/bytedance/NEVC.

Figures

Figures reproduced from arXiv: 2509.04118 by the authors.

Figure 1
Figure 1. Rate-Distortion curves and BD-rate [4] comparisons with other video codecs. The test dataset is HEVC B with intra￾period -1. EHVC surpasses VTM and SOTA NVC. While significant research efforts have been devoted to optimiz￾ing neural network architectures and coding modules in NVCs, the top-level design of NVC such as hierarchical structure (including quality structure and reference structure) remains largely unex￾pl… view at source ↗
Figure 2
Figure 2. Frame-level PSNR on the first 300 frames of [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. All codecs use intra-period -1 (only the first frame uses intra [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figures from the paper (5 more)
Figure 4
Figure 4. Figure 4: The framework of our EHVC. The motion coding branch follows the same structure as DCVC-FM and is therefore [PITH_FULL_IMAGE:figures/full_fig_p005_4.png]
Figure 5
Figure 5. Figure 5: The context generation process for (a) the intra-coding frame and (b) other frames in our EHVC. [PITH_FULL_IMAGE:figures/full_fig_p006_5.png]
Figure 6
Figure 6. Figure 6: An example of the designed hierarchical reference [PITH_FULL_IMAGE:figures/full_fig_p006_6.png]
Figure 7
Figure 7. Figure 7: Rate-Distortion curves on HEVC B, HEVC C, UVG, and MCL-JCV dataset. The intra-period is 32 with 96 frames. [PITH_FULL_IMAGE:figures/full_fig_p007_7.png]
Figure 8
Figure 8. Figure 8: Rate-Distortion curves on HEVC B, HEVC C, UVG, and MCL-JCV dataset. The intra-period is -1 with all frames. [PITH_FULL_IMAGE:figures/full_fig_p008_8.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

48 extracted references · 46 canonical work pages

  1. [1]

    Eirikur Agustsson, David Minnen, Nick Johnston, Johannes Balle, Sung Jin Hwang, and George Toderici. 2020. Scale-space flow for end-to-end optimized video compression. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 8503–8512

    work page 2020

  2. [2]

    David Alexandre, Hsueh-Ming Hang, and Wen-Hsiao Peng. 2023. Hierarchical B- frame video coding using two-layer CANF without motion coding. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition . 10249– 10258

    work page 2023

  3. [3]

    Yifan Bian, Chuanbo Tang, Li Li, and Dong Liu. 2025. Augmented Deep Contexts for Spatially Embedded Video Coding. In Proceedings of the Computer Vision and Pattern Recognition Conference. 2094–2104

    work page 2025

  4. [4]

    Gisle Bjontegaard. 2001. Calculation of average PSNR differences between RD- curves. ITU SG16 Doc. VCEG-M33 (2001)

    work page 2001

  5. [5]

    Frank Bossen, Karsten Sühring, Adam Wieckowski, and Shan Liu. 2021. VVC complexity and software implementation analysis. IEEE Transactions on Circuits and Systems for Video Technology 31, 10 (2021), 3765–3778

    work page 2021

  6. [6]

    Benjamin Bross, Jianle Chen, Jens-Rainer Ohm, Gary J Sullivan, and Ye-Kui Wang. 2021. Developments in international video coding standardization after AVC, with an overview of versatile video coding (VVC). Proc. IEEE 109, 9 (2021), 1463–1493

    work page 2021

  7. [7]

    Benjamin Bross, Ye-Kui Wang, Yan Ye, Shan Liu, Jianle Chen, Gary J Sullivan, and Jens-Rainer Ohm. 2021. Overview of the versatile video coding (VVC) standard and its applications. IEEE Transactions on Circuits and Systems for Video Technology 31, 10 (2021), 3736–3764

    work page 2021

  8. [8]

    Zhenghao Chen, Lucas Relic, Roberto Azevedo, Yang Zhang, Markus Gross, Dong Xu, Luping Zhou, and Christopher Schroers. 2023. Neural video compression with spatio-temporal cross-covariance transformers. In Proceedings of the 31st ACM International Conference on Multimedia . 8543–8551

    work page 2023

Show all 48 references
  1. [9]

    Abdelaziz Djelouah, Joaquim Campos, Simone Schaub-Meyer, and Christopher Schroers. 2019. Neural inter-frame compression for video coding. In Proceedings of the IEEE/CVF international conference on computer vision . 6421–6429

    2019

  2. [10]

    Bernd Girod, Eckehard G Steinbach, and Niko Faerber. 1995. Comparison of the H. 263 and H. 261 video compression standards. InStandards and Common Interfaces for Video Information Systems: A Critical Review , Vol. 10282. SPIE, 230–248

    1995

  3. [11]

    Haifeng Guo, Sam Kwong, Dongjie Ye, and Shiqi Wang. 2023. Enhanced con- text mining and filtering for learned video compression. IEEE Transactions on Multimedia 26 (2023), 3814–3826

    2023

  4. [12]

    Amirhossein Habibian, Ties van Rozendaal, Jakub M Tomczak, and Taco S Cohen

  5. [13]

    Yung-Han Ho, Chih-Peng Chang, Peng-Yu Chen, Alessandro Gnutti, and Wen- Hsiao Peng. 2022. Canf-vc: Conditional augmented normalizing flows for video compression. In European Conference on Computer Vision . Springer, 207–223

    2022

  6. [14]

    Zhihao Hu, Zhenghao Chen, Dong Xu, Guo Lu, Wanli Ouyang, and Shuhang Gu

  7. [15]

    Zhihao Hu, Guo Lu, Jinyang Guo, Shan Liu, Wei Jiang, and Dong Xu. 2022. Coarse- to-fine deep video coding with hyperprior-guided mode prediction. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition . 5921– 5930

    2022

  8. [16]

    Théo Ladune, Pierrick Philippe, Wassim Hamidouche, Lu Zhang, and Olivier Déforges. 2021. Conditional coding for flexible learned video compression. arXiv preprint arXiv:2104.07930 (2021)

    2021 arXiv

  9. [17]

    Jiahao Li, Bin Li, and Yan Lu. 2021. Deep contextual video compression.Advances in Neural Information Processing Systems 34 (2021), 18114–18125

    2021

  10. [18]

    Jiahao Li, Bin Li, and Yan Lu. 2022. Hybrid spatial-temporal entropy modelling for neural video compression. InProceedings of the 30th ACM International Conference on Multimedia. 1503–1511

    2022

  11. [19]

    Jiahao Li, Bin Li, and Yan Lu. 2023. Neural video compression with diverse contexts. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. 22616–22626

    2023

  12. [20]

    Jiahao Li, Bin Li, and Yan Lu. 2024. Neural video compression with feature modulation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 26099–26108

    2024

  13. [21]

    Junqi Liao, Li Li, Dong Liu, and Houqiang Li. 2025. EFVC: Error-Propagation- Free Neural Video Coding with Reversible Transform. In 2025 IEEE International Symposium on Circuits and Systems (ISCAS) . IEEE, 1–5

    2025

  14. [22]

    Jianping Lin, Dong Liu, Houqiang Li, and Feng Wu. 2020. M-LVC: Multiple frames prediction for learned video compression. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition . 3546–3554

    2020

  15. [23]

    Haojie Liu, Ming Lu, Zhan Ma, Fan Wang, Zhihuang Xie, Xun Cao, and Yao Wang. 2020. Neural video coding using multiscale motion compensation and spatiotemporal context model. IEEE Transactions on Circuits and Systems for Video Technology 31, 8 (2020), 3182–3196

    2020

  16. [24]

    Jerry Liu, Shenlong Wang, Wei-Chiu Ma, Meet Shah, Rui Hu, Pranaab Dhawan, and Raquel Urtasun. 2020. Conditional entropy coding for efficient video com- pression. In European Conference on Computer Vision . Springer, 453–468

    2020

  17. [25]

    Salvator Lombardo, Jun Han, Christopher Schroers, and Stephan Mandt. 2019. Deep generative video compression. Advances in Neural Information Processing Systems 32 (2019)

    2019

  18. [26]

    Guo Lu, Wanli Ouyang, Dong Xu, Xiaoyun Zhang, Chunlei Cai, and Zhiyong Gao

  19. [27]

    Guo Lu, Xiaoyun Zhang, Wanli Ouyang, Li Chen, Zhiyong Gao, and Dong Xu

  20. [28]

    Fabian Mentzer, George Toderici, David Minnen, Sung-Jin Hwang, Sergi Caelles, Mario Lucic, and Eirikur Agustsson. 2022. VCT: A video compression transformer. arXiv preprint arXiv:2206.07307 (2022)

    2022 arXiv

  21. [29]

    In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition

    Dvc: An end-to-end deep video compression framework. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition . 11006–11015

  22. [30]

    Alexandre Mercat, Marko Viitanen, and Jarno Vanne. 2020. UVG dataset: 50/120fps 4K sequences for video codec analysis and development. InProceedings of the 11th ACM multimedia systems conference . 297–302

    2020

  23. [31]

    An end-to-end learning framework for video compression.IEEE transactions on pattern analysis and machine intelligence 43, 10 (2020), 3292–3308

    2020

  24. [32]

    Woonsung Park and Munchurl Kim. 2020. Deep predictive video compression using mode-selective uni-and bi-directional predictions based on multi-frame hypothesis. IEEE Access 9 (2020), 72–85

    2020

  25. [33]

    Alexandre Mercat, Arttu Mäkinen, Joose Sainio, Ari Lemmetti, Marko Viitanen, and Jarno Vanne. 2021. Comparative rate-distortion-complexity analysis of VVC and HEVC video codecs. IEEE Access 9 (2021), 67813–67828

    2021

  26. [34]

    Oren Rippel, Alexander G Anderson, Kedar Tatwawadi, Sanjay Nair, Craig Lytle, and Lubomir Bourdev. 2021. Elf-vc: Efficient learned flexible-rate video coding. In Proceedings of the IEEE/CVF International Conference on Computer Vision . 14479– 14488

    2021

  27. [35]

    Woonsung Park and Munchurl Kim. 2019. Deep predictive video compression with bi-directional prediction. arXiv preprint arXiv:1904.02909 (2019)

    2019 arXiv

  28. [36]

    Xihua Sheng, Li Li, Dong Liu, and Shiqi Wang. 2025. Bi-Directional Deep Con- textual Video Compression. IEEE Transactions on Multimedia (2025)

    2025

  29. [37]

    Linfeng Qi, Jiahao Li, Bin Li, Houqiang Li, and Yan Lu. 2023. Motion informa- tion propagation for neural video compression. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition . 6111–6120

    2023

  30. [38]

    Chuanbo Tang, Zhuoyuan Li, Yifan Bian, Li Li, and Dong Liu. 2025. Neural Video Compression with Context Modulation. In Proceedings of the Computer Vision and Pattern Recognition Conference . 12553–12563

    2025

  31. [39]

    Xihua Sheng, Jiahao Li, Bin Li, Li Li, Dong Liu, and Yan Lu. 2022. Temporal context mining for learned video compression. IEEE Transactions on Multimedia 25 (2022), 7311–7322

    2022

  32. [40]

    Ye-Kui Wang, Robert Skupin, Miska M Hannuksela, Sachin Deshpande, Virginie Drugeon, Rickard Sjöberg, Byeongdoo Choi, Vadim Seregin, Yago Sanchez, Jill M Boyce, et al. 2021. The high-level syntax of the versatile video coding (VVC) standard. IEEE Transactions on Circuits and Sy...

    2021

  33. [41]

    Gary J Sullivan, Jens-Rainer Ohm, Woo-Jin Han, and Thomas Wiegand. 2012. Overview of the high efficiency video coding (HEVC) standard.IEEE Transactions on circuits and systems for video technology 22, 12 (2012), 1649–1668

    2012

  34. [42]

    Ren Yang, Fabian Mentzer, Luc Van Gool, and Radu Timofte. 2020. Learning for video compression with hierarchical quality and recurrent enhancement. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 6628–6637

    2020

  35. [43]

    Haiqiang Wang, Weihao Gan, Sudeng Hu, Joe Yuchieh Lin, Lina Jin, Longguang Song, Ping Wang, Ioannis Katsavounidis, Anne Aaron, and C-C Jay Kuo. 2016. MCL-JCV: a JND-based H. 264/AVC video quality assessment dataset. In 2016 IEEE international conference on image processing (IC...

    2016

  36. [44]

    Yiwei Zhang, Guo Lu, Yunuo Chen, Shen Wang, Yibo Shi, Jing Wang, and Li Song. 2023. Neural rate control for learned video compression. In The Twelfth International Conference on Learning Representations

    2023

  37. [45]

    Jinxi Xiang, Kuan Tian, and Jun Zhang. 2022. Mimt: Masked image modeling transformer for video compression. In The Eleventh International Conference on Learning Representations

    2022

  38. [47]

    M Akin Yilmaz and A Murat Tekalp. 2020. End-to-end rate-distortion optimization for bi-directional learned video compression. In2020 IEEE International Conference on Image Processing (ICIP) . IEEE, 1311–1315

    2020

  39. [2019]

    In Proceedings of the IEEE/CVF international conference on computer vision

    Video compression with rate-distortion autoencoders. In Proceedings of the IEEE/CVF international conference on computer vision . 7033–7042

  40. [2020]

    In Computer Vision–ECCV 2020: 16th European Conference, Glasgow, UK, August 23–28, 2020, Proceedings, Part II 16

    Improving deep video compression by resolution-adaptive flow coding. In Computer Vision–ECCV 2020: 16th European Conference, Glasgow, UK, August 23–28, 2020, Proceedings, Part II 16 . Springer, 193–209

    2020

Pith tools

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