acceptodds
Under review as a conference paper at ICLR 2027

VoRTeC: Taming Foundation Flow for one-step Real Time Video Compression

Abstract

Ultra-low bitrate video compression still faces critical challenges: traditional neural video compres- sion inevitably introduces blurring artifacts, while diffusion-based generative video compression suffers from excessive decoding latency and poor temporal consistency. To address these issues, we propose VoRTeC, a Video Compression framework built upon a foundational flow model (Wan2.1). By compactly encoding latent video representations, predicting the positions of compressed represen- tations along flow trajectories, and integrating multi-scale priors, VoRTeC enables the compressor to harness generative video flow priors effectively. Without accessing the parameters or gradients of flow matching networks, our framework achieves one-step decoding and reconstructions with high perceptual fidelity. Meanwhile, we maintain consistency across frame groups via tail-frame reuse and prior caching. Extensive experiments demonstrate that our method reduces bit consumption by 58% compared to prior diffusion-based approaches, with decoding speed boosted by 3 to 197 times: VoRTeC achieves a decoding speed of 13 FPS at 720p and 32 FPS at 480p.

Primary area:
generative models
Keywords:
Diffusion model,generative model, video compression

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_gwuLOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

VoRTeC maps compressed video latents to a flow state so a pretrained Wan2.1 FlowDiT can refine them in one step. FPMF fuses the result with the compressed latents, and caches connect frame groups. The paper reports 32.31 fps decoding at 480p on an A100; VoRTeC+ additionally fine-tunes FlowDiT with LoRA.

Critical strengths
  • S1 — Clear technical goal. The flow-state estimate, FPMF, and inter-group caches address the costs and errors of one-step generative decoding. The base model does not require FlowDiT fine-tuning.
  • S2 — Promising 480p results. Decoding reaches 32.31 fps on an A100. Removing flow-state estimation, FPMF, or the prior cache raises LPIPS BD-rate by 44.31%, 72.59%, or 19.17%, respectively, relative to the full model.
Critical weaknesses
  • W1 — Distinction from GNVC-VD. GNVC-VD already refines compressed spatiotemporal latents with a video prior and adapts to compression artifacts. Please clarify what is technically new in the flow-state estimate, FPMF, and inter-group cache relative to its multi-step refinement and conditioning.
  • W2 — Limited scope of the speed and quality claims. Decoding reaches 32.31 fps32.31\,\mathrm{fps} at 480p, but only 12.60 fps12.60\,\mathrm{fps} at 720p and 3.93 fps3.93\,\mathrm{fps} at 1080p; end-to-end streaming latency is not reported. On HEVC-B 1080p and UVG1080P, GNVC-VD outperforms VoRTeC on LPIPS-VGG according to Table 2 (BD-rate: −34.58%-34.58\% and −23.88%-23.88\%). (Table 2)
  • W3 — Unclear rate comparisons and GOP settings. The claimed ∼58%\sim 58\% rate saving—compared with which method? The paper does not make this clear. VTM's -IntraPeriod is left unspecified, and DCVC's GOP settings are not reported. Thus, key-frame intervals cannot be checked against VoRTeC's 25-frame meta-group. These gaps do not establish an error, but they limit the comparison.
  • W4 — Unclear gradients. The computation of σ\sigma, the scaling in Eqs. (9)–(10), and the gradient paths through zcgz_c^g, z^0g\hat z_0^g, and z~0g\tilde z_0^g are not sufficiently specified.
  • W5 — Incorrect bibliographic references. For example, ECVC's second author is Junru Li, not Jiahao Li; GLC-Video's first author is Linfeng Qi, not Libiao Qi.
Questions for authors
  1. What does each of the flow-state estimate, FPMF, and inter-group cache add over GNVC-VD? Which gains come mainly from fewer FlowDiT evaluations? Please provide matched comparisons or ablations where possible.
  2. Why does GNVC-VD perform better on some 1080p perceptual metrics in Table 2? Can you compare quality and speed at matched rates or quality levels?
  3. Please report the GOP structure, intra period, and first-frame codec for VTM, DCVC, and the other baselines, along with VoRTeC's I-/P-Group ratio and reset interval. How do these settings affect the bitrate comparison?
  4. Please specify the forward and backward paths through zcgz_c^g, z^0g\hat z_0^g, and z~0g\tilde z_0^g. Is FlowDiT parameter-frozen only, or is its input gradient also stopped?
  5. Does Table 1a include the VAE, entropy coding, FPMF, and caches? Please report end-to-end and first-frame latency, peak memory, hardware, and batch size.
  6. Please correct the mis-cited papers and duplicate Wan entry, including any affected in-text citations.

Selected citation checks

These BibTeX entries are reference checks, not a complete bibliography. Check publication metadata against the original sources.

@InProceedings{Jiang_2025_CVPR,
    author    = {Jiang, Wei and Li, Junru and Zhang, Kai and Zhang, Li},
    title     = {ECVC: Exploiting Non-Local Correlations in Multiple Frames for Contextual Video Compression},
    booktitle = {Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)},
    month     = {June},
    year      = {2025},
    pages     = {7331-7341}
}
@inproceedings{careil2024towards,
  title={Towards image compression with perfect realism at ultra-low bitrates},
  author={Careil, Marlene and Muckley, Matthew J and Verbeek, Jakob and Lathuili{\`e}re, St{\'e}phane},
  booktitle={International Conference on Learning Representations},
  year={2024}
}
@inproceedings{lu2024deep,
  title={Deep hierarchical video compression},
  author={Lu, Ming and Duan, Zhihao and Zhu, Fengqing and Ma, Zhan},
  booktitle={Proceedings of the AAAI Conference on Artificial Intelligence},
  volume={38},
  number={8},
  pages={8859--8867},
  year={2024}
}
@article{ma2025diffusion,
  title={Diffusion-based perceptual neural video compression with temporal diffusion information reuse},
  author={Ma, Wenzhuo and Chen, Zhenzhong},
  journal={ACM Transactions on Multimedia Computing, Communications and Applications},
  volume={21},
  number={12},
  pages={1--22},
  year={2025},
  publisher={ACM New York, NY}
}
@inproceedings{mercat2020uvg,
  title={UVG dataset: 50/120fps 4K sequences for video codec analysis and development},
  author={Mercat, Alexandre and Viitanen, Marko and Vanne, Jarno},
  booktitle={Proceedings of the 11th ACM multimedia systems conference},
  pages={297--302},
  year={2020}
}
@article{qi2025generative,
  title={Generative latent coding for ultra-low bitrate image and video compression},
  author={Qi, Linfeng and Jia, Zhaoyang and Li, Jiahao and Li, Bin and Li, Houqiang and Lu, Yan},
  journal={IEEE Transactions on Circuits and Systems for Video Technology},
  volume={35},
  number={10},
  pages={10500--10515},
  year={2025},
  publisher={IEEE}
}
@misc{vonderfecht2025lossy,
  title={Lossy Compression with Pretrained Diffusion Models},
  author={Vonderfecht, Jeremy and Liu, Feng},
  year={2025},
  eprint={2501.09815},
  archivePrefix={arXiv}
}
@inproceedings{yin2024one,
  title={One-step diffusion with distribution matching distillation},
  author={Yin, Tianwei and Gharbi, Micha{\"e}l and Zhang, Richard and Shechtman, Eli and Durand, Fredo and Freeman, William T and Park, Taesung},
  booktitle={2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)},
  pages={6613--6623},
  year={2024},
  organization={IEEE}
}

Duplicate entry

The paper lists Wan (arXiv:2503.20314) twice under different author names. Merge the entries and use one in-text citation.

@article{wan2025wan,
  title={Wan: Open and advanced large-scale video generative models},
  author={Wan, Team and Wang, Ang and Ai, Baole and Wen, Bin and Mao, Chaojie and Xie, Chen-Wei and Chen, Di and Yu, Feiwu and Zhao, Haiming and Yang, Jianxiao and others},
  journal={arXiv preprint arXiv:2503.20314},
  year={2025}
}
8 comments
AuthorsOpenReview ↗

Response to Reviewer gwuL

We thank reviewer gwul for carefully reading our manuscript. You have identified relatively serious oversights in the paper. Fortunately, these oversights are amendable. We will make every effort to correct these errors and address your questions.

W5 & Q6 :Incorrect bibliographic references

R1: We recognize that this is a serious error and are willing to bear the consequences for it. We greatly appreciate you pointing this out, which has prevented bigger problems from arising later.

We have checked and proofread against the references item by item, and revised the incorrect citations.

W1 & Q1:Distinction from GNVC-VD

R2: Thank you for your comment. This is a good question. I will discuss the differences between the two from two aspects: technical details and starting points.

In terms of technical details, vortec is a type of "non-intrusive" solution. It achieves single-step decoding through FSE, FPMF and inter-group communication modules, even when we cannot obtain the gradient of the denoising network FlowDIT. It is worth noting that thanks to these designs, VoRTeC does not require multi-step FlowDIT evaluation like other flow-based methods. Single-step flow-dit evaluation is our goal, rather than the reason for the excellent performance.

In GNVC-VD, similar modules are absent. GNVC treats the flow matching network as part of the compression decoder and injects compressed representations into FlowDit as conditions. We argue that, on the one hand, this tightly couples the compressor and denoising network architectures, limiting transfer flexibility. On the other hand, GNVC requires multiple sampling steps for decoding, which substantially increases latency. (This also explains why VoRTeC achieves a decoding speed 6–8 times faster than GNVC.)

Furthermore, the intra-group correlation of GNVC is captured by the causal convolution of the base model (wanvae): this approach incurs a very high computational cost. The table below records the number of frames per group and the occupied VRAM used by VoRTeC and GNVC

GOPVRAM-train(batch-size=2, resolution=512*512)VRAM-test(1080p)
9 (VoRTeC)41789MB59132MB
1769355MB80254MB
25 (GNVC-VD)OOMOOM

It can be observed that the memory consumption for training and inference increases linearly as the GOP grows, which makes it infeasible to achieve inter-frame interaction merely by increasing the GOP. Most importantly, GNVC lacks the capability of cross-group information interaction. This means that when processing long videos, GNVC can only guarantee continuity by expanding the group size, which is impractical in engineering implementations. Moreover, the base stream model does not support such long inputs (e.g., Wan2.1 can generate a maximum of 81 frames in a single pass).

In contrast, the proposed VoRTeC scheme enables information interaction between groups via prior_cache and frame_cache (as also demonstrated in Figure 14 and Figure 9). It enhances inter-frame continuity without introducing extra training or inference overhead.

In terms of design philosophy, GNVC is a scheme tightly coupled with the Wan2.1 model. VoRTeC aims to implement a compression scheme decoupled from the model architecture, allowing the compression algorithm to genuinely leverage generative priors, rather than treating the generative structure merely as an enhancer for decoding. We believe this is highly significant. It means the compression performance can improve alongside advances in generative capabilities, while boosting the algorithm's transferability across vertical domains. The table below presents partial encoding and decoding performance of VORTEC on Wan2.1 14B and CogVideoX:

Base-flowbpplpipsdists
wan2.1(1.3B)+VoRTeC0.00550.1970.117
wan2.1 (14B)+VoRTeC0.00510.1890.105
CogVideoX(5B) +VoRTeC0.00530.1920.110

It can be observed that as the generative capability of the model improves, the decoding perceptual quality of vortec will also increase accordingly. Due to limited time and computing resources, we have only conducted partial experiments. In subsequent versions, we plan to include more ablation studies on base flow models to fully verify the transferability of our method.

AuthorsOpenReview ↗

Author Response (Part II)

W2 & Q2: Limited scope of the speed and quality claims

R3: Thank you for your suggestions !

We apologize that we have made overclaims to some extent. On one hand, we have honestly reported the decoding rates at 480p and 720p resolutions in the abstract without obfuscating this issue. On the other hand, we argue that achieving real-time decoding even at 480p is non-trivial, especially with the help of a pre-trained diffusion prior: our decoding speed is 5–7 times that of GNVC and 30–100 times that of Free-GVC.

In terms of performance, VoRTeC is indeed inferior to GNVC on certain metrics for the 1080p dataset. However, as stated in R2, VoRTeC is an external, non-intrusive approach. It trades off a slight performance drop for lower decoding latency, better temporal consistency, and stronger transferability. Meanwhile, VoRTeC only requires lightweight fine-tuning with LoRA, and it can outperform GNVC while retaining the aforementioned advantages. We believe this should not be overlooked and constitutes the core contribution of our method.

We have fully presented the comparison between GNVC-VD and our proposed method in Figure 5. The table below shows performance comparisons at selected bitrates, which we hope addresses your questions.

MethodBPPLPIPSDISTS
GNVC-VD0.01630.1040.062
VoRTeC0.01610.1030.061
VoRTeC+0.01150.1080.063

Q3 & W3 : Unclear rate comparisons and GOP settings.

R4: We appreciate the reviewer's comments! Indeed, we made a minor oversight when reporting the rate. The 58 rate saving is the result compared with GLC-Video, where S2VC was used as the base-model for this comparison. We will add further summarized data to the subscript (tested on 1080p, HEVC-B):

BD-LPIPSBD-DistsBD-FID
S2VC000
GLC-Video-52.1-42.93-44.43
DCVC-RT502.081044.61039.84
DiffVC79.2742.59None
SEVC509.45NoneNone
VoRTeC-58.12-49.9-57.32

However, we found that using S2VC as the base-model yields a large number of None values. Therefore, we adopt vortec as the baseline in the main text (Table 2). This also explains why the value of 58 does not appear: this is caused by the inherent calculation characteristics of bd-rate, rather than an experimental error. Nevertheless, we recognize that this may confuse readers, and we will revise it in the updated version.

Furthermore, for the number of GOPs of VTM and DCVC, we adopt the official standard protocol, namely intra_period=-1. We will also update this description in the new version. In fact, intra_period=-1 also delivers the optimal performance for these methods in most scenarios. We also set the GOP of these schemes to 25 in the table below (consistent with VoRTeC).

BPPLPIPSDISTS
VoRTeC (GOP=9+9+8=25)0.00870.1590.093
VTM(GOP=25)0.01440.3820.161
DCVC-RT(GOP=25)0.00920.3010.155

As can be seen, after adjustment, the bpp of the baseline methods has increased substantially, while the improvement in perceptual quality is not obvious. Overall, our method still exhibits significant advantages, which does not affect the conclusion.

For other baselines, we directly report the optimal GOP specified in their papers. For these works, either changing the GOP does not significantly affect performance, or the GOP cannot be adjusted. For instance, the GOP of GNVC-VD is determined during training and cannot be flexibly modified at test time to adjust the bitrate. In contrast, the proposed vortec enables this capability, as illustrated in Figure 12 of the original paper.

It is worth noting that although VoRTeC also supports setting intra_period to -1, the error accumulation of generative methods is more severe than that of reconstruction-based methods. Therefore, this configuration yields worse relative performance compared with periodic refresh settings.

AuthorsOpenReview ↗

Author Response (Part III)

W4 & Q4 : Unclear gradients

R4: Thank you for your suggestions! We will answer your questions in detail.

The calculation process for σ\sigma is actually provided on Line 270:

σ2=∥zcg−z0g∥2\sigma^2=\left\|z_c^g-z_0^g\right\|_2

For the scaling part of Eqs. (9)–(10), this actually originates from the basic setup of the flow model and is a transformation of a formula on line 155 of our paper: xt=(1−t)x0+tx1\mathbf{x}_t=(1-t) \mathbf{x}_0+t \mathbf{x}_1

x0=xt−tvθ(xt)\mathbf{x}_0=\mathbf{x}_t - tv _ \theta (\mathbf{x}_t) Since our estimated ⌊t⌉\lfloor t\rceil comes from the formula on line 264: t⋆=arg⁡min⁡tdist⁡[(1−t)∗zcg,FL(t)],t∈[0,1]t^{\star}=\underset{t}{\arg \min } \operatorname{dist}\left[(1-t) * z_c^g, \mathbf{F L}(t)\right], t \in[0,1] After estimation: zt≈(1−t)zcg=(1−⌊t⌉1000)zcgz_t \approx (1-t)z_c^g=\left(1-\frac{\lfloor t\rceil}{1000}\right)z_c^g

This is where the weighting term comes from. We will add further explanations in subsequent versions to improve the readability of the paper.

Regarding the gradient propagation process. In the VoRTeC version, we do not need to obtain the gradients of FlowDIT, as we described in the introduction section. Therefore, the gradient propagation path at this time is:

z~0g→zcg\tilde{z} _0^g \rightarrow z_c^g

In VoRTeC+, we enable gradient access for FlowDIT and use LoRA for simple fine-tuning. At this point, the gradient propagation path is: z~0g→z^0g→zcg\tilde{z}_0^g \rightarrow \hat{z}_0^g \rightarrow z_c^g We will revise Figure 2 in the new version of the paper to clearly demonstrate the gradient flow under different modes.

Q5: Does Table 1a include...Please report end-to-end and first-frame latency.

R5: Table 1 covers the latency of all modules. Indeed, the existing latency experiments are insufficient. We have added the encoding and decoding speeds for different frames, the peak video memory usage at different resolutions, and the batch_size used as you suggested:

resolutioni_frame_decodep_frame_decodeE2E latencypeak_memory (batch=1,GOP=9)
1080p0.601s0.163s0.514s59132MB
720p0.199s0.05s0.193s38021MB

In the tests, we only used batch_size=1. In fact, decoding can be accelerated via multi-card parallelism. However, the baseline methods all report serial decoding time, which would make the comparison unfair. Therefore, we adopted batch_size=1 only. During training, we used batch_size=2.

Reviewer_wwZvOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes VoRTeC, a one-step generative video compression framework built upon a pre-trained video flow model. It introduces Flow-State Estimation (FSE), Flow Prior Multi-Fusion (FPMF), and Contact Group to Group (CGG) to improve perceptual quality, temporal consistency, and decoding efficiency. Experiments demonstrate competitive rate-perception performance with efficient decoding.

Critical strengths

1.The one-step generative decoding framework significantly reduces the computational complexity of foundation-model-based video compression.

2.The proposed FPMF and CGG modules provide meaningful designs for improving compression performance and temporal consistency.

3.The experiments are relatively comprehensive, including multiple datasets, ablation studies, temporal consistency evaluations, and decoding speed comparisons.

Critical weaknesses

1.Insufficient validation of FSE assumptions. The derivation assumes Gaussian compression errors, which may not accurately characterize the errors introduced by nonlinear transforms and quantization. The current ablations do not directly validate the analytical timestep estimation.

2.Limited evidence for the claimed redundancy reduction. Although the ablations demonstrate the effectiveness of FPMF, it remains unclear whether the compression gains specifically come from removing prior-reconstructable information. More direct quantitative analysis would strengthen this claim.

3.Unclear stability of long prediction chains. CGG introduces dependencies across consecutive P-Groups, which may lead to error accumulation or temporal drift. The current experiments do not sufficiently investigate this potential issue.

Questions for authors

1.Can the authors analyze the empirical distribution of compression errors and compare the estimated flow timestep with an oracle-optimal timestep obtained through exhaustive search?

2.Can the authors provide more direct quantitative evidence of redundancy reduction, such as comparing bitrate allocation with and without the flow prior at matched reconstruction quality?

3.How do reconstruction quality and temporal consistency change with increasing P-Group distance from the most recent I-Group?

5 comments
AuthorsOpenReview ↗

Response to Reviewer wwZv

We sincerely appreciate the reviewers' recognition of the efficient decoding, the novelty of the FPMP/CGG algorithm, and the sufficiency of the experiments in our manuscript. We will address your questions as thoroughly as possible and welcome discussions with you at any time.

W1 & Q1: Insufficient validation of FSE assumptions/ Can the authors analyze the empirical distribution of compression errors...

R1: Thank you for the constructive suggestions from the authors! We will answer your questions in detail.

  1. We have added a visual comparison between compressed noise and Gaussian noise, along with empirical distribution plots in the appendix of the revised paper, to qualitatively analyze the reliability of FSE. In general, compressed noise does differ somewhat from Gaussian noise in form. However, what FSE does is not simply hard embedding of compressed noise; instead, it matches the compressed noise to a flow state with a similar signal-to-noise ratio, and the compressed noise gradually approximates Gaussian noise during end-to-end training. In other words, FSE provides a favorable matching paradigm that enables stable and easier convergence of training. The table below records the MSE between compressed noise and Gaussian noise at the corresponding stages before training and after E2E training:
Before E2EAfter E2E
N_mse0.003180.00103

whereNmse=∥(1−⌊t⌉)(zcg−z0)−⌊t⌉ϵ∥2N_{mse}=\|(1-\lfloor t \rceil)(z_c^g-z_0)-\lfloor t \rceil \epsilon\|_2

It can be observed that end-to-end training makes the compression noise closer to Gaussian noise.

  1. We have followed your suggestion and tested the gap between the oracle-optimal timestep (oot) and the FSE estimation at certain bitrates (tested on HEVC-B 1080p):
method (bpp=0.008)tlpipsdists
FSE ⌊t⌉\lfloor t \rceil1430.15920.0932
OOT (search in [⌊t⌉−50\lfloor t \rceil-50,⌊t⌉+50\lfloor t \rceil+50])1460.15880.0933
method (bpp=0.013)tlpipsdists
FSE ⌊t⌉\lfloor t \rceil1240.13320.0749
OOT (search in [⌊t⌉−50\lfloor t \rceil-50,⌊t⌉+50\lfloor t \rceil+50])1260.13310.0751

Our optimal search strategy finds the flow state closest to the compressed representation, with MSE adopted as the distance metric. It can be observed that the estimates from FSE show little difference from brute-force search, and the perceptual quality after decoding is also similar. The performance gain brought by the searched optimal value is close to random fluctuations. This also verifies the effectiveness of our proposed FSE.

W2 & Q2:Limited evidence for the claimed redundancy reduction

R2: Thank you for your comment! We present the equal bitrate allocation results with and without the flow prior in Figure 11. This figure demonstrates the importance of the flow prior: with the prior incorporated, bitrates can be better allocated to specific objects and textures, while highly repetitive regions such as the background are generated by the prior. The table below also shows an ablation experiment on the flow prior (tested on HEVC-B 1080p).

BD-rate(LPIPS)BD-rate(DISTS)BD-rate(FID)
VoRTeC000
VoRTeC(W/o prior)75.41%69.27%105.38%

Without using the flow prior, its performance is close to that without prior fusion: both significantly degrade perceptual quality. We will add ablation experiments for this part in subsequent versions.

AuthorsOpenReview ↗

Author Response (Part II)

W3&Q3: Unclear stability of long prediction chains.

R3: Thank you for your insightful comment! This is an excellent question. We have actually discussed this issue to some extent in the appendix; please refer to Figure 13. In this ablation experiment, we adjusted the ratio of P-Group: I-group within a meta group. As this ratio increases, the bitrate decreases significantly, at the cost of a certain degree of reduction in perceptual quality. We believe that the decline in perceptual quality is related to error accumulation. Overall, however, the R-D curve changes gently as the ratio is adjusted, which means such error accumulation is acceptable. Meanwhile, we also present the temporal consistency for different ratios in the table below, measured using flolpips (tested on UVG):

bppflo-lpips
VoRTeC (group=0)0.01420.137
VoRTeC (group=1)0.00860.138
VoRTeC (group=2)0.00690.139
VoRTeC(group=3)0.00610.139
VoRTeC (group=4)0.00550.140

Overall, as the number of P-Groups increases, the bpp decreases significantly, while flolpips does not change much. This indicates that intra-group information communication is effective. As the number of P-Groups decreases, the rise in bitrate improves consistency, yet such consistency is also hampered by the high refresh frequency of I-Groups. In general, although empirical judgment is required to balance temporal consistency and bitrate, the overall fluctuation is small and within a controllable range.

AuthorsOpenReview ↗

A Gentle Reminder of the Final Feedback

Dear Reviewer wwZv

We are extremely grateful for your careful review of our article! However, as the review deadline approaches, we are eager to know if your issue has been resolved. Your feedback would be greatly appreciated, and we would be delighted to engage in further discussions if needed

Sincerely,

The Authors

Reviewer_fWgHOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

The paper presents a method for video compression using a generative approach. The main contribution lies in how to use a generative prior, useful for quality in especially low bitrates, while not sacrificing significant temporal stability and increasing decoding time common of previous generative approaches for the task. The method achieves this by using a foundation flow model. The method has been evaluated on multiple common datasets and compared to various compression methods, including different families (traditional, neural, and generative codecs), showing improvement especially for low bitrates.

Critical strengths

S1: Successfully addressing both limitations of previous generative based approaches such as decoding latency and temporal inconsistency, and as a result can benefit from the generative prior leading to improved quality especially for low bit rates. This is a significant and relevant improvement compared to the previous work. This is shown on multiple datasets and in comparison to various methods giving strong empirical evidence for the made claims.

S2: The proposed solution is interesting to read (e.g., how the flow-matching network can be used efficiently for the task) and has some nice benefits (allowing training-free bitrate selection often not the case for learning-based video compression methods). Overall the paper is nicely written.

Critical weaknesses

W1: The paper can benefit from some minor clarifications in various parts:

  • Clarifying when VoRTeC and when VoRTeC+ are used in experiments (e.g., both are mentioned to be shown in Fig. 1).
  • In Table 1b, are there absolute BD-rate numbers or differences to the full method? If it is relative, it would be good to add a '+'.
  • Flow-step estimation (FSE) is used as a term in the caption of Figure 4, but it is not clearly stated in the method. Makes it difficult to find the corresponding description.
  • Why a text embedding is suddenly used is unclear (l. 280).

-> After rebuttal:

  • Line 092 still mentions that VoRTeC+ is shown in Figure 1, which is not the case.
  • addressed
  • addressed
  • partly addressed, improved version promised for final version
Questions for authors

What are the remaining limitations of the presented method? (Not to weaken the contribution, but to give an honest impression of the presented method)

-> Discussed in the reply but not in the revised manuscript.

4 comments
AuthorsOpenReview ↗

Response to Reviewer fWgH

We greatly appreciate the reviewers’ careful review and their recognition of the decoding quality and temporal consistency of our work. We will do our best to address your questions and concerns.

W1: The paper can benefit from some minor clarifications....

R1: Thank you for your review. We will address your questions one by one.

  • Clarifying when VoRTeC and when VoRTeC+ are used in experiments: Thank you for your comment. We have checked each point that might cause confusion and made corrections in the new version of the article.
  • absolute BD-rate numbers.. Thank you for your careful reading. We have corrected Table 1b in the new version.
  • FSE is not clearly stated in the method Thank you for your comment! We have added prominent markings in the Methods section to improve the readability of the manuscript.
  • Why a text embedding is suddenly used is unclear Thank you for your comment. In fact, pre-trained base flow models usually require a text embedding as input. We use the default text string here to align the input. For rigor, we have noted this in the main text. However, we realize that this may cause confusion, and we will provide a clearer explanation in later versions.

Q1: What are the remaining limitations of the presented method?

R2:Thank you for recognizing our work! VoRTeC offers a highly transferable and flexible prior-based generative video compression scheme. We believe this work makes a notable contribution to the community. Nevertheless, we must acknowledge that this work has certain limitations: First, although VoRTeC achieves better inter-frame continuity compared with other generative methods, its temporal consistency is still inferior to that of reconstruction-based schemes at high bitrates. Simply increasing the bitrate cannot fully mitigate the subtle jitter introduced by generative modeling — especially when there are fast-moving objects in the video. We regard this as an important problem to be investigated and resolved in future research.

Second, the encoding of VoRTeC relies on causal convolution encoding layers, which imposes substantial encoding overhead and is not well suited for scenarios such as edge deployment and edge-cloud collaboration. Future work will focus on encoder lightweighting and encoding acceleration.

AuthorsOpenReview ↗

A Gentle Reminder of the Final Feedback

Dear Reviewer fWgH

We highly cherish your initial feedback, recognizing that your schedule may be quite full. Nevertheless, we humbly ask you to spare a moment to examine our responses to your concerns. Your insights, no matter how brief, would be profoundly valued. Moreover, we remain at your disposal to answer any further queries prior to the end of the rebuttal period.

Sincerely,

The Authors

Thanks for the clarifications of my comments. Given that the limitations do not seem to have been added to the revised paper, and there is no video to assess the resulting quality, I remain with my rating of weak acceptance for the paper.

AuthorsOpenReview ↗

Thank You for Your Feedback

Thank you very much for your reply! We apologize for not including the limitations and future work in the manuscript. Your comments are constructive. We have added the limitations of the proposed method and future work to the conclusion section in the revised version. Thank you for helping us improve our work!

Meanwhile, we have provided some visualized video results for your reference at the link https://anonymous.4open.science/r/Anonymous-Submission-1683/.

Due to network constraints and other factors, it is difficult for us to present a large number of raw decoded videos. We appreciate your understanding.

We greatly appreciate your positive assessment of our work!

Comments on the paper

3 comments
AuthorsOpenReview ↗

Modifications to the revision

We sincerely appreciate the suggestions from all reviewers. We have adopted most of the suggestions and made corresponding revisions to the manuscript: including correcting several typo errors, revising figures to make the paper clearer and easier to read, and adding supplementary experiments. All modified parts are highlighted in dark blue.

Experimental Revisions:

  • We modified the numerical presentation format in Table.1(b) by adding a "+" sign before positive values.
  • We added a more detailed runtime analysis in Supplementary Material A.3.1.
  • We added visualization and analysis of the normality of compressed noise in Supplementary Material A.4.
  • We added quantitative ablation experiments on the convection prior in A.5.

Figure and Caption Revisions:

  • We corrected some typo errors and all incorrect citations.
  • We added the VoRTeC logo in Figure 1 to avoid confusion.
  • We revised Figure 2 to better demonstrate the gradient flow across different versions.
Area_Chair_3SunOpenReview ↗

Please join the discussion before it closes

Thank you to the authors for the rebuttal and to the reviewers for engaging in the discussion. As the discussion period draws to a close, I encourage everyone to keep the conversation going. I’d appreciate a brief final assessment from each reviewer. After reading the rebuttal and the discussion so far, what is the main reason you lean toward acceptance or rejection, and what evidence supports that view? Please note which concerns the rebuttal resolved and which remaining concerns, if any, are decisive for your recommendation.

If your assessment differs from another reviewer’s, please point to the specific claim or evidence you interpret differently. Authors are welcome to clarify any factual points that remain open. This will help us identify where we agree and understand the reasons for any remaining disagreement.

AuthorsOpenReview ↗

Summary during the rebuttal period

We would like to thank the reviewers again for their careful review and valuable suggestions. During the review process, we have made the following additions and revisions.

  • We added the normality analysis of compressed noise to verify the reliability of the FSE assumption and address the reviewer wwZv comments.
  • We conducted a search for the optimal time step within a certain range to verify the accuracy of FSE estimation and address the reviewer wwZv comments.
  • We added ablation studies on the prior and CGG; please refer to our responses to the reviewer wwZv.
  • We included the performance of VoRTeC on different basis modes in the appendix to preliminarily verify the transferability of VoRTeC in response to the reviewer gwuL.
  • We reported the frame-by-frame encoding and decoding latency of VoRTeC and the adopted GOP settings in the appendix to address the reviewer gwuL comments.
  • We have added reflections on the limitations of VoRTeC and prospects for future work in the summary section and address the reviewer fWgH comments.
  • We have added some visualization cases of decoding videos. Please refer to our comments on the reviewer fWgH

Finally, we would like to emphasize that VoRTeC is fundamentally different from previous works based on base flows [1][2]. This is reflected in multiple aspects including the construction and intervention of compressed representations, as well as the adjustment and alignment of flow priors. Meanwhile, as a non-intrusive method, VoRTeC can regularize a multi-step generative prior into a single-step one while maintaining performance, without resorting to high-training-cost techniques such as DMD. This is non-trivial.

[1] Mao Q, Cheng H, Yang T, et al. Generative neural video compression via video diffusion prior[C]//Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 2026: 43239-43248.

[2] Ling X, Zhou C, Li C, et al. Free-gvc: Towards training-free extreme generative video compression with temporal coherence[J]. arXiv preprint arXiv:2602.09868, 2026.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.