- 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 at 480p, but only at 720p and 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: and ). (Table 2)
- W3 — Unclear rate comparisons and GOP settings. The claimed rate saving—compared with which method? The paper does not make this clear. VTM's
-IntraPeriodis 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 , the scaling in Eqs. (9)–(10), and the gradient paths through , , and 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
- 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.
- 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?
- 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?
- Please specify the forward and backward paths through , , and . Is FlowDiT parameter-frozen only, or is its input gradient also stopped?
- 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.
- 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
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.
| Method | BPP | LPIPS | DISTS |
|---|---|---|---|
| GNVC-VD | 0.0163 | 0.104 | 0.062 |
| VoRTeC | 0.0161 | 0.103 | 0.061 |
| VoRTeC+ | 0.0115 | 0.108 | 0.063 |
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-LPIPS | BD-Dists | BD-FID | |
|---|---|---|---|
| S2VC | 0 | 0 | 0 |
| GLC-Video | -52.1 | -42.93 | -44.43 |
| DCVC-RT | 502.08 | 1044.6 | 1039.84 |
| DiffVC | 79.27 | 42.59 | None |
| SEVC | 509.45 | None | None |
| 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).
| BPP | LPIPS | DISTS | |
|---|---|---|---|
| VoRTeC (GOP=9+9+8=25) | 0.0087 | 0.159 | 0.093 |
| VTM(GOP=25) | 0.0144 | 0.382 | 0.161 |
| DCVC-RT(GOP=25) | 0.0092 | 0.301 | 0.155 |
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.
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 is actually provided on Line 270:
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:
Since our estimated comes from the formula on line 264: After estimation:
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:
In VoRTeC+, we enable gradient access for FlowDIT and use LoRA for simple fine-tuning. At this point, the gradient propagation path is: 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:
| resolution | i_frame_decode | p_frame_decode | E2E latency | peak_memory (batch=1,GOP=9) |
|---|---|---|---|---|
| 1080p | 0.601s | 0.163s | 0.514s | 59132MB |
| 720p | 0.199s | 0.05s | 0.193s | 38021MB |
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.
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
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:
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.