Free-GVC: Toward Retraining-Free Extreme Generative Video Compression with Temporal Coherence
Listen
Radio episode about this paper
Transcript
Introduction to the show: ident: AI Radio. Generated commentary on the latest Artificial Intelligence papers.
Tom: We're looking at the paper: "Free-GVC: Toward Retraining-Free Extreme Generative Video Compression with Temporal Coherence".
Jane: Here's the gist of the paper. Building on recent advances in video generation, generative video compression has emerged as a new paradigm for achieving visually pleasing reconstructions <ref:2602.09868#pg2> The paper proposes Free-GVC, a training-free generative video compression framework that reformulates video coding as latent trajectory compression guided by a video diffusion prior <ref:2602.09868#pg2> Experiments show that Free-GVC achieves an average of 93.29% BD-Rate reduction in DISTS over the latest neural codec DCVCRT, and a user study further confirms its superior perceptual quality and temporal coherence at ultra-low bitrates <ref:2602.09868#pg2> Pruning redundant reasoning tokens The core innovation of Free-GVC is recasting compression as "progressive coding of noisy latent representations in the diffusion space" <ref:2602.09868#pg4> The framework operates at the groupof-pictures (GOP) level, encoding video segments into a compact latent space and progressively compressing them along the diffusion trajectory <ref:2602.09868#pg2> To ensure perceptually consistent reconstruction across GOPs, an Adaptive Quality Control module dynamically constructs an online rate–perception surrogate model to predict the optimal diffusion step for each GOP <ref:2602.09868#pg2> Additionally, an Inter-GOP Alignment module establishes frame overlap and performs latent fusion between adjacent groups, thereby mitigating flicker and enhancing temporal coherence <ref:2602.09868#pg2> How the training signal is built The framework utilizes a pretrained video diffusion prior to achieve high-quality generative compression with temporally consistent reconstruction <ref:2602.09868#pg2> The compression process begins from a random Gaussian sample zT ∼ N (0, I) and proceeds toward the target latent yk, terminating at an adaptive timestep t∗ determined by the quality control module <ref:2602.09868#pg2> The encoder leverages the video diffusion prior pθ, which operates on the GOP through its spatiotemporal attention and 3D causal convolutional architecture to implicitly preserve temporal coherence <ref:2602.09868#pg2> Where the gains come from The Inter-GOP Alignment mechanism consists of two complementary parts: Frame Overlap at the encoder and Latent Fusion at the decoder <ref:2602.09868#pg2> Frame Overlap involves sharing overlapping frames between adjacent GOPs to maintain temporal continuity across boundaries <ref:2602.09868#pg2> Latent Fusion blends overlapping latents in the latent space during decoding to ensure smooth transitions at GOP boundaries, effectively suppressing boundary artifacts and maintaining temporal coherence across GOPs <ref:2602.09868#pg2> Adaptive Quality Control The Adaptive Quality Control module addresses the challenge of stochastic behavior by formulating quality control as an online inverse rate–perception mapping problem <ref:2602.09868#pg2> This involves Sparse Timestep Sampling, where M candidate timesteps are uniformly sampled from the diffusion schedule to form a set S <ref:2602.09868#pg2> Online Surrogate Fitting constructs a lightweight surrogate model, adopting a power-law formulation P = α · Rβ as the surrogate function <ref:2602.09868#pg2> Finally, Iterative Refinement determines the optimal timestep t∗ by minimizing the deviation between achieved and target perceptual quality <ref:2602.09868#pg2> Inter-GOP Alignment The Inter-GOP Alignment mechanism is crucial for addressing discontinuities at GOP boundaries in dynamic scenes <ref:2602.09868#pg2> It consists of Frame Overlap at the encoder and Latent Fusion at the decoder <ref:2602.09868#pg2> Frame Overlap reuses the last m frames of a previous GOP as the first m frames of the current GOP to maintain temporal continuity across boundaries <ref:2602.09868#pg2> Latent Fusion blends overlapping latents in the latent space using a weighted combination to ensure smooth transitions at GOP boundaries, thereby effectively suppressing boundary artifacts and maintaining temporal coherence across GOPs <ref:2602.09868#pg2> Performance Comparison Quantitative results show that Free-GVC consistently outperforms all competing methods on DISTS and FID, demonstrating superior perceptual quality in terms of structural fidelity and distributional realism <ref:2602.09868#pg2> The method achieves performance comparable to the latest distortion-oriented codecs on the distortion-oriented metric MSSSIM, while substantially outperforming perception-oriented approaches <ref:2602.09868#pg2> A user study confirms that FreeGVC surpasses VVC and DCVC-RT at similar bitrates and even outperforms DCVC-RT at half its bitrate <ref:2602.09868#pg2> Ablation Study Ablation studies on the HEVC Class D dataset summarize the impact of key configurations <ref:2602.09868#pg2> The Inter-GOP Alignment mechanism shows that without alignment, frames exhibit temporal variation across frame indexes <ref:2602.09868#pg2> The quality control mechanism guides subsequent GOPs using the quality metric of the first GOP as a target to maintain stable perceptual quality <ref:2602.09868#pg2> The Inter-GOP Alignment mechanism is shown to improve temporal continuity by fusing overlapping latent representations, and a small overlap with m = 4 improves temporal continuity and perceptual quality <ref:2602.09868#pg2> The GOP Size study indicates that a size l of 48 achieves the best overall performance <ref:2602.09868#pg2> Complexity Analysis The encoding phase involves VAE encoding, diffusion-based distribution estimation, and RCC compression <ref:2602.09868#pg2> The decoding phase mirrors these operations with additional steps for unconditional denoising and VAE decoding <ref:2602.09868#pg2> While the framework achieves 1.25 frames per second (fps) at 0.001 bpp, it remains challenging to achieve real-time processing for high-resolution videos <ref:2602.09868#pg2> The RCC runtime scales more rapidly with bitrate than the diffusion computations themselves <ref:2602.09868#pg2> This paper presents a training-free generative video compression framework, termed Free-GVC, which leverages pretrained video diffusion models to jointly capture both spatial and temporal dependencies <ref:2602.09868#pg2> By reformulating compression as progressive latent coding within the diffusion space, our method achieves perceptually faithful and visually coherent reconstruction at ultra-low bitrates <ref:2602.09868#pg2> The Inter-GOP Alignment module effectively mitigates boundary artifacts and enhances temporal consistency, while the Adaptive Quality Control mechanism dynamically stabilizes perceptual quality across diverse content and GOPs <ref:2602.09868#pg2> Despite its strong performance, achieving real-time compression for high-resolution videos remains challenging for this framework <ref:2602.09868#pg2> REFERENCES
1: T.
Tom: Wiegand, G.
Jane: J.
Tom: Sullivan, G.
Paper summary: Tom: So wrapping up this discussion on "Free-GVC: Toward Retraining-Free Extreme Generative Video Compression with Temporal Coherence." The authors are Xiaoyue Ling and her team, and they propose this framework as a way to achieve extreme generative video compression without needing to retrain the model for the compression task itself.
Jane: They successfully showed that by reformulating coding as latent trajectory compression guided by a video diffusion prior, they can get perceptually faithful and visually coherent reconstructions at ultra-low bitrates <ref:2602.09868#pg2>. The key mechanisms are that the Inter-GOP Alignment module fixes boundary artifacts and improves temporal consistency, while the Adaptive Quality Control mechanism dynamically stabilizes perceptual quality across different GOPs <ref:2602.09868#pg2>.
Lu: What this means for the field is that we can leverage powerful, pretrained video diffusion models to capture both spatial and temporal dependencies simultaneously without needing a new training signal for compression <ref:2602.09868#pg4>. It shifts the focus toward using these generative priors directly in the compression pipeline <ref:2602.09868#pg3>.
Meng: I see how this fits into what we’re doing at scale, though I still have questions about the real-time performance for high resolution videos <ref:2602.09868#pg2>. It’s a great quality result, but we need to see that scaling behavior under heavy production load.
Jane: That's the honest limitation they point out: achieving real-time compression for high resolution videos remains challenging with this framework <ref:2602.09868#pg3>. It’s a testament to how much better the latent trajectory compression is compared to what we’ve seen before.
Tom: So, in simple terms, Free-GVC takes existing generative video models and uses them as a guide for compression, and it does that without retraining the model <ref:2602.09868#pg1>. It’s a way to get high-quality video compression where the quality is dictated by the diffusion process rather than a massive training effort.
Conclusion: Tom: So, Free-GVC is basically taking those big video diffusion models we use—the ones that can generate incredible video—and using them to compress existing videos without having to retrain anything for compression specifically.
Jane: It’s about making the compression process feel like a trajectory through a diffusion space, which lets it keep the visual flow consistent across different parts of the video.
Tom: The authors are Xiaoyue Ling and her team, and they show how this training-free approach can actually beat current neural codecs on things like DISTS by reducing the bitrate by over ninety percent.
Jane: That ninety-three point two nine percent reduction sounds huge when you think about how much data you can squeeze into a video stream while keeping it looking good.
Lu: What’s really interesting is how they build that quality control module, which dynamically adjusts the compression step based on what it thinks will look best at that moment along the diffusion path.
Meng: That sounds like a lot of bookkeeping for the decoder, though I'm still wondering how fast it actually runs when you try to push this on a production server with high-resolution footage.
Lalam: My analysis confirms that this framework successfully combines spatial and temporal information using a pretrained video diffusion prior, which is pretty clever for achieving that visual coherence.
Tom: So, the main thing they’re pointing at is that by focusing on progressive coding in the diffusion space, you get high quality and temporal consistency at ultra-low bitrates.
Jane: And they show how this technique handles those tricky boundaries between video segments by using frame overlap and latent fusion to smooth things out.
Tom: It basically means we can use these massive generative models as a powerful, fixed guide for compression instead of needing a whole new training cycle every time we want to optimize the codec.
Jane: If this holds up in real-world tests like the user studies they ran, it could mean a much more accessible way to get high-fidelity video on devices that really can't handle massive file sizes anymore.
Institute of Image Communication and Network Engineering, Shanghai Jiao Tong University · Shanghai AI Lab
cs.CV
Submitted: 2026-02-10
Updated: 2026-11-13
Comments: Accepted for publication in IEEE Transactions on Broadcasting
License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/
Importance score: 94/100
The gist: Building on recent advances in video generation, generative video compression has emerged as a new paradigm for achieving visually pleasing reconstructions <ref:2602.09868#pg2> The paper proposes…
Key concepts
- Latent Trajectory Compression
- This redefines video coding as compressing noisy latent representations along a path dictated by a video diffusion prior. Instead of traditional block-based encoding, the framework progressively refines these latent representations step-by-step, guided by the learned diffusion process to achieve high perceptual quality.
- Adaptive Quality Control
- This module dynamically determines the optimal compression step for each video segment (GOP). It models quality as an inverse rate-perception mapping and uses sparse timestep sampling and iterative refinement to ensure stable perceptual quality across different GOPs, managing the stochastic nature of the diffusion process.
- Inter-GOP Alignment
- This mechanism prevents visual flickering and temporal discontinuities between consecutive video segments (GOPs). It works by sharing overlapping frames at the encoder and blending overlapping latent representations during decoding. This fusion ensures smooth transitions and maintains temporal coherence across GOP boundaries.
Terminology
Summary
Building on recent advances in video generation, generative video compression has emerged as a new paradigm for achieving visually pleasing reconstructions <ref:2602. p. 009868#pg2> The paper proposes Free-GVC, a training-free generative video compression framework that reformulates video coding as latent trajectory compression guided by a video diffusion prior <ref:2602. p. 009868#pg2> Experiments show that Free-GVC achieves an average of 93. p. 029% BD-Rate reduction in DISTS over the latest neural codec DCVCRT, and a user study further confirms its superior perceptual quality and temporal coherence at ultra-low bitrates <ref:2602. p. 009868#pg2>
Pruning redundant reasoning tokens
The core innovation of Free-GVC is recasting compression as progressive coding of noisy latent representations in the diffusion space
<ref:2602.09868#pg4> The framework operates at the groupof-pictures (GOP) level, encoding video segments into a compact latent space and progressively compressing them along the diffusion trajectory <ref:2602. p. 009868#pg2> To ensure perceptually consistent reconstruction across GOPs, an Adaptive Quality Control module dynamically constructs an online rate–perception surrogate model to predict the optimal diffusion step for each GOP <ref:2602. p. 009868#pg2> Additionally, an Inter-GOP Alignment module establishes frame overlap and performs latent fusion between adjacent groups, thereby mitigating flicker and enhancing temporal coherence <ref:2602. p. 009868#pg2>
How the training signal is built
The framework utilizes a pretrained video diffusion prior to achieve high-quality generative compression with temporally consistent reconstruction <ref:2602.09868#pg2> The compression process begins from a random Gaussian sample zT ∼ N (0, I) and proceeds toward the target latent yk, terminating at an adaptive timestep t∗ determined by the quality control module <ref:2602. p. 409868#pg2> The encoder leverages the video diffusion prior pθ, which operates on the GOP through its spatiotemporal attention and 3D causal convolutional architecture to implicitly preserve temporal coherence <ref:2602. p. 409868#pg2>
Where the gains come from
The Inter-GOP Alignment mechanism consists of two complementary parts: Frame Overlap at the encoder and Latent Fusion at the decoder <ref:2602.09868#pg2> Frame Overlap involves sharing overlapping frames between adjacent GOPs to maintain temporal continuity across boundaries <ref:2602. p. 709868#pg2> Latent Fusion blends overlapping latents in the latent space during decoding to ensure smooth transitions at GOP boundaries, effectively suppressing boundary artifacts and maintaining temporal coherence across GOPs <ref:2602. p. 409868#pg2>
Adaptive Quality Control
The Adaptive Quality Control module addresses the challenge of stochastic behavior by formulating quality control as an online inverse rate–perception mapping problem <ref:2602.09868#pg2> This involves Sparse Timestep Sampling, where M candidate timesteps are uniformly sampled from the diffusion schedule to form a set S <ref:2602. p. 509868#pg2> Online Surrogate Fitting constructs a lightweight surrogate model, adopting a power-law formulation P = α · Rβ as the surrogate function <ref:2602. p. 609868#pg2> Finally, Iterative Refinement determines the optimal timestep t∗ by minimizing the deviation between achieved and target perceptual quality <ref:2602. p. 609868#pg2>
Inter-GOP Alignment
The Inter-GOP Alignment mechanism is crucial for addressing discontinuities at GOP boundaries in dynamic scenes <ref:2602.09868#pg2> It consists of Frame Overlap at the encoder and Latent Fusion at the decoder <ref:2602. p. 709868#pg2> Frame Overlap reuses the last m frames of a previous GOP as the first m frames of the current GOP to maintain temporal continuity across boundaries <ref:2602. p. 709868#pg2> Latent Fusion blends overlapping latents in the latent space using a weighted combination to ensure smooth transitions at GOP boundaries, thereby effectively suppressing boundary artifacts and maintaining temporal coherence across GOPs <ref:2602. p. 709868#pg2>
Performance Comparison
Quantitative results show that Free-GVC consistently outperforms all competing methods on DISTS and FID, demonstrating superior perceptual quality in terms of structural fidelity and distributional realism <ref:2602.09868#pg2> The method achieves performance comparable to the latest distortion-oriented codecs on the distortion-oriented metric MSSSIM, while substantially outperforming perception-oriented approaches <ref:2602. p. 909868#pg2> A user study confirms that FreeGVC surpasses VVC and DCVC-RT at similar bitrates and even outperforms DCVC-RT at half its bitrate <ref:2602. p. 109868#pg2>
Ablation Study
Ablation studies on the HEVC Class D dataset summarize the impact of key configurations <ref:2602.09868#pg2> The Inter-GOP Alignment mechanism shows that without alignment, frames exhibit temporal variation across frame indexes <ref:2602. p. 109868#pg2> The quality control mechanism guides subsequent GOPs using the quality metric of the first GOP as a target to maintain stable perceptual quality <ref:2602. p. 609868#pg2> The Inter-GOP Alignment mechanism is shown to improve temporal continuity by fusing overlapping latent representations, and a small overlap with m = 4 improves temporal continuity and perceptual quality <ref:2602. p. 109868#pg2> The GOP Size study indicates that a size l of 48 achieves the best overall performance <ref:2602. p. 909868#pg2>
Complexity Analysis
The encoding phase involves VAE encoding, diffusion-based distribution estimation, and RCC compression <ref:2602.09868#pg2> The decoding phase mirrors these operations with additional steps for unconditional denoising and VAE decoding <ref:2602.09868#pg2> While the framework achieves 1. p. 125 frames per second (fps) at 0. p. 8001 bpp, it remains challenging to achieve real-time processing for high-resolution videos <ref:2602. p. 909868#pg2> The RCC runtime scales more rapidly with bitrate than the diffusion computations themselves <ref:2602.09868#pg2>
Conclusion
This paper presents a training-free generative video compression framework, termed Free-GVC, which leverages pretrained video diffusion models to jointly capture both spatial and temporal dependencies <ref:2602.09868#pg2> By reformulating compression as progressive latent coding within the diffusion space, our method achieves perceptually faithful and visually coherent reconstruction at ultra-low bitrates <ref:2602. p. 109868#pg2> The Inter-GOP Alignment module effectively mitigates boundary artifacts and enhances temporal consistency, while the Adaptive Quality Control mechanism dynamically stabilizes perceptual quality across diverse content and GOPs <ref:2602. p. 109868#pg2> Despite its strong performance, achieving real-time compression for high-resolution videos remains challenging for this framework <ref:2602. p. 209868#pg2>
REFERENCES
[1] T. Wiegand, G. J. Sullivan, G. Bjontegaard, and A. Luthra, “Overview of the h. p. 3264/avc video coding standard,” IEEE Trans. p. 0 Circuits Syst. Video Technol. p. 0 p. 0 p. 0 p. 0 p. 0, vol. 13, no. 7, pp. 560–576, 2003 <ref:2602. p. 009868#pg2>
[2] G. J. Sullivan, J. Ohm, W. Han, and T. Wiegand, “Overview of the high efficiency video coding (HEVC) standard,” IEEE Trans. p. 0 Circuits Syst. Video Technol., vol. 22, no. 12, pp. 1649–1668, 2012 <ref:2602. p. 009868#pg4>
[3] B. Bross, Y.-K. Wang, Y. Ye, S. Liu, J. Chen, G. J. Sullivan, and J.-R. Ohm, “Overview of the versatile video coding (vvc) standard and its applications,” IEEE Trans. p. 0 Circuits Syst. Video Technol., vol. 31, no. 10, pp. 3736–3764 <ref:2602. p. 009868#pg2>
[4] J. Li, B. Li, and Y. Lu, “Neural video compression with feature modulation,” in Proc. p. 0 IEEE/CVF Conf. p. 0 p. 0 p. 0 p. 0 p. 0 p. 0 p. 0 p. 0 p. 0 Comput. Vis. Pattern Recognit., IEEE, 2024, pp. p. 0 26 099–26 108 <ref:2602. p. 009868#pg2>
[5] Z. Jia, B. Li, J. Li, W. Xie, L. Qi, H. Li, and Y.-Lu “Towards practical real-time neural video compression,” in Proc. p. 0 IEEE/CVF Conf. Comput. Vis., 2025, pp. 12 543–12 552 <ref:2602. p. 009868#pg2>
[6] W. Jiang, J. Li, K. Zhang, and L. p. 0 Zhang “Ecvc: Exploiting non-local correlations in multiple frames for contextual video compression,” in Proc. p. 0 IEEE/CVF Conf. Comput. Vis., 2025, pp. 7331–7341 <ref:2602. p. 009868#pg2>
[7] M. Li, Y. Shi, J.-Wang, and Y.-Huang “High visual-fidelity learned video compression,” in Proc. p. 0 ACM Int. Conf. Multimedia, 2023, pp. 8057–8066 <ref:2602. p. 009868#pg2>
[8] F. Mentzer, E. Agustsson, J.-Balle, D.-Minnen, N. Johnston, and G. Toderici “Neural video compression using gans for detail synthesis and propagation,” in Proc. p. 0 Eur. Conf. Comput. Vis., 2022, pp. 562–578 <ref:2602. p. 009868#pg2>
[9] R.-Yang, R.-Timofte, and L.-Vool “Perceptual learned video compression with recurrent conditional GAN,” in Proc. p. 2 Int. Joint Conf. p. 4 Artif. Intell., 2022, pp. 1537–1544 <ref:2602. p. 009868#pg2>
[10] D.-Podell, Z.-English, K.-Lacey, A.-Blattmann, T.-Dockhorn, J. Muller, J.-Penna, and R. Rombach “SDXL: improving latent diffusion models for high-resolution image synthesis,” in Int. p. 0 Conf. Learn. Represent., 2024 <ref:2602. p. 009868#pg2>
[11] T.-Brooks, B.-Peebles, C.-Holmes, W.-DePue, Y.-Guo, L.-Jing, D.-Schnurr, J. Taylor, T. Luhman, E. Luhman, C. Ng, R. Wang., and A. Ramesh “Video generation models as world simulators,” 2024 <ref:2602. p. 009868#pg2>
[12] B.-F Labs “Flux.1 kontext: Flow matching for in-context image generation and editing in latent space,” arXiv preprint arXiv:2506. p. 015742, 2025 <ref:2602. p. 0 p. 0 p. 0 p. 0 p. 0 p. 009868#pg4>
[13] Z.-Li, Y.-Zhou, H.-Wei, C.-Ge, and J.-Jiang “Toward extreme image compression with latent feature guidance and diffusion prior,” IEEE Trans. p. 1 Circuits Syst. Video Technol., vol. 35, no. 1, pp. 888–899 <ref:2602. p. 009868#pg2>
[14] M.-Careil, M.-Muckley, J.-Verbeek, and S.-Lathuiliere “Towards image compression with perfect realism at ultra-low bitrates,” in Int. p. 0 Conf. Learn. Represent., 2023 <ref:2602. p. 009868#pg2>
[15] L.-Theis, T.-Salimans, M.-Hoffman, and F.-Mentzer “Lossy compression with gaussian diffusion,” arXiv preprint arXiv:2206. p. 008889, 2022 <ref:2602. p. 009868#pg2>
[16] L.-Qi, Z.-Jia, J.-Li, B.-Li, H.-Li, and Y.-Lu “Generative latent coding for ultra-low bitrate image and video compression,” IEEE Trans. p. 0 Circuits Syst. Video Technol., 2025 <ref:2602.09868#pg2>
[17] T.-Zhang, X.-Luo, L.-Li, and D.-Liu “Stablecodec: Taming one-step diffusion for extreme image compression,” in Proc. p. 2 IEEE/CVF Int. p. 0 p. 0 p. 0 Conf. Comput. Vis., 2025 <ref:2602.09868#pg2>
[18] J.-Vonderfecht and F.-Liu “Lossy compression with pretrained diffusion models,” in Int. p. 2 Conf. Learn. Represent., 2025 <ref:2602.09868#pg4>
[19] W.-Ma and Z.-Chen “Diffusion-based perceptual neural video compression with temporal diffusion information reuse,” ACM Trans. p. 5 Multimedia Comput. Commun. Appl., vol. 21, no. 12, pp. 1–22, 2025 <ref:2602. p. 009868#pg2>
[20] B.-Li, Y.-Liu, X.-Niu, B.-Bait, W.-Han, L.-Deng, and D.-Gunduz “Extreme video compression with prediction using pre-trained diffusion models,” in Proc. p. 1 Int. Conf. Wireless Commun. Signal Process. p. 2, IEEE, 2024 <ref:2602. p. 009868#pg2>
[21] Z.-Wang, H.-Man, W.-Li, X.-Wang, X.-Fan, and D.-Zhao “T-gvc: Trajectory-guided generative video coding at ultra-low bitrates,” arXiv preprint arXiv:2507. p. 007633, 2025 <ref:2602. p. 009868#pg2>
[22] Y.-H. Chen, Y.-C. Yao, K.-W. Ho, C.-H. Wu, H.-T. Phung, M.-Benjak, J.-Ostermann, and W.-H. Peng “Hytip: Hybrid temporal information propagation for masked conditional residual video coding,” in Proc. p. 1 IEEE/CVF Int. Conf. Comput. Vis., 2025 <ref:2602.09868#pg2>
[23] J.-Liao, Y.-Wu, C.-Lin, Z.-Deng, L.-Li, D.-Liu, and X.-Sun “Ehvc: Efficient hierarchical reference and quality structure for neural video coding,” in Proc. p. 5 ACM Int. Conf. Multimedia, 2025 <ref:2602. p. 009868#pg2>
[24] X.-Sheng, L.-Li, D.-Liu, and H.-Li “Prediction and reference quality adaptation for learned video compression,” IEEE Trans. p. 0 Image Process. p. 2 p. 2, 2025 <ref:2602.09868#pg2>
[25] Y.-Bian, C.-Tang, L.-Li, and D.-Liu “Augmented deep contexts for spatially embedded video coding,” in Proc. p. 0 IEEE/CVF Conf. Comput. Vis. Pattern Recognit., 2025, pp. 2094–2104 <ref:2602. p. 009868#pg2>
[26] C.-Tang, Z.-Li, Y.-Bian, L.-Li, and D.-Liu “Neural video compression with context modulation,” in Proc. p. 0 IEEE/CVF Conf. Comput. Vis. Pattern Recognit., 2025, pp. 12 553–12 563 <ref:2602. p. 009868#pg2>
[27] G.-Gao, S.-Teng, T.-Peng, F.-Zhang, and D. p. 0-Bull “Givic: Generative implicit video compression,” in Proc. p. 0 IEEE/CVF Int. Conf. Comput. Vis., 2025, pp. 17 356–17 367 <ref:2602. p. 009868#pg2>
[28] G.-Lu, W.-Ouyang, D.-Xu, X.-Zhang, C.-Cai, and Z.-Gao “DVC: an end-to-end deep video compression framework,” in Proc. p. 1 IEEE/CVF Conf. Comput. Vis. Pattern Recognit., 2019, pp. 11 006–11 015 <ref:2602. p. 009868#pg4>
[29] J.-Li, B.-Li, and Y.-Lu “Deep contextual video compression,” in Adv. p. 0 Neural Inf. p. 0 p. 0 Process. Syst., 2021, pp. p. 0 p. 0 18 114–18 125 <ref:2602. p. 009868#pg4>
[30] ——, “Neural video compression with diverse contexts,” in Proc. IEEE/CVF Conf. Comput. Vis., 2023, pp. 22 616–22 626 <ref:2602. p. 009868#pg2>
[31] M.-Lu, Z.-Duan, F.-Zhu, and Z.-Ma “Deep hierarchical video compression,” in Proc. p. 1 AAAI Conf. Artif. Intell., vol. 38, no. 8, 2024, pp. 8859–8867 <ref:2602. p. 009868#pg4>
[32] M.-Lu, Z.-Duan, W.-Cong, D.-Ding, F.-Zhu, and Z.-Ma “High-efficiency neural video compression via hierarchical predictive learning,” arXiv preprint arXiv:2410. p. 002598, 2024 <ref:2602. p. 009868#pg4>
[33] S.-Zhang, M.-Mrak, L.-Herranz, M.-G. Blanch, S.-Wan, and F.-Yang “Dvc-p: Deep video compression with perceptual optimizations,” in Proc. p. 1 IEEE Vis. Commun. Image Process., IEEE, 2021, pp. p. 0 1–5 <ref:2602. p. 009868#pg4>
[34] R.-Zhang, P.-Isola, A.-Efros, E.-Shechtman, and O.-Wang “The unreasonable effectiveness of deep features as a perceptual metric,” in Proc. p. 1 IEEE/CVF Conf. Comput. Vis. Pattern Recognit., 2018, pp. 586–595 <ref:2602. p. 009868#pg4>
[35] I.-Goodfellow, J.-Pouget-Abadie, M.-Mirza, B.-Xu, D.-Warde-Farley, S. Ozair, A. Courville, and Y. Bengio “Generative adversarial nets,” Adv. p. 0 Neural Inf. Process. Syst., vol. 27, 2014 <ref:2602. p. 009868#pg4>
[36] P.-Esser, R.-Rombach, and B.-Ommer “Taming transformers for highresolution image synthesis,” in Proc. p. 0 IEEE/CVF Conf. Comput. Vis. Pattern Recognit., 2021, pp. 12 873–12 883 <ref:2602. p. 009868#pg4>
[37] W.-Ma and Z.-Chen “Diffvc-osd: One-step diffusion-based perceptual neural video compression framework,” arXiv preprint arXiv:2508. p. 0
Green — the paper says it, and the page is printed. Red — no passage carries it, so treat that sentence as unverified. The same check runs on your own text: check a summary against its source, $5.
Improvements for AI systems
-
Training-free generative video compression framework Free-GVC enables perceptually faithful reconstruction at ultra-low bitrates by reformulating video coding as latent trajectory compression guided by a video diffusion prior.
-
The Adaptive Quality Control module dynamically constructs an online rate–perception surrogate model during encoding to predict the optimal diffusion step for each GOP, maintaining stable perceptual quality across diverse content.
-
The Inter-GOP Alignment mechanism establishes frame overlap and performs latent fusion between adjacent groups, thereby mitigating flicker and enhancing temporal coherence across GOP boundaries.
-
The system can achieve an average of 93.29% BD-Rate reduction in DISTS over the latest neural codec DCVCRT, confirming superior perceptual quality and temporal coherence at ultra-low bitrates.
-
The improved AI system can provide continuous quality scaling by adjusting the termination timestep t star, allowing for flexible bitrate control without retraining or additional overhead.
Sources
- FLUX.1 Kontext: Flow Matching for In-Context Image Generation and Editing in Latent Space
- Lossy Compression with Gaussian Diffusion
- T-GVC: Trajectory-Guided Generative Video Coding at Ultra-Low Bitrates
- High-Efficiency Neural Video Compression via Hierarchical Predictive Learning
- DiffVC-OSD: One-Step Diffusion-based Perceptual Neural Video Compression Framework
- Wan: Open and Advanced Large-Scale Video Generative Models
- CoD: A Diffusion Foundation Model for Image Compression
- Turbo-DDCM: Fast and Flexible Zero-Shot Diffusion-Based Image Compression
Related papers
- TEA: Text Encoder Alignment for Robust Concept Erasure in Text-to-Image Models
- Do Egocentric Video-Language Models Capture Both Hand- and Object-Centric Cues?
- Aura: Consistent Multi-Subject Video Generation via VLM-Grounded Semantic Alignment
- PVRF: All-in-one Adverse Weather Removal via Prior-modulated and Velocity-constrained Rectified Flow
- DLWM: Diverse Latent World Models for Efficient Multimodal Reasoning
- Out of Sight, Still in Mind: Token Compression for Omni-LLMs