DiffCVE: Diffusion-based Compressed Video Enhancement
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: "DiffCVE: Diffusion-based Compressed Video Enhancement".
Jane: Here's the gist of the paper. Diffusion-based compressed video enhancement methods are presented to improve perceptual quality of severely compressed videos by jointly modeling compressed video and coding prior conditions <ref:2607.07195#pg2>.
Tom: The main contributions include Coding Prior-enhanced Dual Conditioning (CPDC) branches, a Compression Degradation Semantic Prompting (CDSP) mechanism, and a Coding Prior-guided Weighted Fusion (CPWF) module <ref:2607.07195#pg4>.
Jane: Pruning redundant reasoning tokens The paper addresses the challenge of applying diffusion models directly to compressed video by introducing CPDC branches that jointly model compressed video and coding prior conditions <ref:2607.07195#pg4>.
Tom: Coding priors, including residuals and motion vectors, provide complementary structural and motion guidance during the diffusion denoising process <ref:2607.07195#pg2>.
Paper summary: Tom: So, looking at "DiffCVE: Diffusion-based Compressed Video Enhancement", this paper by Xiao, Ma, Zhang, and Chen is showing us how to use diffusion models effectively for compressed video restoration <ref:2607.07195#pg5>. It highlights that ignoring the compression degradation characteristics when applying diffusion models directly can lead to structural inconsistencies in the output <ref:2607.07195#pg5>.
Jane: The implication here is that for anyone working on video quality enhancement, especially with highly compressed footage, you can't just use a generic diffusion model blindly <ref:2607.07195#pg5>. You need to incorporate information about the compression itself into the restoration process <ref:2607.07195#pg4>.
Lu: What they propose is that we should jointly model the compressed video and those coding prior conditions, like residuals and motion vectors, during the diffusion denoising process <ref:2607.07195#pg2>. This means we’re using the inherent structure of the compression as a guide for generation <ref:2607.07195#pg4>.
Meng: From a deployment standpoint, I'm still looking at how much overhead these mechanisms add to the system, but it seems like their method provides a solid foundation for better quality under severe compression <ref:2607.07195#pg4>. If we can keep the inference time down, that’s what matters for us in production <ref:2607.07195#pg5>.
Lalam: The paper verifies that this joint leverage of diffusion priors and coding priors is effective for improving the perceptual quality of compressed video <ref:2607.07195#pg4>. It’s a solid piece of evidence showing that incorporating both structural and motion guidance from the compression process helps in reconstruction <ref:2607.07195#pg4>.
Tom: So, to summarize the title "DiffCVE: Diffusion-based Compressed Video Enhancement", it's about using diffusion models specifically tailored for compressed video by adding coding prior-enhanced dual conditioning, semantic prompting based on compression severity, and a weighted fusion module in the VAE decoder <ref:2607.07195#pg4>. It’s a method that uses the specific constraints of the compression to guide the diffusion process to produce better perceptual quality <ref:2607.07195#pg4>.
Jane: It really suggests that for compressed video enhancement, we need methods that go beyond just looking at pixels and start incorporating those structural hints from the coding prior <ref:2607.07195#pg4>. The authors are showing that this approach works well for improving temporal consistency too <ref:2607.07195#pg4>.
Lu: Ultimately, it verifies that jointly leveraging diffusion priors and coding priors is a valid way to improve the perceptual quality of compressed video <ref:2607.07195#pg4>. It's a method that leverages the inherent knowledge embedded in the compression process to achieve better visual results <ref:2607.07195#pg4>.
Conclusion: Tom: So, to wrap up, this paper is about DiffCVE by Xiao and his team <ref:2607.07195#pg5>. They’re using diffusion models to fix bad video compression artifacts by giving them extra information from the coding process <ref:2607.07195#pg4>.
Jane: Right, so the core idea is that you don't just throw a generic diffusion model at a compressed video and hope for the best <ref:2607.07195#pg5>. Instead, they’re adding specific guidance from how the video was compressed <ref:2607.07195#pg4>.
Lu: They introduced three main things to make this happen: first, Coding Prior-enhanced Dual Conditioning which models both the compressed video and those prior conditions together <ref:2607.07195#pg4>. Second, a semantic prompting mechanism called CDSP which uses text prompts based on the compression level to guide the model's behavior <ref:2607.07195#pg5>.
Meng: I’m looking at that dual conditioning idea, and if you can get the model to learn from both what it sees and what the compression process left behind, that sounds like it could handle those really nasty distortions <ref:2607.07195#pg4>. But how much extra compute does all that conditioning add for me at scale?
Tom: It does add complexity, Meng. They show a model called DiffCVE-fifty which runs faster than their baseline SeedVR2-3B, which is good <ref:2607.07195#pg5>. The single step version, DiffCVE-one gets the runtime down to about zero point three two nine seconds per frame <ref:2607.07195#pg5>.
Jane: That speed is a big deal for deployment, Tom. It shows that this method isn't just theoretically interesting; it’s actually efficient enough to be considered for real-world use <ref:2607.07195#pg4>.
Lu: The implication here is that we can start expecting diffusion models to perform better on compressed video, not just on clean images <ref:2607.07195#pg4>. They're showing the effectiveness of combining these different types of priors for both perceptual quality and temporal consistency <ref:2607.07195#pg4>.
Tom: So, the big picture is that this approach proves you can use diffusion models to get much better results on severely compressed video, especially when you care about keeping things looking smooth over time <ref:2607.07195#pg4>.
Jane: It does, but they also flag some limits. They mention that the text prompting part has a certain scope, and they still plan to look at how to fix those issues in future work <ref:2607.07195#pg5>. That’s where the next set of research needs to focus.
School of Remote Sensing and Information Engineering, Wuhan University
cs.CV
Submitted: 2026-07-08
Updated: 2026-11-12
Project page: https://wqmaker.github.io/projects/DiffCVE
License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/
Importance score: 74/100
The gist: Diffusion-based compressed video enhancement methods are presented to improve perceptual quality of severely compressed videos by jointly modeling compressed video and coding prior conditions…
Key concepts
- Coding Prior-enhanced Dual Conditioning (CPDC)
- This involves two branches that model the compressed video alongside coding prior conditions like residuals and motion vectors. They jointly generate conditioning features, allowing the diffusion process to use structural and motion guidance from the compression artifacts during denoising.
- Compression Degradation Semantic Prompting (CDSP)
- CDSP uses QP-conditioned textual prompts with structured degradation descriptions. These prompts are encoded using a CLIP text encoder and injected into the U-Net via cross-attention, helping the diffusion model understand and adapt to specific compression severities.
- Coding Prior-guided Weighted Fusion (CPWF)
- This module in the VAE decoder fuses features from different encoders (VAE, coding prior) using weights predicted by QP. This dynamically modulates feature fusion based on compression severity, ensuring the enhanced output maintains both content and codec-side constraints.
- QP-conditioned Textual Prompts
- These are text prompts specifically designed for different quantization parameter (QP) levels (32, 37, 42). They describe the characteristics of artifacts and overall visual quality at each severity level. These prompts guide the model to generate appropriate enhancement based on the compression level.
Terminology
Summary
Diffusion-based compressed video enhancement methods are presented to improve perceptual quality of severely compressed videos by jointly modeling compressed video and coding prior conditions <ref:2607. p. 007195#pg2>. The main contributions include Coding Prior-enhanced Dual Conditioning (CPDC) branches, a Compression Degradation Semantic Prompting (CDSP) mechanism, and a Coding Prior-guided Weighted Fusion (CPWF) module <ref:2607. p. 007195#pg4>.
Pruning redundant reasoning tokens
The paper addresses the challenge of applying diffusion models directly to compressed video by introducing CPDC branches that jointly model compressed video and coding prior conditions <ref:2607.07195#pg4>. Coding priors, including residuals and motion vectors, provide complementary structural and motion guidance during the diffusion denoising process <ref:2607. p. 007195#pg2>. The CPDC branches consist of a compressed video conditioning branch and a coding prior conditioning branch that jointly generate coding prior conditioning features and coding prior-modulated video conditioning features for diffusion denoising <ref:2607. p. 1 p. 107195#pg4>. The Compressed Video Conditioning Module (CVCM) uses residual blocks and spatial attention layers for content representation, while coding priors are injected through Spatial Feature Transform (SFT) to adaptively recalibrate intermediate visual features <ref:2607. p. 207195#pg4>. This allows codec-side evidence to guide condition generation and compensate for distortion-biased and structurally unreliable responses caused by compression artifacts <ref:2607. p. 207195#pg4>.
How the training signal is built
The CDSP mechanism provides semantic guidance for different compression severities within a unified framework <ref:2607.07195#pg5>. It constructs QP-conditioned textual prompts with structured degradation descriptions and injects the corresponding embeddings into the diffusion U-Net through cross-attention, while LoRA finetuning aligns the pretrained model with degradation semantics <ref:2607. p. 007195#pg4>. For each QP level q ∈ 32, 37, and 42, a textual prompt describes artifact characteristics and overall visual quality <ref:2607. p. 307195#pg5>. The corresponding textual prompt is encoded using a pretrained CLIP text encoder to produce the degradation semantic embedding zqqp <ref:2607. p. 307195#pg4>.
Where the gains come from
The CPWF module is designed in the VAE decoder to improve temporal consistency and local fidelity <ref:2607.07195#pg4>. It uses temporal convolution to capture local inter-frame dynamics, then fuses temporally enhanced decoder features, VAE encoder features, and coding prior features with QP-predicted weights <ref:2607. p. 007195#pg4>. The final enhanced decoder feature is obtained via a weighted combination where the adaptive fusion weight w is predicted from QP using a lightweight MLP <ref:2607. p. 407195#pg4>. This dynamically modulates feature fusion according to compression severity, allowing the decoder to receive both content-layout and codec-side constraints for improving local fidelity and structure consistency <ref:2607. p. 407195#pg4>.
Model complexity comparison
The model complexity comparison shows that DiffCVE-50 achieves faster inference with a smaller parameter scale compared to the generic diffusion-based baseline SeedVR2-3B <ref:2607.07195#pg5>. The single-step variant DiffCVE-1 further reduces the runtime to 0. p. 6329s per frame, showing the efficiency advantage of the proposed design among diffusion-based enhancement methods <ref:2607. p. 607195#pg5>. However, DiffCVE-1 remains slower than lightweight non-diffusion methods such as STDR and MW-GAN+ <ref:2607. p. 607195#pg5>.
Conclusion
DiffCVE achieves strong perceptual quality enhancement, especially under severe compression, while also improving temporal consistency by jointly leveraging diffusion priors and coding priors <ref:2607.07195#pg4>. The proposed method verifies the effectiveness of jointly leveraging diffusion priors and coding priors for perceptual quality enhancement of compressed video <ref:2607. p. 807195#pg4>. However, limitations exist regarding computational requirements and the scope of textual prompting <ref:2607. p. 807195#pg5>. The authors plan to consider how to address these issues in future work <ref:2607. p. 807195#pg5>.
Acknowledgment
The numerical calculations in this paper have been done on the supercomputing system in the Supercomputing Center of Wuhan University <ref:2607.07195#pg5>. The project page with enhanced video demonstrations is available at https://wqmaker. p. 0github.io/projects/DiffCVE/ <ref:2607. p. 007195#pg5>.
References
[1] Thomas Wiegand, Gary J Sullivan, Gisle Bjontegaard, and Ajay Luthra. Overview of the H. p. 8264/AVC video coding standard <ref:2607. p. 807195#pg5>.
[2] Gary J Sullivan, Jens-Rainer Ohm, Woo-Jin Han, and Thomas Wiegand. p. 8 Overview of the high efficiency video coding (HEVC) standard <ref:2607. p. 807195#pg5>.
[3] Benjamin Bross, Ye-Kui Wang, Yan Ye, Shan Liu, Jianle Chen, Gary J Sullivan, and Jens-Rainer Ohm. p. 8 Overview of the versatile video coding (VVC) standard and its applications <ref:2607. p. 807195#pg5>.
[4] Chao Dong, Yubin Deng, Chen Change Loy, and Xiaoou Tang. p. 8 Compression artifacts reduction by a deep convolutional network <ref:2607. p. 807195#pg5>.
[5] Tingting Wang, Mingjin Chen, and Hongyang Chao. p. 8 A novel deep learning-based method of improving coding efficiency from the decoder-end for HEVC <ref:2607. p. 807195#pg5>.
[6] Ren Yang, Mai Xu, Zulin Wang, and Tianyi Li. p. 9 Multiframe quality enhancement for compressed video <ref:2607. p. 907195#pg5>.
[7] Qing Ding, Liquan Shen, Liangwei Yu, Hao Yang, and Mai Xu. p. 9 Blind quality enhancement for compressed video <ref:2607. p. 907195#pg5>.
[8] Nanfeng Jiang, Weiling Chen, Jielian Lin, Tiesong Zhao, and Chia-Wen Lin. p. 9 Video compression artifacts removal with spatial-temporal attention-guided enhancement <ref:2607. p. 907195#pg5>.
[9] Li Yu, Wenshuai Chang, Shiyu Wu, and Moncef Gabbouj. p. 9 End-to-end transformer for compressed video quality enhancement <ref:2607. p. 907195#pg5>.
[10] Xinjian Zhang, Su Yang, Wuyang Luo, Longwen Gao, and Weishan Zhang. p. 9 Video compression artifact reduction by fusing motion compensation and global context in a swin-CNN based parallel architecture <ref:2607. p. 907195#pg5>.
[11] Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. p. 9 Image quality assessment: from error visibility to structural similarity <ref:2607. p. 907195#pg5>.
[12] Jian Yue, Mao Ye, Luping Ji, Hongwei Guo, and Ce Zhu. p. 9 A survey of deep-learning-based compressed video quality enhancement <ref:2607. p. 907195#pg5>.
[13] Jianyi Wang, Xin Deng, Mai Xu, Congyong Chen, and Yuhang Song. p. 9 Multi-level wavelet-based generative adversarial network for perceptual quality enhancement of compressed video <ref:2607. p. 907195#pg5>.
[14] Jianyi Wang, Mai Xu, Xin Deng, Liquan Shen, and Yuhang Song. p. 9 MW-GAN+ for perceptual quality enhancement on compressed video <ref:2607. p. 507195#pg5>.
[15] Lorenzo Agnolucci, Leonardo Galteri, Marco Bertini, and Alberto Del Bimbo. p. 9 Perceptual quality improvement in videoconferencing using keyframes-based gan <ref:2607. p. 907195#pg5>.
[16] Jing Chen, Kemi Chen, Huanqiang Zeng, Qi Lin, and Jianqing Zhu. p. 9 Perceptual quality enhancement for compressed video with high-frequency details and high-dimensional features <ref:2607. p. 907195#pg5>.
[17] Cunhui Dong, Haichuan Ma, Zhuoyuan Li, Li Li, and Dong Liu. p. 9 Temporal wavelet transform-based lowcomplexity perceptual quality enhancement of compressed video <ref:2607. p. 907195#pg5>.
[18] Liuhan Peng, Shuai Li, Yanbo Gao, Mao Ye, and Chong Lv. p. 9 Hierarchical frequency-guided alignment transformer for compressed video quality enhancement <ref:2607. p. 907195#pg5>.
[19] Yi Huang, Jiancheng Huang, Jianzhuang Liu, Mingfu Yan, Yu Dong, Jiaxi Lv, Chaoqi Chen, and Shifeng Chen. p. 9 WaveDM: Wavelet-based diffusion models for image restoration <ref:2607. p. 907195#pg5>.
[20] Xi Yang, Chenhang He, Jianqi Ma, and Lei Zhang. p. 9 Motionguided latent diffusion for temporally consistent real-world video super-resolution <ref:2607. p. 907195#pg5>.
[21] Yuhao Qing, Si Liu, Hai Wang, and Yueying Wang. p. 9 DiffUIE: Learning latent global priors in diffusion models for underwater image enhancement <ref:2607. p. 907195#pg5>.
[22] Xiangchen Yin, Zhenda Yu, Longtao Jiang, Xin Gao, Xiao Sun, Zhi Liu, and Xun Yang. p. 9 Structure-guided diffusion transformer for low-light image enhancement <ref:2607. p. 907195#pg5>.
[23] Jianyi Wang, Zhijie Lin, Meng Wei, Yang Zhao, Ceyuan Yang, Chen Change Loy, and Lu Jiang. p. 9 SeedVR: Seeding infinity in diffusion transformer towards generic video restoration <ref:2607. p. 907195#pg5>.
[24] Peilin Chen, Wenhan Yang, Meng Wang, Long Sun, Kangkang Hu, and Shiqi Wang. p. 9 Compressed domain deep video super-resolution <ref:2607. p. 907195#pg5>.
[25] Hengsheng Zhang, Xueyi Zou, Jiaming Guo, Youliang Yan, Rong Xie, and Li Song. p. 9 A codec information assisted framework for efficient compressed video super-resolution <ref:2607. p. 907195#pg5>.
[26] Qiang Zhu, Jinhua Hao, Yukang Ding, Yu Liu, Qiao Mo, Ming Sun, Chao Zhou, and Shuyuan Zhu. p. 9 CPGA: Coding priors-guided aggregation network for compressed video quality enhancement <ref:2607. p. 907195#pg5>.
[27] Zhenyu Guan, Qunliang Xing, Mai Xu, Ren Yang, Tie Liu, and Zulin Wang. p. 9 MFQE 2.0: A new approach for multiframe quality enhancement on compressed video <ref:2607. p. 907195#pg5>.
[28] Dengyan Luo, Mao Ye, Shuai Li, Ce Zhu, and Xue Li. p. 9 Spatio-temporal detail information retrieval for compressed video quality enhancement <ref:2607. p. 907195#pg5>.
[29] Mingxing Wang, Yipeng Liao, Weiling Chen, Liqun Lin, and Tiesong Zhao. p. 9 STFF: Spatio-temporal and frequency fusion for video compression artifact removal <ref:2607. p. 907195#pg5>.
[30] Chitwan Saharia, Jonathan Ho, William Chan, Tim Salimans, David J Fleet, and Mohammad Norouzi. p. 9 Image super-resolution via iterative refinement <ref:2607. p. 907195#pg5>.
[31] Haoying Li, Yifan Yang, Meng Chang, Shiqi Chen, Huajun Feng, Zhihai Xu, Qi Li, and Yueting Chen. p. 9 SRDiff: Single image super-resolution with diffusion probabilistic models <ref:2607. p. 107195#pg5>.
[32] Jianyi Wang, Zongsheng Yue, Shangchen Zhou, Kelvin CK Chan, and Chen Change Loy. p. 9 Exploiting diffusion prior for real-world image super-resolution <ref:2607. p. 107195#pg5>.
[33] Rongyuan Wu, Tao Yang, Lingchen Sun, Zhengqiang Zhang, Shuai Li, and Lei Zhang. p. 9 SeeSR: Towards semantics-aware real-world image super-resolution <ref:2607. p. 107195#pg5>.
[34] Jianyi Wang, Shanchuan Lin, Zhijie Lin, Yuxi Ren, Meng Wei, Zongsheng Yue, Shangchen Zhou, Hao Chen, Yang Zhao, Ceyuan Yang, Xuefeng Xiao, Chen Change Loy, and Lu Jiang. p. 9 SeedVR2: One-step video restoration via diffusion adversarial post-training <ref:2607. p. 107195#pg5>.
[35] Simon Welker, Henry N Chapman, and Timo Gerkmann. DriftRec: Adapting diffusion models to blind jpeg restoration <ref:2607. p. 107195#pg5>.
[36] Hongyu An, Xinfeng Zhang, Shijie Zhao, Li Zhang, and Ruiqin Xiong. p. 9 Spatial degradation-aware and temporal consistent diffusion model for compressed video superresolution <ref:2607. p. 007195#pg5>.
[37] Shiv Gehlot and Guan-Ming Su. p. 9 LViCAR: Diffusion models for perceptual quality enhancement in video compression artifact reduction <ref:2607. p. 107195#pg5>.
[38] Xintao Wang, Ke Yu, Chao Dong, and Chen Change Loy. p. 8 Recovering realistic texture in image super-resolution by deep spatial feature transform <ref:2607. p. 107195#pg5>.
[39] Aaditya Singh, Adam Fry, Adam Perelman, Adam Tart, Adi Ganesh, Ahmed El-Kishky, Aidan McLaughlin, Aiden Low, AJ Ostrow, Akhila Ananthram, et al. OpenAI GPT-5 system card <ref:2607. p. 307195#pg5>.
[40] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision <ref:2607. p. 307195#pg5>.
[41] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan AllenZhu, Yuanzhi Li, Shean Wang, Liang Wang, Weizhu Chen, et al. p. 9 LoRA: Low-rank adaptation of large language models <ref:2607. p. 307195#pg5>.
[42] Xintao Wang, Ke Yu, Shixiang Wu, Jinjin Gu, Yihao Liu, Chao Dong, Yu Qiao, and Chen Change Loy. p. 8 ESRGAN: Enhanced super-resolution generative adversarial networks <ref:2607. p. 107195#pg5>.
[43] Tianfan Xue, Baian Chen, Jiajun Wu, Donglai Wei, and William T Freeman. p. 9 Video enhancement with task-oriented flow <ref:2607. p. 407195#pg5>.
[44] Jens-Rainer Ohm, Gary J Sullivan, Heiko Schwarz, Thiow Keng Tan, and Thomas Wiegand. p. 8 Comparison of the coding efficiency of video coding standards—including high efficiency video coding (HEVC) <ref:2607. p. 807195#pg5>.
[45] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. p. 9 The unreasonable effectiveness of deep features as a perceptual metric <ref:2607. p. 007195#pg5>.
[46] Keyan Ding, Kede Ma, Shiqi Wang, and Eero P Simoncelli. p. 9 Image quality assessment: Unifying structure and texture similarity <ref:2607. p. 907195#pg5>.
[47] Junjie Ke, Qifei Wang, Yilin Wang, Peyman Milanfar, and Feng Yang. p. 9 MUSIQ: Multi-scale image quality transformer <ref:2607. p. 307195#pg5>.
[48] Jianyi Wang, Kelvin CK Chan, and Chen Change Loy. p. 9 Exploring CLIP for assessing the look and feel of images <ref:2607. p. 007195#pg5>.
[49] Haoning Wu, Erli Zhang, Liang Liao, Chaofeng Chen, Jingwen Hou, Annan Wang, Wenxiu Sun, Qiong Yan, and Weisi Lin. p. 9 Exploring video quality assessment on user generated contents from aesthetic and technical perspectives <ref:2607. p. 907195#pg5>.
[50] Anish Mittal, Rajiv Soundararajan, and Alan C Bovik. p. 9 Making a ’completely blind’ image quality analyzer <ref:2607. p. 307195#pg5>.
[51] Mengyu Chu, You Xie, Jonas Mayer, Laura Leal-Taixé, and Nils Thuerey. p. 9 Learning temporal coherence via selfsupervision for GAN-based video generation <ref:2607. p. 107195#pg5>.
--- Page 1 ---
DiffCVE: Diffusion-based Compressed Video Enhancement
Wenqiang Xiao, Wenzhuo Ma, Junxi Zhang, and Zhenzhong Chen <ref:2607.07195#pg5>.
School of Remote Sensing and Information Engineering, Wuhan University <ref:2607. p. 007195#pg5>.
Abstract
Perceptual quality enhancement of severely compressed videos remains challenging due to complex artifact patterns and substantial information loss <ref:2607.07195#pg5>. Recent diffusion models have demonstrated strong generative capability for visual restoration, but directly applying them to compressed video often ignores compression degradation characteristics and may introduce structure-inconsistent hallucinations <ref:2607. p. 0 p. 007195#pg5>. To address this issue, this paper presents a diffusion-based compressed video enhancement method, named DiffCVE <ref:2607. p. 0 p. 007195#pg5>. Coding Prior-enhanced Dual Conditioning (CPDC) branches are designed to jointly model compressed video and coding prior conditions <ref:2607. p. 007195#pg5>. To make the diffusion process aware of compression severity, a Compression Degradation Semantic Prompting (CDSP) mechanism is introduced to leverage QP-conditioned textual prompts together with LoRA fine-tuning <ref:2607. p. 0 p. 007195#pg5>. In addition, a Coding Prior-guided Weighted Fusion (CPWF) module is incorporated into the VAE decoder to fuse VAE encoder and coding prior encoder features with QPpredicted weights <ref:2607. p. 007195#pg5>.
Perceptual quality enhancement of severely compressed videos remains challenging due to complex artifact patterns and substantial information loss <ref:2607.07195#pg5>. Recent diffusion models have demonstrated strong generative capability for visual restoration, but directly applying them to compressed video often ignores compression degradation characteristics and may introduce structure-inconsistent hallucinations <ref:2607.07195#pg5>. To address this issue, this paper presents a diffusion-based compressed video enhancement method, named DiffCVE <ref:2607.07195#pg5>.
Coding Prior-enhanced Dual Conditioning (CPDC) branches are designed to jointly model compressed video and coding prior conditions <ref:2607.07195#pg5>. The CPDC branches consist of a compressed video conditioning branch and a coding prior conditioning branch that jointly generate coding prior conditioning features and coding prior-modulated video conditioning features for diffusion denoising <ref:2607.07195#pg4>. To make the diffusion process aware of compression severity, a Compression Degradation Semantic Prompting (CDSP) mechanism is introduced to leverage QP-conditioned textual prompts together with LoRA fine-tuning <ref:2607.07195#pg5>.
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
-
Coding Prior-enhanced Dual Conditioning (CPDC) branches allow for joint modeling of compressed frames and coding priors by encoding residuals and motion vectors into multi-level features to modulate video conditioning features from compressed frames. This enables
codec-side spatial discrepancy and motion cues to guide denoising while preserving compressed-frame content constraints.
-
A Compression Degradation Semantic Prompting (CDSP) mechanism constructs
QP-conditioned textual prompts with structured degradation descriptions
and uses LoRA fine-tuning to align the diffusion model with these degradation semantics. This providesQP-dependent semantic guidance about compression severity and artifact characteristics, enabling a unified diffusion model to adjust its restoration behavior across different bitrate levels.
-
The Coding Prior-guided Weighted Fusion (CPWF) module in the VAE decoder uses temporal convolution to capture local inter-frame dynamics and fuses
temporally enhanced decoder features, VAE encoder features, and coding prior features with QP-predicted weights.
This allows the system todynamically modulate feature fusion according to compression severity,
leading to a more reliable reconstruction. -
The improved AI system can achieve strong perceptual quality enhancement, especially under severe compression settings, while also improving temporal consistency across different bitrate levels.
Sources
- Spatial Degradation-Aware and Temporal Consistent Diffusion Model for Compressed Video Super-Resolution
- OpenAI GPT-5 System Card
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