Pith. sign in

REVIEW 3 major objections 6 minor 51 references

DiffCVE: Diffusion-based Compressed Video Enhancement

T0 review · 3 major / 6 minor · reviewed 2026-07-09 · glm-5.2

Pith's one-line read Codec internals guide diffusion to restore heavily compressed video

desk verdict Diffusion + codec priors for compressed video: solid engineering, modest perceptual gains over GANs at much higher cost read the letter →

arxiv 2607.07195 v1 pith:KUVLVCZU submitted 2026-07-08 cs.CV

classification cs.CV
keywords compressedvideoenhancementdiffusionmodelscodingpriorsresidualsmotionvectorsquantizationparameterperceptualqualitytemporalconsistency
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

The paper argues that diffusion models, when applied naively to compressed video enhancement, fail because they treat compression as generic degradation and lack structural constraints, producing hallucinated details that do not match the original content. The authors propose DiffCVE, which injects three categories of codec-internal information into the diffusion pipeline: residuals and motion vectors as structural and motion guidance during denoising, QP-conditioned textual prompts that describe the specific artifact characteristics at each compression level, and decoder-side fusion of coding prior features with QP-predicted adaptive weights. The central claim is that these coding priors, which are freely available at decode time but previously underused in generative enhancement, constrain the diffusion model to produce perceptually faithful restorations rather than plausible but structurally inconsistent fabrications. Experiments on standard test sequences show that the method achieves the best or near-best perceptual quality metrics (LPIPS, DISTS) and temporal consistency under heavy compression (QP 42, 37), while remaining competitive at milder compression levels.

What carries the argument

The method has three coupled mechanisms. First, Coding Prior-enhanced Dual Conditioning (CPDC) encodes residuals and motion vectors into a latent that both modulates the compressed video conditioning branch via Spatial Feature Transform and is injected directly into the diffusion U-Net as an independent condition. Second, Compression Degradation Semantic Prompting (CDSP) maps each quantization level to a structured textual description of its artifact characteristics, encodes these via a CLIP text encoder, and injects them through cross-attention layers, with LoRA fine-tuning aligning the pretrained model to degradation semantics. Third, Coding Prior-guided Weighted Fusion (CPWF) in the VAE解码

What would settle it

If one were to evaluate the method on quantization levels not covered by the three prompts (e.g., QP 35 or 40), the degradation semantic guidance would be absent or mismatched, and the model would have to fall back on the nearest available prompt, potentially degrading performance relative to a model trained specifically for that level.

Watch

Extended reading notes

Core claim

Codec-internal signals—residuals, motion vectors, and quantization parameters—can serve as effective conditioning and fusion signals for diffusion-based video enhancement, providing the structural and motion constraints that prevent generative hallucination and enable compression-severity-aware restoration within a single unified model.

Load-bearing premise

The approach assumes that three discrete textual prompts, each describing the artifact characteristics at a specific quantization level, are sufficient to guide a single diffusion model across the full range of compression severities encountered in practice. Intermediate or unseen quantization levels may not be adequately represented by this coarse discrete mapping.

Editorial extensions

If this is right

  • Codec-internal metadata streams (residuals, motion vectors, QP) that are currently discarded after decoding could become standard auxiliary inputs for a broad class of generative video processing tasks including super-resolution, denoising, and frame interpolation.
  • Textual degradation descriptions as a conditioning modality could generalize to other distortion types with continuous severity scales, such as network packet loss or bandwidth-adaptive streaming artifacts, provided that sufficiently granular prompt mappings are constructed.
  • The QP-predicted adaptive fusion weight mechanism suggests that optimal feature fusion strategies vary systematically with compression severity, which could inform adaptive architectures in other degradation-aware restoration pipelines.

Reading between the lines

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

  • The discrete QP-to-prompt mapping (three levels) likely limits generalization to intermediate or out-of-distribution compression settings; a continuous or interpolated prompt embedding space might extend coverage without requiring per-level retraining.
  • The approach is codec-specific in that residuals and motion vectors are defined by the coding standard; extending it to end-to-end learned or neural codecs would require redefining what constitutes a coding prior, potentially using learned latent-space residuals instead.
  • The decoder-side fusion of coding prior features suggests that generative models for restoration may benefit from late-stage injection of task-specific auxiliary signals more broadly, not only for compression artifacts.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

3 major / 6 minor

Summary. This paper proposes DiffCVE, a diffusion-based method for perceptual quality enhancement of compressed video. The approach introduces three components: (1) Coding Prior-enhanced Dual Conditioning (CPDC), which injects codec-internal residuals and motion vectors into the diffusion U-Net via SFT-based modulation; (2) Compression Degradation Semantic Prompting (CDSP), which uses QP-conditioned textual prompts and LoRA fine-tuning to make the model aware of compression severity; and (3) Coding Prior-guided Weighted Fusion (CPWF), which fuses VAE encoder, coding prior encoder, and temporally enhanced decoder features with QP-predicted weights in the VAE decoder. Experiments on 18 JCT-VC sequences under H.264 compression at QP 32, 37, and 42 show that DiffCVE achieves competitive or superior perceptual metrics (LPIPS, DISTS, no-reference metrics) and temporal consistency (tOF) compared to five baselines, with the most pronounced gains at QP 42. The ablation study in Table 6 isolates the contributions of CDSP, CPDC, and CPWF. The central claim—that codec-internal priors and degradation semantics improve diffusion-based compressed video enhancement—is well-motivated and supported by the experimental design.

Significance. The paper addresses a relevant problem: perceptual enhancement of heavily compressed video, where diffusion models risk structure-inconsistent hallucinations. The idea of leveraging codec-internal information (residuals, MVs, QP) as structural and severity-aware constraints for diffusion denoising is a reasonable and novel contribution to the compressed-video enhancement literature. The ablation study (Table 6) provides internal validation for each component, and the inclusion of both 50-step and single-step variants demonstrates awareness of the efficiency concern. The project page with video demonstrations is a positive step toward reproducibility. However, the practical significance of the perceptual gains over simpler GAN baselines, given the substantial computational cost, requires clearer justification (see Major Comment 1).

major comments (3)
  1. Section 4.2, Table 2: The cost-benefit tradeoff between DiffCVE and the GAN-based baseline MW-GAN+ is not adequately addressed. At QP 42, DiffCVE-50 improves average ΔLPIPS to -0.128 vs. MW-GAN+'s -0.099 (a gain of 0.029) and ΔDISTS to -0.085 vs. -0.067 (a gain of 0.018), but degrades ΔPSNR to -0.370 vs. -0.352 (a loss of 0.018 dB). Meanwhile, the computational cost is orders of magnitude higher: 2.955s vs. 0.080s per frame and 1.6B vs. 6.41M parameters (Table 5). The no-reference metric gains are larger (e.g., ΔMUSIQ 17.402 vs. 9.011), but no-reference metrics can be inflated by generative models producing visually pleasing but not necessarily faithful details—a risk the paper itself acknowledges in the Introduction. The paper should explicitly discuss whether the modest perceptual gains over the much simpler GAN baseline justify the complexity, and ideally provide a faithful-detail vs.
  2. Section 4.1: The evaluation is limited to H.264/AVC compression on Vimeo-90k training data. The paper's title and abstract refer to 'compressed video enhancement' broadly, and the Introduction mentions H.264, H.265/HEVC, and H.266/VVC. However, no experiments are conducted with HEVC or VVC codecs, and the coding priors (residuals, MVs) may have different characteristics under these standards. The paper should either narrow its claims to H.264 or provide at least a preliminary cross-codec evaluation to support the generality of the approach. This is load-bearing for the paper's positioning as a general compressed video enhancement method.
  3. Section 4.2, Table 3: The large gains on no-reference metrics (MUSIQ, CLIPIQA, DOVER) relative to the more modest gains on full-reference perceptual metrics (LPIPS, DISTS) raise a concern about whether the diffusion model is generating visually pleasing details that may not correspond to the original content. The paper acknowledges this risk in the Introduction ('structure-inconsistent hallucinations') but does not provide a quantitative analysis of fidelity vs. perceptual quality tradeoff. A simple test—e.g., reporting the percentage of frames where DiffCVE improves LPIPS but worsens PSNR by more than a threshold, compared to baselines—would help clarify whether the gains reflect genuine restoration or generative hallucination.
minor comments (6)
  1. Table 1: The textual prompts are generated using 'OpenAI GPT-5 model family' (reference [39]). The reference appears to be a system card, not a model description. It would be helpful to clarify which specific model version was used and whether the prompts were manually curated or used verbatim from the LLM output.
  2. Section 3.2: The QP-conditioned switch (Eq. 7) selects from three precomputed embeddings for QP 32, 37, 42. For QP values outside this set (e.g., QP 35), the mechanism is undefined. The paper acknowledges this in Section 5, but a brief note in Section 3.2 clarifying that the current design is limited to these three levels would improve clarity.
  3. Section 3.4, Eq. (12): The MLP predicting the fusion weight w takes QP as input. It is unclear whether QP is normalized before input, and what the MLP architecture is (number of layers, hidden dimensions). These details would aid reproducibility.
  4. Table 2: The ΔPSNR and ΔSSIM values for DiffCVE-50 and DiffCVE-1 are negative at QP 42 and 37, indicating that the enhanced results have lower PSNR/SSIM than the compressed input. While this is expected for perceptual-oriented methods, the paper does not explicitly acknowledge or explain this in the text discussion. A brief statement noting this tradeoff would be helpful.
  5. Section 4.1: The training data is generated via H.264 compression under 'Low Delay P at multiple bitrates.' The specific bitrate points or the mapping from QP to bitrate are not specified. Including this information would improve reproducibility.
  6. Figure 1: The architecture diagram is dense and some labels are difficult to read (e.g., the CPWF module internals, the flow of coding prior features). A higher-resolution or simplified version with clearer data flow arrows would improve readability.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity found; derivation chain is self-contained and validated against external benchmarks

full rationale

The paper proposes three architectural components (CPDC, CDSP, CPWF) for a diffusion-based compressed video enhancement model. Walking the derivation chain: (1) CPDC extracts coding priors (residuals R, motion vectors V) from the codec bitstream via a learned encoder E_p(R,V) → z_p, then uses SFT modulation to inject these into video conditioning features. The priors come from the input bitstream, not from the model's output. (2) CDSP constructs QP-conditioned textual prompts offline (Table 1), encodes them with a pretrained CLIP text encoder, and selects the appropriate embedding at inference via a QP-conditioned switch. No step defines the prompt in terms of the enhancement result. (3) CPWF fuses decoder features with VAE encoder features and coding prior encoder features using a weight w = σ(MLP(QP)) predicted from the quantization parameter, not from the output. The denoising equation ε̂ = ε_θ(z_t, t, {f̃_v^l}, {f_pu^l}, z_qp) is a standard conditional diffusion formulation. The ablation study (Table 6) incrementally adds components and evaluates each on external JCT-VC sequences using standard metrics (PSNR, SSIM, LPIPS, DISTS, MUSIQ, CLIPIQA, DOVER, NIQE, tOF) computed against ground truth. No cited work is authored by the present paper's authors (Wenqiang Xiao, Wenzhuo Ma, Junxi Zhang, Zhenzhong Chen), so there are no self-citations, load-bearing or otherwise. No equation reduces to its inputs by construction, and no 'prediction' is a renamed fit. The paper is a standard empirical deep learning contribution with no circular reasoning in its derivation chain.

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

The ledger captures the key design choices and parameters that define the method. The free parameters are standard for such architectures. The axioms represent the foundational assumptions, with the discrete QP representation being the most ad-hoc. The invented entities are the core contributions, each with ablation-based evidence.

free parameters (3)
  • LoRA rank
    The rank of the LoRA adaptation is a free parameter chosen for fine-tuning the U-Net.
  • MLP weights for QP prediction
    The weights of the lightweight MLP used to predict the fusion weight w in the CPWF module are learned during training.
  • Textual prompts for QP levels = See Table 1
    The specific textual descriptions for QP 32, 37, and 42 are manually designed/generated and serve as fixed conditions.
assumptions (3)
  • domain assumption Stable Diffusion v2.1 provides a strong generative prior for video frame restoration.
    The entire method is built upon adapting a pretrained Stable Diffusion model, assuming its latent space and generative capabilities are suitable for this task.
  • domain assumption Codec-internal residuals and motion vectors provide complementary and useful structural and motion guidance for restoration.
    This is the core motivation for the CPDC and CPWF modules, assuming these signals are beneficial beyond the compressed frame itself.
  • ad hoc to paper Discrete QP levels (32, 37, 42) and their corresponding text prompts are sufficient to represent compression degradation semantics.
    The CDSP mechanism relies on this discretization to provide degradation guidance within a unified model.
invented entities (3)
  • Coding Prior-enhanced Dual Conditioning (CPDC) branches independent evidence
    purpose: To jointly model compressed video and coding prior conditions for diffusion denoising.
    The module's effectiveness is supported by the ablation study in Table 6 (models C, D, E, F).
  • Compression Degradation Semantic Prompting (CDSP) mechanism independent evidence
    purpose: To provide semantic guidance for different compression severities using QP-conditioned text prompts.
    Effectiveness is shown in Table 6 (models A, B, C) and Figure 4.
  • Coding Prior-guided Weighted Fusion (CPWF) module independent evidence
    purpose: To fuse VAE encoder and coding prior encoder features with QP-predicted weights in the decoder.
    Effectiveness is supported by the ablation study in Table 6 (models F, G, H).

how reviews work

0 comments
Cite this review

Pith. "Pith review of DiffCVE: Diffusion-based Compressed Video Enhancement." pith.science (2026). https://pith.science/paper/KUVLVCZU

@misc{pith2026260707195,
  author       = {Pith},
  title        = {Pith review of: DiffCVE: Diffusion-based Compressed Video Enhancement},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/KUVLVCZU}},
  note         = {Machine review of arXiv:2607.07195}
}
read the original abstract

Perceptual quality enhancement of severely compressed videos remains challenging due to complex artifact patterns and substantial information loss. 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. To address this issue, this paper presents a diffusion-based compressed video enhancement method, named DiffCVE. Coding Prior-enhanced Dual Conditioning (CPDC) branches are designed to jointly model compressed video and coding prior conditions, where coding priors including residuals and motion vectors provide complementary structural and motion guidance during the diffusion denoising process. 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. 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 QP-predicted weights. Extensive experiments demonstrate the effectiveness of the proposed method in improving perceptual quality, especially under severe compression settings. The project page with enhanced video demonstrations is available at https://wqmaker.github.io/projects/DiffCVE/.

Figures

Figures reproduced from arXiv: 2607.07195 by the authors.

Figure 1
Figure 1. The architecture of the proposed DiffCVE framework. Coding priors (e.g., residuals and motion vectors) are incorporated to provide structure- and motion-aware guidance, while a large language model is used to generate QP-conditioned textual prompts for degradation semantics, which are further encoded as semantic embeddings to guide the diffusion process. During decoding, coding prior-guided weighted fusion is furthe… view at source ↗
Figure 2
Figure 2. Subjective visual comparison under QP 42. All examples correspond to the 10th frame of each sequence. The visual [PITH_FULL_IMAGE:figures/full_fig_p006_2.png] view at source ↗
Figure 3
Figure 3. Temporal consistency visualization under QP 42. Consecutive frames of each method are aligned in the same row for [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗
Figures from the paper (2 more)
Figure 4
Figure 4. Figure 4: Cross-attention visualization of CDSP. From left to [PITH_FULL_IMAGE:figures/full_fig_p008_4.png]
Figure 5
Figure 5. Figure 5: Visualization of the effect of coding priors on perceptual quality enhancement and structure consistency. The first and second columns show the residual and motion-vector priors, respectively. The third column is the compressed frame. The fourth column corresponds to t…

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

51 extracted references · 51 canonical work pages

  1. [1]

    Overview of the H.264/A VC video coding standard.IEEE Transactions on Circuits and Systems for Video Technology, 13(7):560–576, 2003

    Thomas Wiegand, Gary J Sullivan, Gisle Bjontegaard, and Ajay Luthra. Overview of the H.264/A VC video coding standard.IEEE Transactions on Circuits and Systems for Video Technology, 13(7):560–576, 2003

  2. [2]

    Overview of the high efficiency video coding (HEVC) standard.IEEE Transactions on Circuits and Systems for Video Technology, 22(12):1649–1668, 2012

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

  3. [3]

    Overview of the versatile video coding (VVC) standard and its appli- cations.IEEE Transactions on Circuits and Systems for Video Technology, 31(10):3736–3764, 2021

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

  4. [4]

    Compression artifacts reduction by a deep convo- lutional network

    Chao Dong, Yubin Deng, Chen Change Loy, and Xiaoou Tang. Compression artifacts reduction by a deep convo- lutional network. InProceedings of the IEEE/CVF Inter- national Conference on Computer Vision, pages 576–584, 2015

  5. [5]

    A novel deep learning-based method of improving coding9 efficiency from the decoder-end for HEVC

    Tingting Wang, Mingjin Chen, and Hongyang Chao. A novel deep learning-based method of improving coding9 efficiency from the decoder-end for HEVC. InData Com- pression Conference, pages 410–419, 2017

  6. [6]

    Multi- frame quality enhancement for compressed video

    Ren Yang, Mai Xu, Zulin Wang, and Tianyi Li. Multi- frame quality enhancement for compressed video. InPro- ceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6664–6673, 2018

  7. [7]

    Blind quality enhancement for compressed video

    Qing Ding, Liquan Shen, Liangwei Yu, Hao Yang, and Mai Xu. Blind quality enhancement for compressed video. IEEE Transactions on Multimedia, 26:5782–5794, 2023

  8. [8]

    Video compression artifacts removal with spatial-temporal attention-guided enhancement.IEEE Transactions on Multimedia, 26:5657–5669, 2023

    Nanfeng Jiang, Weiling Chen, Jielian Lin, Tiesong Zhao, and Chia-Wen Lin. Video compression artifacts removal with spatial-temporal attention-guided enhancement.IEEE Transactions on Multimedia, 26:5657–5669, 2023

Show all 51 references
  1. [9]

    End-to-end transformer for compressed video quality en- hancement.IEEE Transactions on Broadcasting, 70(1): 197–207, 2023

    Li Yu, Wenshuai Chang, Shiyu Wu, and Moncef Gabbouj. End-to-end transformer for compressed video quality en- hancement.IEEE Transactions on Broadcasting, 70(1): 197–207, 2023

  2. [10]

    Video compression artifact reduction by fusing motion compensation and global context in a swin-CNN based parallel architecture

    Xinjian Zhang, Su Yang, Wuyang Luo, Longwen Gao, and Weishan Zhang. Video compression artifact reduction by fusing motion compensation and global context in a swin-CNN based parallel architecture. InProceedings of the AAAI Conference on Artificial Intelligence, volume 37, page...

  3. [11]

    Image quality assessment: from error visibil- ity to structural similarity.IEEE Transactions on Image Processing, 13(4):600–612, 2004

    Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image quality assessment: from error visibil- ity to structural similarity.IEEE Transactions on Image Processing, 13(4):600–612, 2004

  4. [12]

    A survey of deep-learning-based compressed video quality enhancement.IEEE Transactions on Broadcasting, pages 977–992, 2025

    Jian Yue, Mao Ye, Luping Ji, Hongwei Guo, and Ce Zhu. A survey of deep-learning-based compressed video quality enhancement.IEEE Transactions on Broadcasting, pages 977–992, 2025

  5. [13]

    Multi-level wavelet-based generative ad- versarial network for perceptual quality enhancement of compressed video

    Jianyi Wang, Xin Deng, Mai Xu, Congyong Chen, and Yuhang Song. Multi-level wavelet-based generative ad- versarial network for perceptual quality enhancement of compressed video. InEuropean Conference on Computer Vision, pages 405–421, 2020

  6. [14]

    MW-GAN+for perceptual quality enhancement on compressed video.IEEE Transactions on Circuits and Systems for Video Technology, 32(7):4224–4237, 2021

    Jianyi Wang, Mai Xu, Xin Deng, Liquan Shen, and Yuhang Song. MW-GAN+for perceptual quality enhancement on compressed video.IEEE Transactions on Circuits and Systems for Video Technology, 32(7):4224–4237, 2021

  7. [15]

    Perceptual quality improvement in videoconferencing using keyframes-based gan.IEEE Transactions on Multimedia, 26:339–352, 2023

    Lorenzo Agnolucci, Leonardo Galteri, Marco Bertini, and Alberto Del Bimbo. Perceptual quality improvement in videoconferencing using keyframes-based gan.IEEE Transactions on Multimedia, 26:339–352, 2023

  8. [16]

    Perceptual quality enhancement for compressed video with high-frequency details and high-dimensional features.IEEE Transactions on Instrumentation and Mea- surement, 74:1–13, 2025

    Jing Chen, Kemi Chen, Huanqiang Zeng, Qi Lin, and Jian- qing Zhu. Perceptual quality enhancement for compressed video with high-frequency details and high-dimensional features.IEEE Transactions on Instrumentation and Mea- surement, 74:1–13, 2025

  9. [17]

    Cunhui Dong, Haichuan Ma, Zhuoyuan Li, Li Li, and Dong Liu. Temporal wavelet transform-based low- complexity perceptual quality enhancement of compressed video.IEEE Transactions on Circuits and Systems for Video Technology, 34(5):4040–4053, 2023

  10. [18]

    Hierarchical frequency-guided alignment transformer for compressed video quality enhancement

    Liuhan Peng, Shuai Li, Yanbo Gao, Mao Ye, and Chong Lv. Hierarchical frequency-guided alignment transformer for compressed video quality enhancement. InProceedings of the AAAI Conference on Artificial Intelligence, volume 40, pages 8349–8357, 2026

  11. [19]

    WaveDM: Wavelet-based diffusion models for im- age restoration.IEEE Transactions on Multimedia, 26: 7058–7073, 2024

    Yi Huang, Jiancheng Huang, Jianzhuang Liu, Mingfu Yan, Yu Dong, Jiaxi Lv, Chaoqi Chen, and Shifeng Chen. WaveDM: Wavelet-based diffusion models for im- age restoration.IEEE Transactions on Multimedia, 26: 7058–7073, 2024

  12. [20]

    Motion- guided latent diffusion for temporally consistent real-world video super-resolution

    Xi Yang, Chenhang He, Jianqi Ma, and Lei Zhang. Motion- guided latent diffusion for temporally consistent real-world video super-resolution. InEuropean Conference on Com- puter Vision, pages 224–242, 2024

  13. [21]

    Dif- fUIE: Learning latent global priors in diffusion models for underwater image enhancement.IEEE Transactions on Multimedia, 27:2516–2529, 2024

    Yuhao Qing, Si Liu, Hai Wang, and Yueying Wang. Dif- fUIE: Learning latent global priors in diffusion models for underwater image enhancement.IEEE Transactions on Multimedia, 27:2516–2529, 2024

  14. [22]

    Structure-guided diffusion transformer for low-light image enhancement.IEEE Trans- actions on Multimedia, 27:9505–9515, 2025

    Xiangchen Yin, Zhenda Yu, Longtao Jiang, Xin Gao, Xiao Sun, Zhi Liu, and Xun Yang. Structure-guided diffusion transformer for low-light image enhancement.IEEE Trans- actions on Multimedia, 27:9505–9515, 2025

  15. [23]

    SeedVR: Seed- ing infinity in diffusion transformer towards generic video restoration

    Jianyi Wang, Zhijie Lin, Meng Wei, Yang Zhao, Ceyuan Yang, Chen Change Loy, and Lu Jiang. SeedVR: Seed- ing infinity in diffusion transformer towards generic video restoration. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2161– 2172, 2025

  16. [24]

    Compressed domain deep video super-resolution.IEEE Transactions on Image Pro- cessing, 30:7156–7169, 2021

    Peilin Chen, Wenhan Yang, Meng Wang, Long Sun, Kangkang Hu, and Shiqi Wang. Compressed domain deep video super-resolution.IEEE Transactions on Image Pro- cessing, 30:7156–7169, 2021

  17. [25]

    A codec information assisted framework for efficient compressed video super-resolution

    Hengsheng Zhang, Xueyi Zou, Jiaming Guo, Youliang Yan, Rong Xie, and Li Song. A codec information assisted framework for efficient compressed video super-resolution. InEuropean Conference on Computer Vision, pages 220– 235, 2022

  18. [26]

    CPGA: Coding priors-guided aggregation network for compressed video quality enhancement

    Qiang Zhu, Jinhua Hao, Yukang Ding, Yu Liu, Qiao Mo, Ming Sun, Chao Zhou, and Shuyuan Zhu. CPGA: Coding priors-guided aggregation network for compressed video quality enhancement. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2964–2...

  19. [27]

    MFQE 2.0: A new approach for multi- frame quality enhancement on compressed video.IEEE Transactions on Pattern Analysis and Machine Intelligence, 43(3):949–963, 2019

    Zhenyu Guan, Qunliang Xing, Mai Xu, Ren Yang, Tie Liu, and Zulin Wang. MFQE 2.0: A new approach for multi- frame quality enhancement on compressed video.IEEE Transactions on Pattern Analysis and Machine Intelligence, 43(3):949–963, 2019

  20. [28]

    Spatio-temporal detail information retrieval for com- pressed video quality enhancement.IEEE Transactions on Multimedia, 25:6808–6820, 2023

    Dengyan Luo, Mao Ye, Shuai Li, Ce Zhu, and Xue Li. Spatio-temporal detail information retrieval for com- pressed video quality enhancement.IEEE Transactions on Multimedia, 25:6808–6820, 2023

  21. [29]

    STFF: Spatio-temporal and frequency fusion for video compression artifact removal.IEEE Trans- actions on Broadcasting, 71(2):542–554, 2025

    Mingxing Wang, Yipeng Liao, Weiling Chen, Liqun Lin, and Tiesong Zhao. STFF: Spatio-temporal and frequency fusion for video compression artifact removal.IEEE Trans- actions on Broadcasting, 71(2):542–554, 2025

  22. [30]

    Image super-resolution via iterative refinement.IEEE Transac- tions on Pattern Analysis and Machine Intelligence, 45(4): 4713–4726, 2022.10

    Chitwan Saharia, Jonathan Ho, William Chan, Tim Sal- imans, David J Fleet, and Mohammad Norouzi. Image super-resolution via iterative refinement.IEEE Transac- tions on Pattern Analysis and Machine Intelligence, 45(4): 4713–4726, 2022.10

  23. [31]

    SRDiff: Single image super-resolution with diffusion probabilistic models

    Haoying Li, Yifan Yang, Meng Chang, Shiqi Chen, Huajun Feng, Zhihai Xu, Qi Li, and Yueting Chen. SRDiff: Single image super-resolution with diffusion probabilistic models. Neurocomputing, 479:47–59, 2022

  24. [32]

    Exploiting diffusion prior for real-world image super-resolution.International Journal of Computer Vision, 132(12):5929–5949, 2024

    Jianyi Wang, Zongsheng Yue, Shangchen Zhou, Kelvin CK Chan, and Chen Change Loy. Exploiting diffusion prior for real-world image super-resolution.International Journal of Computer Vision, 132(12):5929–5949, 2024

  25. [33]

    SeeSR: Towards semantics-aware real-world image super-resolution

    Rongyuan Wu, Tao Yang, Lingchen Sun, Zhengqiang Zhang, Shuai Li, and Lei Zhang. SeeSR: Towards semantics-aware real-world image super-resolution. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 25456–25467, 2024

  26. [34]

    SeedVR2: One-step video restoration via diffu- sion adversarial post-training

    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. SeedVR2: One-step video restoration via diffu- sion adversarial post-training. InInternational Conference o...

  27. [35]

    DriftRec: Adapting diffusion models to blind jpeg restora- tion.IEEE Transactions on Image Processing, 33:2795– 2807, 2024

    Simon Welker, Henry N Chapman, and Timo Gerkmann. DriftRec: Adapting diffusion models to blind jpeg restora- tion.IEEE Transactions on Image Processing, 33:2795– 2807, 2024

  28. [36]

    Spatial degradation-aware and temporal consistent diffusion model for compressed video super- resolution.arXiv preprint arXiv:2502.07381, 2025

    Hongyu An, Xinfeng Zhang, Shijie Zhao, Li Zhang, and Ruiqin Xiong. Spatial degradation-aware and temporal consistent diffusion model for compressed video super- resolution.arXiv preprint arXiv:2502.07381, 2025

  29. [37]

    LViCAR: Diffusion mod- els for perceptual quality enhancement in video compres- sion artifact reduction

    Shiv Gehlot and Guan-Ming Su. LViCAR: Diffusion mod- els for perceptual quality enhancement in video compres- sion artifact reduction. InInternational Workshop on Deep Multimodal Generation and Retrieval, pages 1–10, 2025

  30. [38]

    Recovering realistic texture in image super-resolution by deep spatial feature transform

    Xintao Wang, Ke Yu, Chao Dong, and Chen Change Loy. Recovering realistic texture in image super-resolution by deep spatial feature transform. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 606–615, 2018

  31. [39]

    OpenAI GPT-5 system card.arXiv preprint arXiv:2601.03267, 2025

    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.arXiv preprint arXiv:2601.03267, 2025

  32. [40]

    Learn- ing transferable visual models from natural language super- vision

    Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learn- ing transferable visual models from natural language super- vision. InInternational Conference on Machine Learning,...

  33. [41]

    LoRA: Low-rank adaptation of large language mod- els

    Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen- Zhu, Yuanzhi Li, Shean Wang, Liang Wang, Weizhu Chen, et al. LoRA: Low-rank adaptation of large language mod- els. InInternational Conference on Learning Representa- tions, 2022

  34. [42]

    ESRGAN: Enhanced super-resolution generative adversarial networks

    Xintao Wang, Ke Yu, Shixiang Wu, Jinjin Gu, Yihao Liu, Chao Dong, Yu Qiao, and Chen Change Loy. ESRGAN: Enhanced super-resolution generative adversarial networks. InEuropean Conference on Computer Vision Workshops, pages 63–79, 2018

  35. [43]

    Video enhancement with task-oriented flow.International Journal of Computer Vision, 127(8): 1106–1125, 2019

    Tianfan Xue, Baian Chen, Jiajun Wu, Donglai Wei, and William T Freeman. Video enhancement with task-oriented flow.International Journal of Computer Vision, 127(8): 1106–1125, 2019

  36. [44]

    Jens-Rainer Ohm, Gary J Sullivan, Heiko Schwarz, Thiow Keng Tan, and Thomas Wiegand. Comparison of the coding efficiency of video coding standards—including high efficiency video coding (HEVC).IEEE Transactions on Circuits and Systems for Video Technology, 22(12): 1669–1684, 2012

  37. [45]

    The unreasonable effectiveness of deep features as a perceptual metric

    Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shecht- man, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 586–595, 2018

  38. [46]

    Image quality assessment: Unifying structure and texture similarity.IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(5):2567–2581, 2020

    Keyan Ding, Kede Ma, Shiqi Wang, and Eero P Simoncelli. Image quality assessment: Unifying structure and texture similarity.IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(5):2567–2581, 2020

  39. [47]

    MUSIQ: Multi-scale image quality trans- former

    Junjie Ke, Qifei Wang, Yilin Wang, Peyman Milanfar, and Feng Yang. MUSIQ: Multi-scale image quality trans- former. InProceedings of the IEEE/CVF International Conference on Computer Vision, pages 5148–5157, 2021

  40. [48]

    Ex- ploring CLIP for assessing the look and feel of images

    Jianyi Wang, Kelvin CK Chan, and Chen Change Loy. Ex- ploring CLIP for assessing the look and feel of images. In Proceedings of the AAAI Conference on Artificial Intelli- gence, volume 37, pages 2555–2563, 2023

  41. [49]

    Exploring video quality assessment on user gen- erated contents from aesthetic and technical perspectives

    Haoning Wu, Erli Zhang, Liang Liao, Chaofeng Chen, Jingwen Hou, Annan Wang, Wenxiu Sun, Qiong Yan, and Weisi Lin. Exploring video quality assessment on user gen- erated contents from aesthetic and technical perspectives. InProceedings of the IEEE/CVF International Conference o...

  42. [50]

    Making a ’completely blind’ image quality analyzer.IEEE Signal Processing Letters, 20(3):209–212, 2012

    Anish Mittal, Rajiv Soundararajan, and Alan C Bovik. Making a ’completely blind’ image quality analyzer.IEEE Signal Processing Letters, 20(3):209–212, 2012

  43. [51]

    Learning temporal coherence via self- supervision for GAN-based video generation.ACM Trans- actions on Graphics, 39(4):75–1, 2020

    Mengyu Chu, You Xie, Jonas Mayer, Laura Leal-Taixé, and Nils Thuerey. Learning temporal coherence via self- supervision for GAN-based video generation.ACM Trans- actions on Graphics, 39(4):75–1, 2020. 11

Pith tools

Reviewed July 9, 2026 · model on record in the stance chip above.