Junru Li
Research Subjects
Research Interests
Junru Li has a total of 32 co-authors. Here are the top co-authors by number of shared publications:
- Li Zhang
ByteDance Inc., Beijing, China
14 shared publications - Kai Zhang
ByteDance Inc., Beijing, China
10 shared publications - Shiqi Wang
City University of Hong Kong, Hong Kong, China
10 shared publications - Meng Wang
School of Data Science, Lingnan University, Hong Kong SAR, China
8 shared publications - Siwei Ma
National Engineering Research Center of Visual Technology, Peking University, Beijing, China
7 shared publications - Yue Li
Bytedance Inc., San Diego, USA
5 shared publications
Latest Publications
- Open Access
- 9 November 2026
- Published by Association for Computing Machinery (ACM)
- p. 5963-5971
Semantic-Embedded Bitstream: An INR-based Image Compression Framework Supporting Instant Editing
- 9 November 2026
- Published by Association for Computing Machinery (ACM)
- p. 5963-5971
Open AccessSemantic-Embedded Bitstream: An INR-based Image Compression Framework Supporting Instant Editing
- 3 December 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Circuits and Systems for Video Technology
- Vol. 36 (4) , 5525-5537
CNVC: A Compact Neural Video Codec With Instance-Level Adaptation
Autoencoder-based neural compression methods leverage expressive models to fit large datasets but often incur considerable decoding complexity. Recently, overfitted codecs with reduced decoding complexity have gained attention as an alternative. However, they usually require access to entire videos or multiple frames simultaneously for encoding, resulting in substantial system delays. To address these limitations, we propose CNVC, a compact neural video codec that employs instance-level adaptation for efficient and flexible video compression. CNVC is fully overfitted (each frame is optimized independently using up to 45k iterations for maximum performance in this paper), building on the COOL-CHIC video model with substantial architectural and training enhancements. At a decoding complexity of just 1300 MACs per pixel, CNVC provides a more compact solution than previous autoencoder-based and overfitted codecs. Additionally, CNVC inherits the frame-wise overfitting mechanism of COOL-CHIC video, enabling flexible encoding configurations (e.g., low-delay1). In terms of compression efficiency, CNVC achieves significant bitrate reductions on HEVC and UVG datasets compared to COOL-CHIC video. To our knowledge, CNVC is the first compact neural codec to match HEVC (x265 slow setting) performance at such a decoding complexity level.
- 3 December 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Circuits and Systems for Video Technology
- Vol. 36 (4) , 5525-5537
CNVC: A Compact Neural Video Codec With Instance-Level Adaptation
Autoencoder-based neural compression methods leverage expressive models to fit large datasets but often incur considerable decoding complexity. Recently, overfitted codecs with reduced decoding complexity have gained attention as an alternative. However, they usually require access to entire videos or multiple frames simultaneously for encoding, resulting in substantial system delays. To address these limitations, we propose CNVC, a compact neural video codec that employs instance-level adaptation for efficient and flexible video compression. CNVC is fully overfitted (each frame is optimized independently using up to 45k iterations for maximum performance in this paper), building on the COOL-CHIC video model with substantial architectural and training enhancements. At a decoding complexity of just 1300 MACs per pixel, CNVC provides a more compact solution than previous autoencoder-based and overfitted codecs. Additionally, CNVC inherits the frame-wise overfitting mechanism of COOL-CHIC video, enabling flexible encoding configurations (e.g., low-delay1). In terms of compression efficiency, CNVC achieves significant bitrate reductions on HEVC and UVG datasets compared to COOL-CHIC video. To our knowledge, CNVC is the first compact neural codec to match HEVC (x265 slow setting) performance at such a decoding complexity level.
- Open Access
- 27 October 2025
- Published by Association for Computing Machinery (ACM)
- p. 7248-7257
BiECVC: Gated Diversification of Bidirectional Contexts for Learned Video Compression
- 27 October 2025
- Published by Association for Computing Machinery (ACM)
- p. 7248-7257
Open AccessBiECVC: Gated Diversification of Bidirectional Contexts for Learned Video Compression
- Open Access
- 26 October 2025
- Published by Association for Computing Machinery (ACM)
- p. 147-155
Objects Disentangled Implicit Neural Representation for Image Coding
- 26 October 2025
- Published by Association for Computing Machinery (ACM)
- p. 147-155
Open AccessObjects Disentangled Implicit Neural Representation for Image Coding
- Open Access
- 1 May 2025
- Published by Association for Computing Machinery (ACM) in ACM Transactions on Multimedia Computing, Communications, and Applications
Advanced Neural Network-based Video Coding Technologies for Intra Prediction and In-Loop Filtering
The past decade has witnessed the huge success of deep learning in well-known artificial intelligence applications such as face recognition, autonomous driving, and large language model like ChatGPT. Recently, the application of deep learning has been extended to a much wider range, with neural network-based video coding being one of them. Neural network-based video coding can be performed at two different levels: embedding neural network-based (NN-based) coding tools into a classical video compression framework or building the entire compression framework upon neural networks. This paper elaborates our studies in response to the recent exploration efforts in JVET (Joint Video Experts Team of ITU-T SG 16 WP 3 and ISO/IEC JTC 1/SC29) in the name of neural network-based video coding (NNVC), falling in the former category. Specifically, in this paper we propose two advanced NN-based video coding technologies, i.e. neural network-based intra prediction and neural network-based in-loop filtering, which have been investigated for several meeting cycles in JVET and then adopted into the reference software, i.e. NNVC 1 . In addition, we further propose a small ad-hoc deep-learning library (SADL), which provides integer-based inference capabilities for neural networks to ensure interoperability across different systems. SADL has been adopted as the inference platform of all neural networks in NNVC. Extensive experiments on top of the NNVC have been conducted to evaluate the effectiveness of the proposed techniques. Compared with VTM-11.0_nnvc 2 , the proposed two NN-based coding tools jointly achieve {11.94%, 21.86%, 22.59%}, {9.18%, 19.76%, 20.92%}, and {10.63%, 21.56%, 23.02%} BD-rate reductions on average for {Y, Cb, Cr} under random-access, low-delay, and all-intra configurations respectively.
- 1 May 2025
- Published by Association for Computing Machinery (ACM) in ACM Transactions on Multimedia Computing, Communications, and Applications
Open AccessAdvanced Neural Network-based Video Coding Technologies for Intra Prediction and In-Loop Filtering
The past decade has witnessed the huge success of deep learning in well-known artificial intelligence applications such as face recognition, autonomous driving, and large language model like ChatGPT. Recently, the application of deep learning has been extended to a much wider range, with neural network-based video coding being one of them. Neural network-based video coding can be performed at two different levels: embedding neural network-based (NN-based) coding tools into a classical video compression framework or building the entire compression framework upon neural networks. This paper elaborates our studies in response to the recent exploration efforts in JVET (Joint Video Experts Team of ITU-T SG 16 WP 3 and ISO/IEC JTC 1/SC29) in the name of neural network-based video coding (NNVC), falling in the former category. Specifically, in this paper we propose two advanced NN-based video coding technologies, i.e. neural network-based intra prediction and neural network-based in-loop filtering, which have been investigated for several meeting cycles in JVET and then adopted into the reference software, i.e. NNVC 1 . In addition, we further propose a small ad-hoc deep-learning library (SADL), which provides integer-based inference capabilities for neural networks to ensure interoperability across different systems. SADL has been adopted as the inference platform of all neural networks in NNVC. Extensive experiments on top of the NNVC have been conducted to evaluate the effectiveness of the proposed techniques. Compared with VTM-11.0_nnvc 2 , the proposed two NN-based coding tools jointly achieve {11.94%, 21.86%, 22.59%}, {9.18%, 19.76%, 20.92%}, and {10.63%, 21.56%, 23.02%} BD-rate reductions on average for {Y, Cb, Cr} under random-access, low-delay, and all-intra configurations respectively.
- 26 March 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Circuits and Systems for Video Technology
- Vol. PP (10518215) , 1
A Theoretical and Experimental Study for Dependent Learned Rate-Distortion Optimization
Recent advancements in learned rate-distortion optimization (RDO) showcase that by making the intra coding decisions based on a learned measure, the encoding can be significantly accelerated without incurring much coding loss. Despite great progress in complexity reduction, the dependency issue has been largely neglected in the current learned RDO research. In this study, aiming to tap the full potential of dependent learned RDO, we first derive a probabilistic RDO framework for theoretical analysis, under which the classic and the learned RDO problems are equivalent to the maximum a posteriori (MAP) inference and the distribution imitation, respectively. Subsequently, we probabilistically revisit dependency considerations in the intra RDO research. Our key finding is that the existing learned RDO scheme can only produce a measure that indicates the local “goodness” of coding decisions. We therefore further discuss the opportunities for learning a dependent measure that is more optimal in the long run. Finally, as learning an accurate measure for the full decision space could be extremely challenging, taking the High Efficiency Video Coding (HEVC) intra coding as a case study, we experimentally identify that the prediction decision accounts for the majority of the dependent optimization gain and is of the utmost value to be learned, paving the way for future research on dependent learned RDO.
- 26 March 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Circuits and Systems for Video Technology
- Vol. PP (10518215) , 1
A Theoretical and Experimental Study for Dependent Learned Rate-Distortion Optimization
Recent advancements in learned rate-distortion optimization (RDO) showcase that by making the intra coding decisions based on a learned measure, the encoding can be significantly accelerated without incurring much coding loss. Despite great progress in complexity reduction, the dependency issue has been largely neglected in the current learned RDO research. In this study, aiming to tap the full potential of dependent learned RDO, we first derive a probabilistic RDO framework for theoretical analysis, under which the classic and the learned RDO problems are equivalent to the maximum a posteriori (MAP) inference and the distribution imitation, respectively. Subsequently, we probabilistically revisit dependency considerations in the intra RDO research. Our key finding is that the existing learned RDO scheme can only produce a measure that indicates the local “goodness” of coding decisions. We therefore further discuss the opportunities for learning a dependent measure that is more optimal in the long run. Finally, as learning an accurate measure for the full decision space could be extremely challenging, taking the High Efficiency Video Coding (HEVC) intra coding as a case study, we experimentally identify that the prediction decision accounts for the majority of the dependent optimization gain and is of the utmost value to be learned, paving the way for future research on dependent learned RDO.
- 26 June 2024
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Circuits and Systems for Video Technology
- Vol. 34 (11) , 12086-12091
Enhanced Motion Compensated Temporal Filter for VVenC
Motion Compensated Temporal Filter (MCTF) has been repeatedly proven to be an effective pre-processing tool that improves the coding performance. The philosophy is that by smoothing with temporal filter, the noise of the to-be-coded image can be reduced, thereby shrinking the prediction residuals and improving the rate-distortion (RD) performance. While abundant efforts have been devoted to the design of the MCTF filter weights, how motion vector variance and texture complexity influence MCTF has been relatively under-explored. In this work, we propose an enhanced MCTF method (EMCTF) based on multi-hypothesis reference, motion vector variance, and texture complexity. We take initial steps towards the incorporation of motion vector variance and texture complexity in the filtering weights design. Motion compensation blocks based on multi-hypothesis reference, can be efficiently aggregated in an effort to obtain the final inference. The proposed method is implemented on the top of Versatile Video Encoder (VVenC). Experimental results show that for faster preset, fast preset, medium preset and slow preset, the proposed EMCTF achieves 0.85%, 0.91%, 0.85% and 0.82% Bjøntegaard delta rate (BD-rate) savings, respectively. Moreover, the EMCTF introduces little additional encoding complexity increase, facilitating its future applications in real-world scenarios.
- 26 June 2024
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Circuits and Systems for Video Technology
- Vol. 34 (11) , 12086-12091
Enhanced Motion Compensated Temporal Filter for VVenC
Motion Compensated Temporal Filter (MCTF) has been repeatedly proven to be an effective pre-processing tool that improves the coding performance. The philosophy is that by smoothing with temporal filter, the noise of the to-be-coded image can be reduced, thereby shrinking the prediction residuals and improving the rate-distortion (RD) performance. While abundant efforts have been devoted to the design of the MCTF filter weights, how motion vector variance and texture complexity influence MCTF has been relatively under-explored. In this work, we propose an enhanced MCTF method (EMCTF) based on multi-hypothesis reference, motion vector variance, and texture complexity. We take initial steps towards the incorporation of motion vector variance and texture complexity in the filtering weights design. Motion compensation blocks based on multi-hypothesis reference, can be efficiently aggregated in an effort to obtain the final inference. The proposed method is implemented on the top of Versatile Video Encoder (VVenC). Experimental results show that for faster preset, fast preset, medium preset and slow preset, the proposed EMCTF achieves 0.85%, 0.91%, 0.85% and 0.82% Bjøntegaard delta rate (BD-rate) savings, respectively. Moreover, the EMCTF introduces little additional encoding complexity increase, facilitating its future applications in real-world scenarios.
- 1 November 2023
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Image Processing
- Vol. 32 (10577149) , 6020-6031
Toward the Achievable Rate-Distortion Bound of VVC Intra Coding: A Beam Search-Based Joint Optimization Scheme
In this paper, we present the first attempt at determining where the achievable rate-distortion (R-D) performance bound in versatile video coding (VVC) intra coding is when considering the mutual dependency in the rate-distortion optimization (RDO) process. In particular, the abundant search space of encoding parameters in VVC intra coding is practically explored with a beam search-based joint rate-distortion optimization (BSJRDO) scheme. As such, the partitioning, prediction and transform decisions are jointly optimized across different coding units (CUs) with a customized search subset instead of the full space. To make the beam search process implementation-friendly for VVC, the dependencies among the CUs are truncated at different depths. To facilitate finer computational scalability, the beam size is flexibly adjusted based on the characteristics of the CUs, such that the operational points that satisfy different complexity demands for diverse applications can be practically obtained. The proposed BSJRDO approach, which fully conforms to the VVC decoding syntax, can serve as both the way toward the optimal RDO bound and a practical performance-boosting solution. BSJRDO is further implemented on a VVC coding platform (VVC Test model (VTM) 12.0), and extensive experiments show that BSJRDO can achieve 1.30% and 3.22% bit rate savings compared to the VTM anchor under the common test condition and low-bit-rate coding scenarios, respectively. Moreover, the performance gain can also be flexibly customized with different computational overheads.
- 1 November 2023
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Image Processing
- Vol. 32 (10577149) , 6020-6031
Toward the Achievable Rate-Distortion Bound of VVC Intra Coding: A Beam Search-Based Joint Optimization Scheme
In this paper, we present the first attempt at determining where the achievable rate-distortion (R-D) performance bound in versatile video coding (VVC) intra coding is when considering the mutual dependency in the rate-distortion optimization (RDO) process. In particular, the abundant search space of encoding parameters in VVC intra coding is practically explored with a beam search-based joint rate-distortion optimization (BSJRDO) scheme. As such, the partitioning, prediction and transform decisions are jointly optimized across different coding units (CUs) with a customized search subset instead of the full space. To make the beam search process implementation-friendly for VVC, the dependencies among the CUs are truncated at different depths. To facilitate finer computational scalability, the beam size is flexibly adjusted based on the characteristics of the CUs, such that the operational points that satisfy different complexity demands for diverse applications can be practically obtained. The proposed BSJRDO approach, which fully conforms to the VVC decoding syntax, can serve as both the way toward the optimal RDO bound and a practical performance-boosting solution. BSJRDO is further implemented on a VVC coding platform (VVC Test model (VTM) 12.0), and extensive experiments show that BSJRDO can achieve 1.30% and 3.22% bit rate savings compared to the VTM anchor under the common test condition and low-bit-rate coding scenarios, respectively. Moreover, the performance gain can also be flexibly customized with different computational overheads.
- Open Access
- 12 July 2023
- Published by Association for Computing Machinery (ACM) in ACM Transactions on Multimedia Computing, Communications, and Applications
- Vol. 19 (6) , 1-20
Compressed Screen Content Image Super Resolution
Screen content has become one of the prominent mediums in the increasingly connected world. With the prevalence of remote collaboration and communication such as virtual conferences and online education, recent years have witnessed a dramatic increase in the data volume of the screen content. Screen content compression serves as the fundamental technology in fostering the storage, transmission, and exhibition of screen content. In this article, we target the super-resolution of the compressed screen content images, intending to tackle the real-world challenge problems. A dataset is proposed for the super-resolution of the screen contents contaminated with different compression distortion levels. Subsequently, we introduce the principle of the multi-hypothesis into the super resolution and propose a new paradigm for the restoration of the compressed screen content images. The luminance and sharpness similarity metric is adopted in the network learning to better adapt to the screen content characteristic and ensure perceptual fidelity. Experimental results verify the superiority and effectiveness of the proposed method.
- 12 July 2023
- Published by Association for Computing Machinery (ACM) in ACM Transactions on Multimedia Computing, Communications, and Applications
- Vol. 19 (6) , 1-20
Open AccessCompressed Screen Content Image Super Resolution
Screen content has become one of the prominent mediums in the increasingly connected world. With the prevalence of remote collaboration and communication such as virtual conferences and online education, recent years have witnessed a dramatic increase in the data volume of the screen content. Screen content compression serves as the fundamental technology in fostering the storage, transmission, and exhibition of screen content. In this article, we target the super-resolution of the compressed screen content images, intending to tackle the real-world challenge problems. A dataset is proposed for the super-resolution of the screen contents contaminated with different compression distortion levels. Subsequently, we introduce the principle of the multi-hypothesis into the super resolution and propose a new paradigm for the restoration of the compressed screen content images. The luminance and sharpness similarity metric is adopted in the network learning to better adapt to the screen content characteristic and ensure perceptual fidelity. Experimental results verify the superiority and effectiveness of the proposed method.
- 13 December 2022
- Published by Institute of Electrical and Electronics Engineers (IEEE)
- p. 1-5
End-to-end Image Compression with Swin-Transformer
In this paper, we propose an end-to-end image compression framework, which cooperates with the swin-transformer modules to capture the localized and non-localized similarities in image compression. In particular, the swin-transformer modules are deployed in the analysis and synthesis stages, interleaving with convolution layers. The transformer layers are expected to perceive more flexible receptive fields, such that the spatially localized and non-localized redundancies could be more effectively eliminated. The proposed method reveals the excellent capability of signal conjunction and prediction, leading to the improvement of the rate and distortion performance. Experimental results show that the proposed method is superior to the existing methods on both natural scene and screen content images, where 22.46% BD-Rate savings are achieved when compared with the BPG. Over 30% BD-Rate gains could be observed with screen content images when compared with the classical hyper-prior end-to-end coding method.
- 13 December 2022
- Published by Institute of Electrical and Electronics Engineers (IEEE)
- p. 1-5
End-to-end Image Compression with Swin-Transformer
In this paper, we propose an end-to-end image compression framework, which cooperates with the swin-transformer modules to capture the localized and non-localized similarities in image compression. In particular, the swin-transformer modules are deployed in the analysis and synthesis stages, interleaving with convolution layers. The transformer layers are expected to perceive more flexible receptive fields, such that the spatially localized and non-localized redundancies could be more effectively eliminated. The proposed method reveals the excellent capability of signal conjunction and prediction, leading to the improvement of the rate and distortion performance. Experimental results show that the proposed method is superior to the existing methods on both natural scene and screen content images, where 22.46% BD-Rate savings are achieved when compared with the BPG. Over 30% BD-Rate gains could be observed with screen content images when compared with the classical hyper-prior end-to-end coding method.