Ziyi Zhang
Current Organizations / Institutes
The organizations/institutes from the scholar's latest article
Previous Organizations / Institutes
The organizations / institutes from the scholar's other articles in the past ten years
Research Subjects
Ziyi Zhang has a total of 23 co-authors. Here are the top co-authors by number of shared publications:
- Bo Du
School of Computer Science, Wuhan University, Wuhan, China
2 shared publications Caishan Guo
School of Electric Power, South China University of Technology, Guangzhou, China
2 shared publications- Dacheng Tao
Nanyang Technological University
2 shared publications - Guolong Ma
Department of Electronic, Electrical and Systems Engineering, School of Engineering, University of Birmingham, Edgbaston, Birmingham, U.K.
2 shared publications Haizhu Wang
Power Dispatching Control Center, Guangdong Power Grid Company, Ltd., Guangzhou, China
2 shared publications- Li Shen
Shenzhen Campus of Sun Yat-sen University
2 shared publications
Latest Publications
- Top Cited Papers
- 1 July 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Pattern Analysis and Machine Intelligence
- Vol. 48 (7) , 7375-7386
Aligning Few-Step Diffusion Models With Dense Reward Difference Learning
Few-step diffusion models enable efficient high-resolution image synthesis but struggle to align with specific downstream objectives due to limitations of existing reinforcement learning (RL) methods in low-step regimes with limited state spaces and suboptimal sample quality. To address this, we propose Stepwise Diffusion Policy Optimization (SDPO), a novel RL framework tailored for few-step diffusion models. SDPO introduces a dual-state trajectory sampling mechanism, tracking both noisy and predicted clean states at each step to provide dense reward feedback and enable low-variance, mixed-step optimization. For further efficiency, we develop a latent similarity-based dense reward prediction strategy to minimize costly dense reward queries. Leveraging these dense rewards, SDPO optimizes a dense reward difference learning objective that enables more frequent and granular policy updates. Additional refinements, including stepwise advantage estimates, temporal importance weighting, and step-shuffled gradient updates, further enhance long-term dependency, low-step priority, and gradient stability. Our experiments demonstrate that SDPO consistently delivers superior reward-aligned results across diverse few-step settings and tasks.
- 1 July 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Pattern Analysis and Machine Intelligence
- Vol. 48 (7) , 7375-7386
Top Cited PapersAligning Few-Step Diffusion Models With Dense Reward Difference Learning
Few-step diffusion models enable efficient high-resolution image synthesis but struggle to align with specific downstream objectives due to limitations of existing reinforcement learning (RL) methods in low-step regimes with limited state spaces and suboptimal sample quality. To address this, we propose Stepwise Diffusion Policy Optimization (SDPO), a novel RL framework tailored for few-step diffusion models. SDPO introduces a dual-state trajectory sampling mechanism, tracking both noisy and predicted clean states at each step to provide dense reward feedback and enable low-variance, mixed-step optimization. For further efficiency, we develop a latent similarity-based dense reward prediction strategy to minimize costly dense reward queries. Leveraging these dense rewards, SDPO optimizes a dense reward difference learning objective that enables more frequent and granular policy updates. Additional refinements, including stepwise advantage estimates, temporal importance weighting, and step-shuffled gradient updates, further enhance long-term dependency, low-step priority, and gradient stability. Our experiments demonstrate that SDPO consistently delivers superior reward-aligned results across diverse few-step settings and tasks.
- 18 July 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Pattern Analysis and Machine Intelligence
- Vol. 47 (11) , 9550-9562
Aligning Text-to-Image Diffusion Models With Constrained Reinforcement Learning
Reward finetuning has emerged as a powerful technique for aligning diffusion models with specific downstream objectives or user preferences. However, current approaches suffer from a persistent challenge of reward overoptimization, where models exploit imperfect reward feedback at the expense of overall performance. In this work, we identify three key contributors to overoptimization: (1) a granularity mismatch between the multi-step diffusion process and sparse rewards; (2) a loss of plasticity that limits the model’s ability to adapt and generalize; and (3) an overly narrow focus on a single reward objective that neglects complementary performance criteria. Accordingly, we introduce Constrained Diffusion Policy Optimization (CDPO), a novel reinforcement learning framework that addresses reward overoptimization from multiple angles. Firstly, CDPO tackles the granularity mismatch through a temporal policy optimization strategy that delivers step-specific rewards throughout the entire diffusion trajectory, thereby reducing the risk of overfitting to sparse final-step rewards. Then we incorporate a neuron reset strategy that selectively resets overactive neurons in the model, preventing overoptimization induced by plasticity loss. Finally, to avoid overfitting to a narrow reward objective, we integrate constrained reinforcement learning with auxiliary reward objectives serving as explicit constraints, ensuring a balanced optimization across diverse performance metrics.
- 18 July 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Pattern Analysis and Machine Intelligence
- Vol. 47 (11) , 9550-9562
Aligning Text-to-Image Diffusion Models With Constrained Reinforcement Learning
Reward finetuning has emerged as a powerful technique for aligning diffusion models with specific downstream objectives or user preferences. However, current approaches suffer from a persistent challenge of reward overoptimization, where models exploit imperfect reward feedback at the expense of overall performance. In this work, we identify three key contributors to overoptimization: (1) a granularity mismatch between the multi-step diffusion process and sparse rewards; (2) a loss of plasticity that limits the model’s ability to adapt and generalize; and (3) an overly narrow focus on a single reward objective that neglects complementary performance criteria. Accordingly, we introduce Constrained Diffusion Policy Optimization (CDPO), a novel reinforcement learning framework that addresses reward overoptimization from multiple angles. Firstly, CDPO tackles the granularity mismatch through a temporal policy optimization strategy that delivers step-specific rewards throughout the entire diffusion trajectory, thereby reducing the risk of overfitting to sparse final-step rewards. Then we incorporate a neuron reset strategy that selectively resets overactive neurons in the model, preventing overoptimization induced by plasticity loss. Finally, to avoid overfitting to a narrow reward objective, we integrate constrained reinforcement learning with auxiliary reward objectives serving as explicit constraints, ensuring a balanced optimization across diverse performance metrics.
- 23 February 2023
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Multimedia
- Vol. 25 (15209210) , 1521-1528
HRNeXt: High-Resolution Context Network for Crowd Pose Estimation
Occlusion handling in crowded scenes is an intractable challenge for human pose estimation. To address this problem, we propose two novel feed-forward network structures named Global Feed-Forward Network (GFFN) and Dynamic Feed-Forward Network (DFFN), which are specifically designed for image-based tasks to capture both local and global contextual information within intermediate features and update feature representations with high adaptability for occlusions. By exploiting the context modeling ability of the proposed GFFN and DFFN, we present a novel backbone network, namely High-Resolution Context Network (HRNeXt), which learns high-resolution representations with abundant contextual information to better estimate poses of occluded human bodies. Compared to state-of-the-art pose estimation networks, our HRNeXt absorbs advantages of convolution operation and attention mechanism, and it is more efficient in terms of training data sizes, network parameters and computational costs. Experimental results show that our HRNeXt significantly outperforms state-of-the-art backbone networks on challenging pose estimation datasets with high occurrence of crowds and occlusions. Code is available at: https://github.com/ZiyiZhang27/HRNeXt .
- 23 February 2023
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Multimedia
- Vol. 25 (15209210) , 1521-1528
HRNeXt: High-Resolution Context Network for Crowd Pose Estimation
Occlusion handling in crowded scenes is an intractable challenge for human pose estimation. To address this problem, we propose two novel feed-forward network structures named Global Feed-Forward Network (GFFN) and Dynamic Feed-Forward Network (DFFN), which are specifically designed for image-based tasks to capture both local and global contextual information within intermediate features and update feature representations with high adaptability for occlusions. By exploiting the context modeling ability of the proposed GFFN and DFFN, we present a novel backbone network, namely High-Resolution Context Network (HRNeXt), which learns high-resolution representations with abundant contextual information to better estimate poses of occluded human bodies. Compared to state-of-the-art pose estimation networks, our HRNeXt absorbs advantages of convolution operation and attention mechanism, and it is more efficient in terms of training data sizes, network parameters and computational costs. Experimental results show that our HRNeXt significantly outperforms state-of-the-art backbone networks on challenging pose estimation datasets with high occurrence of crowds and occlusions. Code is available at: https://github.com/ZiyiZhang27/HRNeXt .
- 13 July 2021
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Internet of Things Journal
- Vol. 9 (4) , 3112-3124
Collaborative Dynamic Task Allocation With Demand Response in Cloud-Assisted Multiedge System for Smart Grids
Collaborative cloud–edge Power Internet of Things technology is required to support the development of smart grids, which have become intelligent, green, and regionally autonomous systems. The diversity of electricity customer behaviors and different computational intensities of energy management applications present challenges for task allocation among computing resources that belong to different agents. In this article, we propose a novel trilevel collaborative optimization model to comprehensively consider the relation among various agents, including users, edge nodes (ENs), a cloud center (CC), and a multiedge league (MEL). We first formulate a Stackelberg game between users and ENs modeled as the lower level and middle level. In addition, with the assistance of the CC, we propose a MEL cooperation scheme to analyze the collaborative task allocation problem among multiple edges, which is modeled as the upper level to maximize the social welfare of the multiedge system (MES) without damaging the interests of the various ENs. The proposed trilevel model is equivalent to a bilevel program, solved by the proposed collaborative dynamic task allocation (CDTA) algorithm. Numerical simulations are presented to verify the proposed scheme and the results show that this scheme is effective for task allocation among users, ENs, the cloud, and the MEL in a cloud-assisted MES.
- 13 July 2021
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Internet of Things Journal
- Vol. 9 (4) , 3112-3124
Collaborative Dynamic Task Allocation With Demand Response in Cloud-Assisted Multiedge System for Smart Grids
Collaborative cloud–edge Power Internet of Things technology is required to support the development of smart grids, which have become intelligent, green, and regionally autonomous systems. The diversity of electricity customer behaviors and different computational intensities of energy management applications present challenges for task allocation among computing resources that belong to different agents. In this article, we propose a novel trilevel collaborative optimization model to comprehensively consider the relation among various agents, including users, edge nodes (ENs), a cloud center (CC), and a multiedge league (MEL). We first formulate a Stackelberg game between users and ENs modeled as the lower level and middle level. In addition, with the assistance of the CC, we propose a MEL cooperation scheme to analyze the collaborative task allocation problem among multiple edges, which is modeled as the upper level to maximize the social welfare of the multiedge system (MES) without damaging the interests of the various ENs. The proposed trilevel model is equivalent to a bilevel program, solved by the proposed collaborative dynamic task allocation (CDTA) algorithm. Numerical simulations are presented to verify the proposed scheme and the results show that this scheme is effective for task allocation among users, ENs, the cloud, and the MEL in a cloud-assisted MES.
- 1 December 2019
- Published by Institute of Electrical and Electronics Engineers (IEEE)
- p. 7-12
Resource Provision Method of Cloud Platform for Power Distribution System via Workload Prediction
With the development of smartness of power distribution system (PDS), PDS is confronted with multiple events and its operation state changes greatly and frequently. The consequent applications vary over time leading to dynamic workload, which challenges resource provision in cloud platform for PDS. In this paper, we propose a resource provision method for cloud platform in PDS via workload prediction. Starting from the causation of workload change, we forecast the operation state of DS by Markov Chain. We introduce the knowledge-based system (KBS) that contains logic relationship with operation states, events and applications, by which we predict the future workload under different operation states. We proposed a proactive dynamic computing resource provision model considering the predicted workload aiming at high-speed execution and economic operation in the cloud. The optimal solution is obtained from Pareto sets by technique for order preference by similarity to an ideal solution (TOPSIS). Results show that our proposed approach can not only improve the execution speed for DS application but also utilize the resources efficiently and economically.
- 1 December 2019
- Published by Institute of Electrical and Electronics Engineers (IEEE)
- p. 7-12
Resource Provision Method of Cloud Platform for Power Distribution System via Workload Prediction
With the development of smartness of power distribution system (PDS), PDS is confronted with multiple events and its operation state changes greatly and frequently. The consequent applications vary over time leading to dynamic workload, which challenges resource provision in cloud platform for PDS. In this paper, we propose a resource provision method for cloud platform in PDS via workload prediction. Starting from the causation of workload change, we forecast the operation state of DS by Markov Chain. We introduce the knowledge-based system (KBS) that contains logic relationship with operation states, events and applications, by which we predict the future workload under different operation states. We proposed a proactive dynamic computing resource provision model considering the predicted workload aiming at high-speed execution and economic operation in the cloud. The optimal solution is obtained from Pareto sets by technique for order preference by similarity to an ideal solution (TOPSIS). Results show that our proposed approach can not only improve the execution speed for DS application but also utilize the resources efficiently and economically.