Fuyuan Hu
Current Organizations / Institutes
The organizations/institutes from the scholar's latest article
Fuyuan Hu has a total of 116 co-authors. Here are the top co-authors by number of shared publications:
- Fan Lyu
Institute of Automation, Chinese Academy of Sciences
11 shared publications Fan Lyu
Institute of Automation New Laboratory of Pattern Recognition Chinese Academy of Sciences 100190 Beijing China
6 shared publicationsLinyan Li
Suzhou Institute of Trade & Commerce, Suzhou, China
6 shared publications- Tao Zhou
School of Computer Science and Engineering, North Minzu University, Yinchuan 750021, China
6 shared publications - Huiling Lu
School of Medical Information Engineering, Ningxia Medical University, Yinchuan, China
3 shared publications - Jinchang Ren
Guangdong Polytechnic Normal University
3 shared publications
Latest Publications
- 15 September 2026
- Published by Springer Nature
- p. 1-18
VD-LoRA: Adaptive Reuse of Low-Rank Directions for Continual Learning
- 15 September 2026
- Published by Springer Nature
- p. 1-18
VD-LoRA: Adaptive Reuse of Low-Rank Directions for Continual Learning
- 1 May 2026
- Published by Elsevier in Expert Systems with Applications
- Vol. 325
VRQA: Context-adaptive view routing for long-document question–answer generation
- 1 May 2026
- Published by Elsevier in Expert Systems with Applications
- Vol. 325
VRQA: Context-adaptive view routing for long-document question–answer generation
- Open Access
- 21 April 2026
- Published by MDPI AG in Electronics
- Vol. 15 (8) , 1755
Diffusion Models: Unlocking the “4 Secrets” of High-Quality Image Generation
The diffusion model (DM) is a hot topic in deep generative models and is widely applied in image generation. In diffusion models, there are four main “secrets” that affect high-quality image generation: constructing the diffusion model, improving the sampling velocity, designing the diffusion process, and guiding diffusion models. How should one construct the diffusion model? How can one improve the sampling velocity? How should one design the diffusion process? How should one guide diffusion models? These questions are critical to enhancing diffusion model performance. However, most existing review papers focus on applications, while discussion of the four key technical aspects remains limited. In response, this paper summarizes four key technologies and six representative application directions. First, the basic principles of diffusion models are reviewed from three perspectives: denoising diffusion probabilistic models, noise conditional score network models, and stochastic differential equation models. Second, key techniques for improving sampling velocity are summarized from three perspectives: non-Markovian sampling, knowledge distillation sampling, and discrete optimization sampling. Third, the diffusion process design is summarized from three perspectives: latent space, Transformer-based diffusion, and non-Euclidean space. Fourth, guidance strategies are summarized from three perspectives: classifier guidance, classifier-free guidance, and multimodal guidance. Fifth, the advantages and applications of diffusion models are discussed in high-quality text-to-image generation, high-quality text-to-video generation, and high-quality image-to-image generation. Finally, this paper discusses the challenges faced by diffusion models in image generation. Overall, this review systematically discusses the four “secrets” of diffusion models for image generation and provides a useful reference for future research in this field.
- 21 April 2026
- Published by MDPI AG in Electronics
- Vol. 15 (8) , 1755
Open AccessDiffusion Models: Unlocking the “4 Secrets” of High-Quality Image Generation
The diffusion model (DM) is a hot topic in deep generative models and is widely applied in image generation. In diffusion models, there are four main “secrets” that affect high-quality image generation: constructing the diffusion model, improving the sampling velocity, designing the diffusion process, and guiding diffusion models. How should one construct the diffusion model? How can one improve the sampling velocity? How should one design the diffusion process? How should one guide diffusion models? These questions are critical to enhancing diffusion model performance. However, most existing review papers focus on applications, while discussion of the four key technical aspects remains limited. In response, this paper summarizes four key technologies and six representative application directions. First, the basic principles of diffusion models are reviewed from three perspectives: denoising diffusion probabilistic models, noise conditional score network models, and stochastic differential equation models. Second, key techniques for improving sampling velocity are summarized from three perspectives: non-Markovian sampling, knowledge distillation sampling, and discrete optimization sampling. Third, the diffusion process design is summarized from three perspectives: latent space, Transformer-based diffusion, and non-Euclidean space. Fourth, guidance strategies are summarized from three perspectives: classifier guidance, classifier-free guidance, and multimodal guidance. Fifth, the advantages and applications of diffusion models are discussed in high-quality text-to-image generation, high-quality text-to-video generation, and high-quality image-to-image generation. Finally, this paper discusses the challenges faced by diffusion models in image generation. Overall, this review systematically discusses the four “secrets” of diffusion models for image generation and provides a useful reference for future research in this field.
- 16 March 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Circuits and Systems for Video Technology
- Vol. 36 (7) , 10019-10035
GAIN: Global-Atomic INteraction Graph for Few-Shot Class-Incremental Learning
Few-shot class-incremental learning (FSCIL) aims to develop models that can continually learn novel classes from only a few training samples while retaining knowledge of previously learned classes. The primary challenges of FSCIL lie in the limited available data in novel sessions, which can cause severe overfitting and exacerbate the catastrophic forgetting of novel classes. Prior studies have demonstrated the effectiveness of modeling graph relationships in few-shot learning, which can alleviate overfitting through feature aggregation. Building on these insights, we extend the concept to the incremental learning scenario and propose a Global-Atomic INteraction (GAIN) graph learning framework for FSCIL. Specifically, we introduce an Atomic-session Graph Network (AGN) to capture relationships among samples within each session, enabling the aggregation of similar samples and the extraction of more representative global-session features. Additionally, we design a Global-session Graph Network (GGN) that connects global-session features across base and novel classes, facilitating knowledge transfer between sessions and enhancing overall learning performance. Finally, we utilize a multi-stage strategy for training the GAIN model, which suppresses the training challenges posed by limited data in incremental learning sessions. The multi-stage training strategy is designed to build a GAIN graph from base to novel sessions and improve the capacity via an extra pseudo-incremental stage. Extensive experiments on three widely used FSCIL benchmarks demonstrate that our approach consistently outperforms existing methods and establishes new state-of-the-art performance. Our code will be released upon acceptance.
- 16 March 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Circuits and Systems for Video Technology
- Vol. 36 (7) , 10019-10035
GAIN: Global-Atomic INteraction Graph for Few-Shot Class-Incremental Learning
Few-shot class-incremental learning (FSCIL) aims to develop models that can continually learn novel classes from only a few training samples while retaining knowledge of previously learned classes. The primary challenges of FSCIL lie in the limited available data in novel sessions, which can cause severe overfitting and exacerbate the catastrophic forgetting of novel classes. Prior studies have demonstrated the effectiveness of modeling graph relationships in few-shot learning, which can alleviate overfitting through feature aggregation. Building on these insights, we extend the concept to the incremental learning scenario and propose a Global-Atomic INteraction (GAIN) graph learning framework for FSCIL. Specifically, we introduce an Atomic-session Graph Network (AGN) to capture relationships among samples within each session, enabling the aggregation of similar samples and the extraction of more representative global-session features. Additionally, we design a Global-session Graph Network (GGN) that connects global-session features across base and novel classes, facilitating knowledge transfer between sessions and enhancing overall learning performance. Finally, we utilize a multi-stage strategy for training the GAIN model, which suppresses the training challenges posed by limited data in incremental learning sessions. The multi-stage training strategy is designed to build a GAIN graph from base to novel sessions and improve the capacity via an extra pseudo-incremental stage. Extensive experiments on three widely used FSCIL benchmarks demonstrate that our approach consistently outperforms existing methods and establishes new state-of-the-art performance. Our code will be released upon acceptance.
- 13 January 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Geoscience and Remote Sensing
- Vol. 64 (01962892) , 1-12
IRPNet: Infrared Small Target Detection via RGB Prior Guidance and Physics Feature Fusion
Infrared small target detection (IRSTD) aims to tackle the tough issue of identifying dim and tiny targets in low signal-to-noise ratio (SNR) images, which is crucial for real-world applications such as surveillance, military, and marine rescue. Recent breakthroughs in deep learning have driven IRSTD advancements. However, these methods treat it as a pure black-box paradigm, neglecting the inherent physics of infrared images. This issue significantly affects the model’s accuracy, resulting in lower performance and increased false alarms. Furthermore, existing datasets are limited to about 1,000 samples per benchmark, whereas RGB datasets offer richer samples and information. Inspired by the fact that low SNR and the scarcity of datasets greatly challenge performance, we propose a novel IRSTD network with RGB prior guidance and physics feature fusion (IRPNet). Specifically, an image encoder of CLIP including rich RGB prior information is first employed to integrate RGB knowledge into IRSTD to enrich the model’s representational capacity. Subsequently, infrared-specific visual features are extracted using convolutional block attention mechanisms to capture complex pixel-level relationships. In parallel, a dedicated physical feature extraction (PFE) block is applied to capture the essential properties of infrared images, complementing purely data-driven approaches. Finally, these features are progressively fused through a physical-visual feature fusion (PVFF) block and a multi-scale feature fusion (MSFF) module to enhance representations. Extensive experiments on NUAA-SIRST, IRSTD-1k, NUDT-SIRST and SIRST-Aug demonstrate that IRPNet outperforms existing state-of-the-art (SOTA) methods, providing a more robust and accurate solution. The source code is available at https://github.com/rayyao/IRPNet.
- 13 January 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Geoscience and Remote Sensing
- Vol. 64 (01962892) , 1-12
IRPNet: Infrared Small Target Detection via RGB Prior Guidance and Physics Feature Fusion
Infrared small target detection (IRSTD) aims to tackle the tough issue of identifying dim and tiny targets in low signal-to-noise ratio (SNR) images, which is crucial for real-world applications such as surveillance, military, and marine rescue. Recent breakthroughs in deep learning have driven IRSTD advancements. However, these methods treat it as a pure black-box paradigm, neglecting the inherent physics of infrared images. This issue significantly affects the model’s accuracy, resulting in lower performance and increased false alarms. Furthermore, existing datasets are limited to about 1,000 samples per benchmark, whereas RGB datasets offer richer samples and information. Inspired by the fact that low SNR and the scarcity of datasets greatly challenge performance, we propose a novel IRSTD network with RGB prior guidance and physics feature fusion (IRPNet). Specifically, an image encoder of CLIP including rich RGB prior information is first employed to integrate RGB knowledge into IRSTD to enrich the model’s representational capacity. Subsequently, infrared-specific visual features are extracted using convolutional block attention mechanisms to capture complex pixel-level relationships. In parallel, a dedicated physical feature extraction (PFE) block is applied to capture the essential properties of infrared images, complementing purely data-driven approaches. Finally, these features are progressively fused through a physical-visual feature fusion (PVFF) block and a multi-scale feature fusion (MSFF) module to enhance representations. Extensive experiments on NUAA-SIRST, IRSTD-1k, NUDT-SIRST and SIRST-Aug demonstrate that IRPNet outperforms existing state-of-the-art (SOTA) methods, providing a more robust and accurate solution. The source code is available at https://github.com/rayyao/IRPNet.
- 2 December 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Circuits and Systems for Video Technology
- Vol. 36 (4) , 4433-4449
Mitigating Catastrophic Forgetting in Online Continual Learning With Dual-Margin Contrastive Replay
Online Continual Learning (OCL) enables machine learning models to learn from a stream of non-stationary tasks, making it more aligned with real-world scenarios. However, OCL faces a significant challenge: catastrophic forgetting, wherein the model learned in previous tasks is substantially overwritten upon encountering new tasks, leading to a biased forgetting of prior knowledge. Among various OCL strategies, replay-based methods have proven particularly effective in mitigating catastrophic forgetting by maintaining a small buffer of past samples and retraining them alongside new data. However, due to strict memory constraints, these replay buffers often fail to adequately represent the true data distribution of previous tasks. This leads to distributional shifts in the feature space, amplifying forgetting and degrading model performance. To address the problem, in this paper, we propose a novel replay strategy, termed Dual-Margin Contrastive Replay (DMCR), to anchor the distribution of old tasks and reduce the negative transfer effects. First, we propose to select memory for more representative samples guided by constructed centroids in a data stream. Then, to keep the model from distribution chaos in biased replay, a two-level angular cross-task Contrastive Margin Loss (CML) is proposed, to encourage the intra-class and intra-task compactness, and increase the inter-class and inter-task discrepancy. Finally, to further suppress the distributional drift, we present an optional Centroid Distillation Loss (CDL) on the replay memory to anchor the knowledge in feature space for each previous old task. Extensive experimental results on five benchmark datasets validate that the proposed DMCR can effectively mitigate the catastrophic forgetting and achieve state-of-the-art (SOTA) performance in OCL.
- 2 December 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Circuits and Systems for Video Technology
- Vol. 36 (4) , 4433-4449
Mitigating Catastrophic Forgetting in Online Continual Learning With Dual-Margin Contrastive Replay
Online Continual Learning (OCL) enables machine learning models to learn from a stream of non-stationary tasks, making it more aligned with real-world scenarios. However, OCL faces a significant challenge: catastrophic forgetting, wherein the model learned in previous tasks is substantially overwritten upon encountering new tasks, leading to a biased forgetting of prior knowledge. Among various OCL strategies, replay-based methods have proven particularly effective in mitigating catastrophic forgetting by maintaining a small buffer of past samples and retraining them alongside new data. However, due to strict memory constraints, these replay buffers often fail to adequately represent the true data distribution of previous tasks. This leads to distributional shifts in the feature space, amplifying forgetting and degrading model performance. To address the problem, in this paper, we propose a novel replay strategy, termed Dual-Margin Contrastive Replay (DMCR), to anchor the distribution of old tasks and reduce the negative transfer effects. First, we propose to select memory for more representative samples guided by constructed centroids in a data stream. Then, to keep the model from distribution chaos in biased replay, a two-level angular cross-task Contrastive Margin Loss (CML) is proposed, to encourage the intra-class and intra-task compactness, and increase the inter-class and inter-task discrepancy. Finally, to further suppress the distributional drift, we present an optional Centroid Distillation Loss (CDL) on the replay memory to anchor the knowledge in feature space for each previous old task. Extensive experimental results on five benchmark datasets validate that the proposed DMCR can effectively mitigate the catastrophic forgetting and achieve state-of-the-art (SOTA) performance in OCL.
- Open Access
Ambiguity-Guided Learnable Distribution Calibration for Semi-Supervised Few-Shot Class-Incremental Learning
Few-Shot Class-Incremental Learning (FSCIL) focuses on models learning new concepts from limited data while retaining knowledge of previous classes. Recently, many studies have started to leverage unlabeled samples to assist models in learning from few-shot samples, giving rise to the field of Semi-supervised Few-shot Class-Incremental Learning (Semi-FSCIL). However, these studies often assume that the source of unlabeled data is only confined to novel classes of the current session, which presents a narrow perspective and cannot align well with practical scenarios. To better reflect real-world scenarios, we redefine Semi-FSCIL as Generalized Semi-FSCIL (GSemi-FSCIL) by incorporating both base and all the ever-seen novel classes in the unlabeled set. This change in the composition of unlabeled samples poses a new challenge for existing methods, as they struggle to distinguish between unlabeled samples from base and novel classes. To address this issue, we propose an Ambiguity-guided Learnable Distribution Calibration (ALDC) strategy. ALDC dynamically uses abundant base samples to correct biased feature distributions for few-shot novel classes. Experiments on three benchmark datasets show that our method outperforms existing works, setting new state-of-the-art results.
Open AccessAmbiguity-Guided Learnable Distribution Calibration for Semi-Supervised Few-Shot Class-Incremental Learning
Few-Shot Class-Incremental Learning (FSCIL) focuses on models learning new concepts from limited data while retaining knowledge of previous classes. Recently, many studies have started to leverage unlabeled samples to assist models in learning from few-shot samples, giving rise to the field of Semi-supervised Few-shot Class-Incremental Learning (Semi-FSCIL). However, these studies often assume that the source of unlabeled data is only confined to novel classes of the current session, which presents a narrow perspective and cannot align well with practical scenarios. To better reflect real-world scenarios, we redefine Semi-FSCIL as Generalized Semi-FSCIL (GSemi-FSCIL) by incorporating both base and all the ever-seen novel classes in the unlabeled set. This change in the composition of unlabeled samples poses a new challenge for existing methods, as they struggle to distinguish between unlabeled samples from base and novel classes. To address this issue, we propose an Ambiguity-guided Learnable Distribution Calibration (ALDC) strategy. ALDC dynamically uses abundant base samples to correct biased feature distributions for few-shot novel classes. Experiments on three benchmark datasets show that our method outperforms existing works, setting new state-of-the-art results.
- Open Access
- 21 May 2025
- Published by Association for Computing Machinery (ACM) in ACM Transactions on Multimedia Computing, Communications, and Applications
Historical Object-Aware Prompt Learning for Universal Hyperspectral Object Tracking
Hyperspectral Object Tracking (HOT), utilizing rich spectral information from hyperspectral video (HSV), holds significant importance for object tracking. We identify that a major obstacle in improving HOT performance lies in effectively leveraging spectral and historical information. Furthermore, due to the mismatch in band dimensions between hyperspectral and RGB images, state-of-the-art RGB-based trackers struggle to adapt to unified HOT tasks. To address this, we propose a Historical Object-aware Prompt Learning (HOPL) method for universal hyperspectral object tracking. Initially, we transform hyperspectral image ( \(N\) bands) into multiple sets of 3 bands with different combinations and feed them into a backbone network to generate base features. Subsequently, we introduce a historical object-aware prompter, where historical object-aware images are input to generate prompt features that enhance the representation of object information when combined with base features. Additionally, we design a band information fusion module to integrate the multiple sets of base features. By introducing historical object-aware prompts, HOPL significantly enhances tracking performance without retraining the backbone network. Experimental results on the HOT2023 dataset (comprising HSV with 25-band, 16-band, and 15-band wavelength ranges) and HOT2022 dataset validate the superiority of HOPL over state-of-the-art methods. The source code is available at https://github.com/rayyao/HOPL .- 21 May 2025
- Published by Association for Computing Machinery (ACM) in ACM Transactions on Multimedia Computing, Communications, and Applications
Open AccessHistorical Object-Aware Prompt Learning for Universal Hyperspectral Object Tracking
Hyperspectral Object Tracking (HOT), utilizing rich spectral information from hyperspectral video (HSV), holds significant importance for object tracking. We identify that a major obstacle in improving HOT performance lies in effectively leveraging spectral and historical information. Furthermore, due to the mismatch in band dimensions between hyperspectral and RGB images, state-of-the-art RGB-based trackers struggle to adapt to unified HOT tasks. To address this, we propose a Historical Object-aware Prompt Learning (HOPL) method for universal hyperspectral object tracking. Initially, we transform hyperspectral image ( \(N\) bands) into multiple sets of 3 bands with different combinations and feed them into a backbone network to generate base features. Subsequently, we introduce a historical object-aware prompter, where historical object-aware images are input to generate prompt features that enhance the representation of object information when combined with base features. Additionally, we design a band information fusion module to integrate the multiple sets of base features. By introducing historical object-aware prompts, HOPL significantly enhances tracking performance without retraining the backbone network. Experimental results on the HOT2023 dataset (comprising HSV with 25-band, 16-band, and 15-band wavelength ranges) and HOT2022 dataset validate the superiority of HOPL over state-of-the-art methods. The source code is available at https://github.com/rayyao/HOPL . - Open Access
CALA: A Class-Aware Logit Adapter for Few-Shot Class-Incremental Learning
Few-Shot Class-Incremental Learning (FSCIL) defines a practical but challenging task where models are required to continuously learn novel concepts with only a few training samples. Due to data scarcity, existing FSCIL methods resort to training a backbone with abundant base data and then keeping it frozen afterward. However, the above operation often causes the backbone to overfit to base classes while overlooking the novel ones, leading to severe confusion between them. To address this issue, we propose Class-Aware Logit Adapter (CALA). Our method involves a lightweight adapter that learns to rectify biased predictions through a pseudo-incremental learning paradigm. In the real FSCIL process, we use the learned adapter to dynamically generate robust balancing factors. These factors can adjust confused novel instances back to their true label space based on their similarity to base classes. Specifically, when confusion is more likely to occur in novel instances that closely resemble base classes, greater rectification is required. Notably, CALA operates on the classifier level, preserving the original feature space, thus it can be flexibly plugged into most of the existing FSCIL works for improved performance. Experiments on three benchmark datasets consistently validate the effectiveness and flexibility of CALA. Codes will be available upon acceptance.
Open AccessCALA: A Class-Aware Logit Adapter for Few-Shot Class-Incremental Learning
Few-Shot Class-Incremental Learning (FSCIL) defines a practical but challenging task where models are required to continuously learn novel concepts with only a few training samples. Due to data scarcity, existing FSCIL methods resort to training a backbone with abundant base data and then keeping it frozen afterward. However, the above operation often causes the backbone to overfit to base classes while overlooking the novel ones, leading to severe confusion between them. To address this issue, we propose Class-Aware Logit Adapter (CALA). Our method involves a lightweight adapter that learns to rectify biased predictions through a pseudo-incremental learning paradigm. In the real FSCIL process, we use the learned adapter to dynamically generate robust balancing factors. These factors can adjust confused novel instances back to their true label space based on their similarity to base classes. Specifically, when confusion is more likely to occur in novel instances that closely resemble base classes, greater rectification is required. Notably, CALA operates on the classifier level, preserving the original feature space, thus it can be flexibly plugged into most of the existing FSCIL works for improved performance. Experiments on three benchmark datasets consistently validate the effectiveness and flexibility of CALA. Codes will be available upon acceptance.
- Open Access
- 13 December 2024
- Published by Springer Nature in Research Square
Diffusion Models: Unlocking the “4 secrets” of High-quality Image Generation
Diffusion Model (DM) is a hot topic in deep generative models, and it is widely applied in the image generation fields. In diffusion models, there are 4 main “secrets” that affect the generation high-quality image generation: constructing diffusion model, improving the sampling speed, designing diffusion process, and guiding diffusion models. However, how to construct the diffusion model? How to improve the sampling speed? How to design the diffusion process? How to guide diffusion models? These are critical to enhancing the performance of diffusion models. However, so far, most of the review papers are summarized from the application aspect, and the 4 key technologies of diffusion model are few. In response to the above issues, this paper summarizes 4 key technologies and 6 applications. The main innovative works are as following: Firstly, how to construct diffusion models? The basic principle of the diffusion model are summarized from 3 aspects: denoising diffusion probabilistic models, noise-conditioned score networks, and stochastic differential equations. Secondly, how to improve the sampling speed? Some key techniques for improving the sampling speed are summarized from 3 aspects: non-Markovian sampling, knowledge distillation, and discrete optimization. Thirdly, how to design the diffusion process? This paper summarizes how to design diffusion process from 3 aspects: Latent Space, diffusion process based on transformer and non-Euclidean space. Fourthly, how to guide diffusion models? This paper summarizes how to guide the diffusion model from 3 aspects: classifier guidance, classifier-free guidance, and multimodal guidance. Fifthly, the applications of diffusion models in various fields are discussed from 6 aspects: image fusion, medical image segmentation, image restoration, text-to-image generation, image super-resolution and text-to-video generation. Finally, this paper discusses the challenges faced by diffusion model in image generation, compares the diffusion model with other generation models, and looks forward to the future development direction of diffusion model. This paper systematically points out the "4 secrets" of diffusion models in the image generation fields, providing significant reference value for their research in this field.
- 13 December 2024
- Published by Springer Nature in Research Square
Open AccessDiffusion Models: Unlocking the “4 secrets” of High-quality Image Generation
Diffusion Model (DM) is a hot topic in deep generative models, and it is widely applied in the image generation fields. In diffusion models, there are 4 main “secrets” that affect the generation high-quality image generation: constructing diffusion model, improving the sampling speed, designing diffusion process, and guiding diffusion models. However, how to construct the diffusion model? How to improve the sampling speed? How to design the diffusion process? How to guide diffusion models? These are critical to enhancing the performance of diffusion models. However, so far, most of the review papers are summarized from the application aspect, and the 4 key technologies of diffusion model are few. In response to the above issues, this paper summarizes 4 key technologies and 6 applications. The main innovative works are as following: Firstly, how to construct diffusion models? The basic principle of the diffusion model are summarized from 3 aspects: denoising diffusion probabilistic models, noise-conditioned score networks, and stochastic differential equations. Secondly, how to improve the sampling speed? Some key techniques for improving the sampling speed are summarized from 3 aspects: non-Markovian sampling, knowledge distillation, and discrete optimization. Thirdly, how to design the diffusion process? This paper summarizes how to design diffusion process from 3 aspects: Latent Space, diffusion process based on transformer and non-Euclidean space. Fourthly, how to guide diffusion models? This paper summarizes how to guide the diffusion model from 3 aspects: classifier guidance, classifier-free guidance, and multimodal guidance. Fifthly, the applications of diffusion models in various fields are discussed from 6 aspects: image fusion, medical image segmentation, image restoration, text-to-image generation, image super-resolution and text-to-video generation. Finally, this paper discusses the challenges faced by diffusion model in image generation, compares the diffusion model with other generation models, and looks forward to the future development direction of diffusion model. This paper systematically points out the "4 secrets" of diffusion models in the image generation fields, providing significant reference value for their research in this field.