In this paper,we propose a new privacy-aware transmission scheduling algorithm for 6G ad hoc networks.This system enables end nodes to select the optimum time and scheme to transmit private data safely.In 6G dynamic h...In this paper,we propose a new privacy-aware transmission scheduling algorithm for 6G ad hoc networks.This system enables end nodes to select the optimum time and scheme to transmit private data safely.In 6G dynamic heterogeneous infrastructures,unstable links and non-uniform hardware capabilities create critical issues regarding security and privacy.Traditional protocols are often too computationally heavy to allow 6G services to achieve their expected Quality-of-Service(QoS).As the transport network is built of ad hoc nodes,there is no guarantee about their trustworthiness or behavior,and transversal functionalities are delegated to the extreme nodes.However,while security can be guaranteed in extreme-to-extreme solutions,privacy cannot,as all intermediate nodes still have to handle the data packets they are transporting.Besides,traditional schemes for private anonymous ad hoc communications are vulnerable against modern intelligent attacks based on learning models.The proposed scheme fulfills this gap.Findings show the probability of a successful intelligent attack reduces by up to 65%compared to ad hoc networks with no privacy protection strategy when used the proposed technology.While congestion probability can remain below 0.001%,as required in 6G services.展开更多
Human motion modeling is a core technology in computer animation,game development,and humancomputer interaction.In particular,generating natural and coherent in-between motion using only the initial and terminal frame...Human motion modeling is a core technology in computer animation,game development,and humancomputer interaction.In particular,generating natural and coherent in-between motion using only the initial and terminal frames remains a fundamental yet unresolved challenge.Existing methods typically rely on dense keyframe inputs or complex prior structures,making it difficult to balance motion quality and plausibility under conditions such as sparse constraints,long-term dependencies,and diverse motion styles.To address this,we propose a motion generation framework based on a frequency-domain diffusion model,which aims to better model complex motion distributions and enhance generation stability under sparse conditions.Our method maps motion sequences to the frequency domain via the Discrete Cosine Transform(DCT),enabling more effective modeling of low-frequency motion structures while suppressing high-frequency noise.A denoising network based on self-attention is introduced to capture long-range temporal dependencies and improve global structural awareness.Additionally,a multi-objective loss function is employed to jointly optimize motion smoothness,pose diversity,and anatomical consistency,enhancing the realism and physical plausibility of the generated sequences.Comparative experiments on the Human3.6M and LaFAN1 datasets demonstrate that our method outperforms state-of-the-art approaches across multiple performance metrics,showing stronger capabilities in generating intermediate motion frames.This research offers a new perspective and methodology for human motion generation and holds promise for applications in character animation,game development,and virtual interaction.展开更多
Crack detection accuracy in computer vision is often constrained by limited annotated datasets.Although Generative Adversarial Networks(GANs)have been applied for data augmentation,they frequently introduce blurs and ...Crack detection accuracy in computer vision is often constrained by limited annotated datasets.Although Generative Adversarial Networks(GANs)have been applied for data augmentation,they frequently introduce blurs and artifacts.To address this challenge,this study leverages Denoising Diffusion Probabilistic Models(DDPMs)to generate high-quality synthetic crack images,enriching the training set with diverse and structurally consistent samples that enhance the crack segmentation.The proposed framework involves a two-stage pipeline:first,DDPMs are used to synthesize high-fidelity crack images that capture fine structural details.Second,these generated samples are combined with real data to train segmentation networks,thereby improving accuracy and robustness in crack detection.Compared with GAN-based approaches,DDPM achieved the best fidelity,with the highest Structural Similarity Index(SSIM)(0.302)and lowest Learned Perceptual Image Patch Similarity(LPIPS)(0.461),producing artifact-free images that preserve fine crack details.To validate its effectiveness,six segmentation models were tested,among which LinkNet consistently achieved the best performance,excelling in both region-level accuracy and structural continuity.Incorporating DDPM-augmented data further enhanced segmentation outcomes,increasing F1 scores by up to 1.1%and IoU by 1.7%,while also improving boundary alignment and skeleton continuity compared with models trained on real images alone.Experiments with varying augmentation ratios showed consistent improvements,with F1 rising from 0.946(no augmentation)to 0.957 and IoU from 0.897 to 0.913 at the highest ratio.These findings demonstrate the effectiveness of diffusion-based augmentation for complex crack detection in structural health monitoring.展开更多
Scalable simulation leveraging real-world data plays an essential role in advancing autonomous driving,owing to its efficiency and applicability in both training and evaluating algorithms.Consequently,there has been i...Scalable simulation leveraging real-world data plays an essential role in advancing autonomous driving,owing to its efficiency and applicability in both training and evaluating algorithms.Consequently,there has been increasing attention on generating highly realistic and consistent driving videos,particularly those involving viewpoint changes guided by the control commands or trajectories of ego vehicles.However,current reconstruction approaches,such as Neural Radiance Fields and 3D Gaussian Splatting,frequently suffer from limited generalization and depend on substantial input data.Meanwhile,2D generative models,though capable of producing unknown scenes,still have room for improvement in terms of coherence and visual realism.To overcome these challenges,we introduce GenScene,a world model that synthesizes front-view driving videos conditioned on trajectories.A new temporal module is presented to improve video consistency by extracting the global context of each frame,calculating relationships of frames using these global representations,and fusing frame contexts accordingly.Moreover,we propose an innovative attention mechanism that computes relations of pixels within each frame and pixels in the corresponding window range of the initial frame.Extensive experiments show that our approach surpasses various state-of-the-art models in driving video generation,and the introduced modules contribute significantly to model performance.This work establishes a new paradigm for goal-oriented video synthesis in autonomous driving,which facilitates on-demand simulation to expedite algorithm development.展开更多
With the development of technology,diffusion model-based solvers have shown significant promise in solving Combinatorial Optimization(CO)problems,particularly in tackling Non-deterministic Polynomial-time hard(NP-hard...With the development of technology,diffusion model-based solvers have shown significant promise in solving Combinatorial Optimization(CO)problems,particularly in tackling Non-deterministic Polynomial-time hard(NP-hard)problems such as the Traveling Salesman Problem(TSP).However,existing diffusion model-based solvers typically employ a fixed,uniform noise schedule(e.g.,linear or cosine annealing)across all training instances,failing to fully account for the unique characteristics of each problem instance.To address this challenge,we present GraphGuided Diffusion Solvers(GGDS),an enhanced method for improving graph-based diffusion models.GGDS leverages Graph Neural Networks(GNNs)to capture graph structural information embedded in node coordinates and adjacency matrices,dynamically adjusting the noise levels in the diffusion model.This study investigates the TSP by examining two distinct time-step noise generation strategies:cosine annealing and a Neural Network(NN)-based approach.We evaluate their performance across different problem scales,particularly after integrating graph structural information.Experimental results indicate that GGDS outperforms previous methods with average performance improvements of 18.7%,6.3%,and 88.7%on TSP-500,TSP-100,and TSP-50,respectively.Specifically,GGDS demonstrates superior performance on TSP-500 and TSP-50,while its performance on TSP-100 is either comparable to or slightly better than that of previous methods,depending on the chosen noise schedule and decoding strategy.展开更多
Aim The purpose of this study was to develop a mathe-matical model to quantitatively describe the passive trans-port of macromolecules within dental biofilms. Methodology Fluorescently labeled dextrans with different ...Aim The purpose of this study was to develop a mathe-matical model to quantitatively describe the passive trans-port of macromolecules within dental biofilms. Methodology Fluorescently labeled dextrans with different molecular mass （3 kD,10 kD,40 kD,70 kD,2 000 kD） were used as a series of diffusion probes. Streptococcus mutans,Streptococcus sanguinis,Actinomyces naeslundii and Fusobacterium nucleatum were used as inocula for biofilm formation. The diffusion processes of different probes through the in vitro biofilm were recorded with a confocal laser microscope. Results Mathematical function of biofilm penetration was constructed on the basis of the inverse problem method. Based on this function,not only the relationship between average concentration of steady-state and molecule weights can be analyzed,but also that between penetrative time and molecule weights. Conclusion This can be used to predict the effective concentration and the penetrative time of anti-biofilm medicines that can diffuse through oral biofilm. Further-more,an improved model for large molecule is proposed by considering the exchange time at the upper boundary of the dental biofilm.展开更多
In this study,the removal of monovalent and divalent cations,Nat,Kt,Mg2t,and Ca2t,in a diluted solution from Chott-El Jerid Lake,Tunisia,was investigated with the electrodialysis technique.The process was tested using...In this study,the removal of monovalent and divalent cations,Nat,Kt,Mg2t,and Ca2t,in a diluted solution from Chott-El Jerid Lake,Tunisia,was investigated with the electrodialysis technique.The process was tested using two cation-exchange membranes:sulfonated polyether sulfone cross-linked with 10%hexamethylenediamine(HEXCl)and sulfonated polyether sulfone grafted with octylamine(S-PESOS).The commercially available membrane Nafion®was used for comparison.The results showed that Nafion®and S-PESOS membranes had similar removal behaviors,and the investigated cations were ranked in the following descending order in terms of their demineralization rates:Nat>Ca2t>Mg2t>Kt.Divalent cations were more effectively removed by HEXCl than by monovalent cations.The plots based on the WebereMorris model showed a strong linearity.This reveals that intra-particle diffusion was not the removal rate-determining step,and the removal process was controlled by two or more concurrent mechanisms.The Boyd plots did not pass through their origin,and the sole controlling step was determined by film-diffusion resistance,especially after a long period of electrodialysis.Additionally,a semi-empirical model was established to simulate the temporal variation of the treatment process,and the physical significance and values of model parameters were compared for the three membranes.The findings of this study indicate that HEXCl and S-PESOS membranes can be efficiently utilized for water softening,especially when effluents are highly loaded with calcium and magnesium ions.展开更多
Accurately identifying building distribution from remote sensing images with complex background information is challenging.The emergence of diffusion models has prompted the innovative idea of employing the reverse de...Accurately identifying building distribution from remote sensing images with complex background information is challenging.The emergence of diffusion models has prompted the innovative idea of employing the reverse denoising process to distill building distribution from these complex backgrounds.Building on this concept,we propose a novel framework,building extraction diffusion model(BEDiff),which meticulously refines the extraction of building footprints from remote sensing images in a stepwise fashion.Our approach begins with the design of booster guidance,a mechanism that extracts structural and semantic features from remote sensing images to serve as priors,thereby providing targeted guidance for the diffusion process.Additionally,we introduce a cross-feature fusion module(CFM)that bridges the semantic gap between different types of features,facilitating the integration of the attributes extracted by booster guidance into the diffusion process more effectively.Our proposed BEDiff marks the first application of diffusion models to the task of building extraction.Empirical evidence from extensive experiments on the Beijing building dataset demonstrates the superior performance of BEDiff,affirming its effectiveness and potential for enhancing the accuracy of building extraction in complex urban landscapes.展开更多
Imputation of missing data has long been an important topic and an essential application for intelligent transportation systems(ITS)in the real world.As a state-of-the-art generative model,the diffusion model has prov...Imputation of missing data has long been an important topic and an essential application for intelligent transportation systems(ITS)in the real world.As a state-of-the-art generative model,the diffusion model has proven highly successful in image generation,speech generation,time series modelling etc.and now opens a new avenue for traffic data imputation.In this paper,we propose a conditional diffusion model,called the implicit-explicit diffusion model,for traffic data imputation.This model exploits both the implicit and explicit feature of the data simultaneously.More specifically,we design two types of feature extraction modules,one to capture the implicit dependencies hidden in the raw data at multiple time scales and the other to obtain the long-term temporal dependencies of the time series.This approach not only inherits the advantages of the diffusion model for estimating missing data,but also takes into account the multiscale correlation inherent in traffic data.To illustrate the performance of the model,extensive experiments are conducted on three real-world time series datasets using different missing rates.The experimental results demonstrate that the model improves imputation accuracy and generalization capability.展开更多
Recently,diffusion models have emerged as a promising paradigm for molecular design and optimization.However,most diffusion-based molecular generative models focus on modeling 2D graphs or 3D geom-etries,with limited ...Recently,diffusion models have emerged as a promising paradigm for molecular design and optimization.However,most diffusion-based molecular generative models focus on modeling 2D graphs or 3D geom-etries,with limited research on molecular sequence diffusion models.The International Union of Pure and Applied Chemistry(IUPAC)names are more akin to chemical natural language than the simplified molecular input line entry system(SMILES)for organic compounds.In this work,we apply an IUPAC-guided conditional diffusion model to facilitate molecular editing from chemical natural language to chemical language(SMILES)and explore whether the pre-trained generative performance of diffusion models can be transferred to chemical natural language.We propose DiffIUPAC,a controllable molecular editing diffusion model that converts IUPAC names to SMILES strings.Evaluation results demonstrate that our model out-performs existing methods and successfully captures the semantic rules of both chemical languages.Chemical space and scaffold analysis show that the model can generate similar compounds with diverse scaffolds within the specified constraints.Additionally,to illustrate the model’s applicability in drug design,we conducted case studies in functional group editing,analogue design and linker design.展开更多
Air target intent recognition holds significant importance in aiding commanders to assess battlefield situations and secure a competitive edge in decision-making.Progress in this domain has been hindered by challenges...Air target intent recognition holds significant importance in aiding commanders to assess battlefield situations and secure a competitive edge in decision-making.Progress in this domain has been hindered by challenges posed by imbalanced battlefield data and the limited robustness of traditional recognition models.Inspired by the success of diffusion models in addressing visual domain sample imbalances,this paper introduces a new approach that utilizes the Markov Transfer Field(MTF)method for time series data visualization.This visualization,when combined with the Denoising Diffusion Probabilistic Model(DDPM),effectively enhances sample data and mitigates noise within the original dataset.Additionally,a transformer-based model tailored for time series visualization and air target intent recognition is developed.Comprehensive experimental results,encompassing comparative,ablation,and denoising validations,reveal that the proposed method achieves a notable 98.86%accuracy in air target intent recognition while demonstrating exceptional robustness and generalization capabilities.This approach represents a promising avenue for advancing air target intent recognition.展开更多
The application of generative artificial intelligence(AI)is bringing about notable changes in anime creation.This paper surveys recent advancements and applications of diffusion and language models in anime generation...The application of generative artificial intelligence(AI)is bringing about notable changes in anime creation.This paper surveys recent advancements and applications of diffusion and language models in anime generation,focusing on their demonstrated potential to enhance production efficiency through automation and personalization.Despite these benefits,it is crucial to acknowledge the substantial initial computational investments required for training and deploying these models.We conduct an in-depth survey of cutting-edge generative AI technologies,encompassing models such as Stable Diffusion and GPT,and appraise pivotal large-scale datasets alongside quantifiable evaluation metrics.Review of the surveyed literature indicates the achievement of considerable maturity in the capacity of AI models to synthesize high-quality,aesthetically compelling anime visual images from textual prompts,alongside discernible progress in the generation of coherent narratives.However,achieving perfect long-form consistency,mitigating artifacts like flickering in video sequences,and enabling fine-grained artistic control remain critical ongoing challenges.Building upon these advancements,research efforts have increasingly pivoted towards the synthesis of higher-dimensional content,such as video and three-dimensional assets,with recent studies demonstrating significant progress in this burgeoning field.Nevertheless,formidable challenges endure amidst these advancements.Foremost among these are the substantial computational exigencies requisite for training and deploying these sophisticated models,particularly pronounced in the realm of high-dimensional generation such as video synthesis.Additional persistent hurdles include maintaining spatial-temporal consistency across complex scenes and mitigating ethical considerations surrounding bias and the preservation of human creative autonomy.This research underscores the transformative potential and inherent complexities of AI-driven synergy within the creative industries.We posit that future research should be dedicated to the synergistic fusion of diffusion and autoregressive models,the integration of multimodal inputs,and the balanced consideration of ethical implications,particularly regarding bias and the preservation of human creative autonomy,thereby establishing a robust foundation for the advancement of anime creation and the broader landscape of AI-driven content generation.展开更多
The internal structures of cells as the basic units of life are a major wonder of the microscopic world.Cellular images provide an intriguing window to help explore and understand the composition and function of these...The internal structures of cells as the basic units of life are a major wonder of the microscopic world.Cellular images provide an intriguing window to help explore and understand the composition and function of these structures.Scientific imagery combined with artistic expression can further expand the potential of imaging in educational dissemination and interdisciplinary applications.展开更多
Deep learning has achieved great progress in image recognition,segmentation,semantic recognition and game theory.In this study,a latest deep learning model,a conditional diffusion model was adopted as a surrogate mode...Deep learning has achieved great progress in image recognition,segmentation,semantic recognition and game theory.In this study,a latest deep learning model,a conditional diffusion model was adopted as a surrogate model to predict the heat transfer during the casting process instead of numerical simulation.The conditional diffusion model was established and trained with the geometry shapes,initial temperature fields and temperature fields at t_(i) as the condition and random noise sampled from standard normal distribution as the input.The output was the temperature field at t_(i+1).Therefore,the temperature field at t_(i+1)can be predicted as the temperature field at t_(i) is known,and the continuous temperature fields of all the time steps can be predicted based on the initial temperature field of an arbitrary 2D geometry.A training set with 3022D shapes and their simulated temperature fields at different time steps was established.The accuracy for the temperature field for a single time step reaches 97.7%,and that for continuous time steps reaches 69.1%with the main error actually existing in the sand mold.The effect of geometry shape and initial temperature field on the prediction accuracy was investigated,the former achieves better result than the latter because the former can identify casting,mold and chill by different colors in the input images.The diffusion model has proved the potential as a surrogate model for numerical simulation of the casting process.展开更多
Multi-scale problems in Computational Fluid Dynamics(CFD)often require numerous simulations across various design parameters.Using a fixed mesh for all cases may fail to capture critical physical features.Moving mesh ...Multi-scale problems in Computational Fluid Dynamics(CFD)often require numerous simulations across various design parameters.Using a fixed mesh for all cases may fail to capture critical physical features.Moving mesh adaptation provides an optimal resource allocation to obtain high-resolution flow-fields on low-resolution meshes.However,most existing methods require manual experience and the flow posteriori information poses great challenges to practical applications.In addition,generating adaptive meshes directly from design parameters is difficult due to highly nonlinear relationships.The diffusion model is currently the most popular model in generative tasks that integrates the diffusion principle into deep learning to capture the complex nonlinear correlations.A dual diffusion framework,Para2Mesh,is proposed to predict the adaptive meshes from design parameters by exploiting the robust data distribution learning ability of the diffusion model.Through iterative denoising,the proposed dual networks accurately reconstruct the flow-field to provide flow features as supervised information,and then achieve rapid and reliable mesh movement.Experiments in CFD scenarios demonstrate that Para2Mesh predicts similar meshes directly from design parameters with much higher efficiency than traditional method.It could become a real-time adaptation tool to assist engineering design and optimization,providing a promising solution for high-resolution flow-field analysis.展开更多
Finding suitable initial noise that retains the original image’s information is crucial for image-to-image(I2I)translation using text-to-image(T2I)diffusion models.A common approach is to add random noise directly to...Finding suitable initial noise that retains the original image’s information is crucial for image-to-image(I2I)translation using text-to-image(T2I)diffusion models.A common approach is to add random noise directly to the original image,as in SDEdit.However,we have observed that this can result in“semantic discrepancy”issues,wherein T2I diffusion models misinterpret the semantic relationships and generate content not present in the original image.We identify that the noise introduced by SDEdit disrupts the semantic integrity of the image,leading to unintended associations between unrelated regions after U-Net upsampling.Building on the widely-used latent diffusion model,Stable Diffusion,we propose a training-free,plugand-play method to alleviate semantic discrepancy and enhance the fidelity of the translated image.By leveraging the deterministic nature of denoising diffusion implicit models(DDIMs)inversion,we correct the erroneous features and correlations from the original generative process with accurate ones from DDIM inversion.This approach alleviates semantic discrepancy and surpasses recent DDIM-inversion-based methods such as PnP with fewer priors,achieving a speedup of 11.2 times in experiments conducted on COCO,ImageNet,and ImageNet-R datasets across multiple I2I translation tasks.展开更多
Objective This study aimed to explore a novel method that integrates the segmentation guidance classification and the dif-fusion model augmentation to realize the automatic classification for tibial plateau fractures(...Objective This study aimed to explore a novel method that integrates the segmentation guidance classification and the dif-fusion model augmentation to realize the automatic classification for tibial plateau fractures(TPFs).Methods YOLOv8n-cls was used to construct a baseline model on the data of 3781 patients from the Orthopedic Trauma Center of Wuhan Union Hospital.Additionally,a segmentation-guided classification approach was proposed.To enhance the dataset,a diffusion model was further demonstrated for data augmentation.Results The novel method that integrated the segmentation-guided classification and diffusion model augmentation sig-nificantly improved the accuracy and robustness of fracture classification.The average accuracy of classification for TPFs rose from 0.844 to 0.896.The comprehensive performance of the dual-stream model was also significantly enhanced after many rounds of training,with both the macro-area under the curve(AUC)and the micro-AUC increasing from 0.94 to 0.97.By utilizing diffusion model augmentation and segmentation map integration,the model demonstrated superior efficacy in identifying SchatzkerⅠ,achieving an accuracy of 0.880.It yielded an accuracy of 0.898 for SchatzkerⅡandⅢand 0.913 for SchatzkerⅣ;for SchatzkerⅤandⅥ,the accuracy was 0.887;and for intercondylar ridge fracture,the accuracy was 0.923.Conclusion The dual-stream attention-based classification network,which has been verified by many experiments,exhibited great potential in predicting the classification of TPFs.This method facilitates automatic TPF assessment and may assist surgeons in the rapid formulation of surgical plans.展开更多
With the growing demand for high-precision flow field simulations in computational science and engineering,the super-resolution reconstruction of physical fields has attracted considerable research interest.However,tr...With the growing demand for high-precision flow field simulations in computational science and engineering,the super-resolution reconstruction of physical fields has attracted considerable research interest.However,tradi-tional numerical methods often entail high computational costs,involve complex data processing,and struggle to capture fine-scale high-frequency details.To address these challenges,we propose an innovative super-resolution reconstruction framework that integrates a Fourier neural operator(FNO)with an enhanced diffusion model.The framework employs an adaptively weighted FNO to process low-resolution flow field inputs,effectively capturing global dependencies and high-frequency features.Furthermore,a residual-guided diffusion model is introduced to further improve reconstruction performance.This model uses a Markov chain for noise injection in phys-ical fields and integrates a reverse denoising procedure,efficiently solved by an adaptive time-step ordinary differential equation solver,thereby ensuring both stability and computational efficiency.Experimental results demonstrate that the proposed framework significantly outperforms existing methods in terms of accuracy and efficiency,offering a promising solution for fine-grained data reconstruction in scientific simulations.展开更多
Obtaining unsteady hydrodynamic performance is of great significance for seaplane design.Common methods for obtaining unsteady hydrodynamic performance data include tank test and Computational Fluid Dynamics(CFD)numer...Obtaining unsteady hydrodynamic performance is of great significance for seaplane design.Common methods for obtaining unsteady hydrodynamic performance data include tank test and Computational Fluid Dynamics(CFD)numerical simulation,which are costly and time-consuming.Therefore,it is necessary to obtain unsteady hydrodynamic performance in a low-cost and high-precision manner.Due to the strong nonlinearity,complex data distribution,and temporal characteristics of unsteady hydrodynamic performance,the prediction of it is challenging.This paper proposes a Temporal Convolutional Diffusion Model(TCDM)for predicting the unsteady hydrodynamic performance of seaplanes given design parameters.Under the framework of a classifier-free guided diffusion model,TCDM learns the distribution patterns of unsteady hydrodynamic performance data with the designed denoising module based on temporal convolutional network and captures the temporal features of unsteady hydrodynamic performance data.Using CFD simulation data,the proposed method is compared with the alternative methods to demonstrate its accuracy and generalization.This paper provides a method that enables the rapid and accurate prediction of unsteady hydrodynamic performance data,expecting to shorten the design cycle of seaplanes.展开更多
Multi-instance image generation remains a challenging task in the field of computer vision.While existing diffusionmodels demonstrate impressive fidelity in image generation,they often struggle with precisely controll...Multi-instance image generation remains a challenging task in the field of computer vision.While existing diffusionmodels demonstrate impressive fidelity in image generation,they often struggle with precisely controlling each object’s shape,pose,and size.Methods like layout-to-image and mask-to-image provide spatial guidance but frequently suffer from object shape distortion,overlaps,and poor consistency,particularly in complex scenes with multiple objects.To address these issues,we introduce PolyDiffusion,a contour-based diffusion framework that encodes each object’s contour as a boundary-coordinate sequence,decoupling object shapes and positions.This approach allows for better control over object geometry and spatial positioning,which is critical for achieving high-quality multiinstance generation.We formulate the training process as a multi-objective optimization problem,balancing three key objectives:a denoising diffusion loss to maintain overall image fidelity,a cross-attention contour alignment loss to ensure precise shape adherence,and a reward-guided denoising objective that minimizes the Fréchet distance to real images.In addition,the Object Space-Aware Attention module fuses contour tokens with visual features,while a prior-guided fusion mechanism utilizes inter-object spatial relationships and class semantics to enhance consistency across multiple objects.Experimental results on benchmark datasets such as COCO-Stuff and VOC-2012 demonstrate that PolyDiffusion significantly outperforms existing layout-to-image and mask-to-image methods,achieving notable improvements in both image quality and instance-level segmentation accuracy.The implementation of Poly Diffusion is available at https://github.com/YYYYYJS/PolyDiffusion(accessed on 06 August 2025).展开更多
基金funding from the European Commission by the Ruralities project(grant agreement no.101060876).
文摘In this paper,we propose a new privacy-aware transmission scheduling algorithm for 6G ad hoc networks.This system enables end nodes to select the optimum time and scheme to transmit private data safely.In 6G dynamic heterogeneous infrastructures,unstable links and non-uniform hardware capabilities create critical issues regarding security and privacy.Traditional protocols are often too computationally heavy to allow 6G services to achieve their expected Quality-of-Service(QoS).As the transport network is built of ad hoc nodes,there is no guarantee about their trustworthiness or behavior,and transversal functionalities are delegated to the extreme nodes.However,while security can be guaranteed in extreme-to-extreme solutions,privacy cannot,as all intermediate nodes still have to handle the data packets they are transporting.Besides,traditional schemes for private anonymous ad hoc communications are vulnerable against modern intelligent attacks based on learning models.The proposed scheme fulfills this gap.Findings show the probability of a successful intelligent attack reduces by up to 65%compared to ad hoc networks with no privacy protection strategy when used the proposed technology.While congestion probability can remain below 0.001%,as required in 6G services.
基金supported by the National Natural Science Foundation of China(Grant No.72161034).
文摘Human motion modeling is a core technology in computer animation,game development,and humancomputer interaction.In particular,generating natural and coherent in-between motion using only the initial and terminal frames remains a fundamental yet unresolved challenge.Existing methods typically rely on dense keyframe inputs or complex prior structures,making it difficult to balance motion quality and plausibility under conditions such as sparse constraints,long-term dependencies,and diverse motion styles.To address this,we propose a motion generation framework based on a frequency-domain diffusion model,which aims to better model complex motion distributions and enhance generation stability under sparse conditions.Our method maps motion sequences to the frequency domain via the Discrete Cosine Transform(DCT),enabling more effective modeling of low-frequency motion structures while suppressing high-frequency noise.A denoising network based on self-attention is introduced to capture long-range temporal dependencies and improve global structural awareness.Additionally,a multi-objective loss function is employed to jointly optimize motion smoothness,pose diversity,and anatomical consistency,enhancing the realism and physical plausibility of the generated sequences.Comparative experiments on the Human3.6M and LaFAN1 datasets demonstrate that our method outperforms state-of-the-art approaches across multiple performance metrics,showing stronger capabilities in generating intermediate motion frames.This research offers a new perspective and methodology for human motion generation and holds promise for applications in character animation,game development,and virtual interaction.
基金the National Natural Science Foundation of China(Grant No.:52508343)the Fundamental Research Funds for the Central Universities(Grant No.:B250201004).
文摘Crack detection accuracy in computer vision is often constrained by limited annotated datasets.Although Generative Adversarial Networks(GANs)have been applied for data augmentation,they frequently introduce blurs and artifacts.To address this challenge,this study leverages Denoising Diffusion Probabilistic Models(DDPMs)to generate high-quality synthetic crack images,enriching the training set with diverse and structurally consistent samples that enhance the crack segmentation.The proposed framework involves a two-stage pipeline:first,DDPMs are used to synthesize high-fidelity crack images that capture fine structural details.Second,these generated samples are combined with real data to train segmentation networks,thereby improving accuracy and robustness in crack detection.Compared with GAN-based approaches,DDPM achieved the best fidelity,with the highest Structural Similarity Index(SSIM)(0.302)and lowest Learned Perceptual Image Patch Similarity(LPIPS)(0.461),producing artifact-free images that preserve fine crack details.To validate its effectiveness,six segmentation models were tested,among which LinkNet consistently achieved the best performance,excelling in both region-level accuracy and structural continuity.Incorporating DDPM-augmented data further enhanced segmentation outcomes,increasing F1 scores by up to 1.1%and IoU by 1.7%,while also improving boundary alignment and skeleton continuity compared with models trained on real images alone.Experiments with varying augmentation ratios showed consistent improvements,with F1 rising from 0.946(no augmentation)to 0.957 and IoU from 0.897 to 0.913 at the highest ratio.These findings demonstrate the effectiveness of diffusion-based augmentation for complex crack detection in structural health monitoring.
基金supported by the Cultivation Program for Major Scientific Research Projects of Harbin Institute of Technology(ZDXMPY20180109).
文摘Scalable simulation leveraging real-world data plays an essential role in advancing autonomous driving,owing to its efficiency and applicability in both training and evaluating algorithms.Consequently,there has been increasing attention on generating highly realistic and consistent driving videos,particularly those involving viewpoint changes guided by the control commands or trajectories of ego vehicles.However,current reconstruction approaches,such as Neural Radiance Fields and 3D Gaussian Splatting,frequently suffer from limited generalization and depend on substantial input data.Meanwhile,2D generative models,though capable of producing unknown scenes,still have room for improvement in terms of coherence and visual realism.To overcome these challenges,we introduce GenScene,a world model that synthesizes front-view driving videos conditioned on trajectories.A new temporal module is presented to improve video consistency by extracting the global context of each frame,calculating relationships of frames using these global representations,and fusing frame contexts accordingly.Moreover,we propose an innovative attention mechanism that computes relations of pixels within each frame and pixels in the corresponding window range of the initial frame.Extensive experiments show that our approach surpasses various state-of-the-art models in driving video generation,and the introduced modules contribute significantly to model performance.This work establishes a new paradigm for goal-oriented video synthesis in autonomous driving,which facilitates on-demand simulation to expedite algorithm development.
基金supported by the National Science and Technology Council,Taiwan,under grant no.NSTC 114-2221-E-197-005-MY3.
文摘With the development of technology,diffusion model-based solvers have shown significant promise in solving Combinatorial Optimization(CO)problems,particularly in tackling Non-deterministic Polynomial-time hard(NP-hard)problems such as the Traveling Salesman Problem(TSP).However,existing diffusion model-based solvers typically employ a fixed,uniform noise schedule(e.g.,linear or cosine annealing)across all training instances,failing to fully account for the unique characteristics of each problem instance.To address this challenge,we present GraphGuided Diffusion Solvers(GGDS),an enhanced method for improving graph-based diffusion models.GGDS leverages Graph Neural Networks(GNNs)to capture graph structural information embedded in node coordinates and adjacency matrices,dynamically adjusting the noise levels in the diffusion model.This study investigates the TSP by examining two distinct time-step noise generation strategies:cosine annealing and a Neural Network(NN)-based approach.We evaluate their performance across different problem scales,particularly after integrating graph structural information.Experimental results indicate that GGDS outperforms previous methods with average performance improvements of 18.7%,6.3%,and 88.7%on TSP-500,TSP-100,and TSP-50,respectively.Specifically,GGDS demonstrates superior performance on TSP-500 and TSP-50,while its performance on TSP-100 is either comparable to or slightly better than that of previous methods,depending on the chosen noise schedule and decoding strategy.
基金supported by a grant from the National Natural Science Foundation of China (NSFC) No. 81070826/30872886/30400497Sponsored by Shanghai Rising-Star Program No. 09QA1403700+1 种基金funded by Shanghai Leading Academic Discipline Project (Project Number: S30206)the Science and Technology Commission of Shanghai (08DZ2271100)
文摘Aim The purpose of this study was to develop a mathe-matical model to quantitatively describe the passive trans-port of macromolecules within dental biofilms. Methodology Fluorescently labeled dextrans with different molecular mass （3 kD,10 kD,40 kD,70 kD,2 000 kD） were used as a series of diffusion probes. Streptococcus mutans,Streptococcus sanguinis,Actinomyces naeslundii and Fusobacterium nucleatum were used as inocula for biofilm formation. The diffusion processes of different probes through the in vitro biofilm were recorded with a confocal laser microscope. Results Mathematical function of biofilm penetration was constructed on the basis of the inverse problem method. Based on this function,not only the relationship between average concentration of steady-state and molecule weights can be analyzed,but also that between penetrative time and molecule weights. Conclusion This can be used to predict the effective concentration and the penetrative time of anti-biofilm medicines that can diffuse through oral biofilm. Further-more,an improved model for large molecule is proposed by considering the exchange time at the upper boundary of the dental biofilm.
文摘In this study,the removal of monovalent and divalent cations,Nat,Kt,Mg2t,and Ca2t,in a diluted solution from Chott-El Jerid Lake,Tunisia,was investigated with the electrodialysis technique.The process was tested using two cation-exchange membranes:sulfonated polyether sulfone cross-linked with 10%hexamethylenediamine(HEXCl)and sulfonated polyether sulfone grafted with octylamine(S-PESOS).The commercially available membrane Nafion®was used for comparison.The results showed that Nafion®and S-PESOS membranes had similar removal behaviors,and the investigated cations were ranked in the following descending order in terms of their demineralization rates:Nat>Ca2t>Mg2t>Kt.Divalent cations were more effectively removed by HEXCl than by monovalent cations.The plots based on the WebereMorris model showed a strong linearity.This reveals that intra-particle diffusion was not the removal rate-determining step,and the removal process was controlled by two or more concurrent mechanisms.The Boyd plots did not pass through their origin,and the sole controlling step was determined by film-diffusion resistance,especially after a long period of electrodialysis.Additionally,a semi-empirical model was established to simulate the temporal variation of the treatment process,and the physical significance and values of model parameters were compared for the three membranes.The findings of this study indicate that HEXCl and S-PESOS membranes can be efficiently utilized for water softening,especially when effluents are highly loaded with calcium and magnesium ions.
基金supported by the National Natural Science Foundation of China(Nos.61906168,62202429 and 62272267)the Zhejiang Provincial Natural Science Foundation of China(No.LY23F020023)the Construction of Hubei Provincial Key Laboratory for Intelligent Visual Monitoring of Hydropower Projects(No.2022SDSJ01)。
文摘Accurately identifying building distribution from remote sensing images with complex background information is challenging.The emergence of diffusion models has prompted the innovative idea of employing the reverse denoising process to distill building distribution from these complex backgrounds.Building on this concept,we propose a novel framework,building extraction diffusion model(BEDiff),which meticulously refines the extraction of building footprints from remote sensing images in a stepwise fashion.Our approach begins with the design of booster guidance,a mechanism that extracts structural and semantic features from remote sensing images to serve as priors,thereby providing targeted guidance for the diffusion process.Additionally,we introduce a cross-feature fusion module(CFM)that bridges the semantic gap between different types of features,facilitating the integration of the attributes extracted by booster guidance into the diffusion process more effectively.Our proposed BEDiff marks the first application of diffusion models to the task of building extraction.Empirical evidence from extensive experiments on the Beijing building dataset demonstrates the superior performance of BEDiff,affirming its effectiveness and potential for enhancing the accuracy of building extraction in complex urban landscapes.
基金partially supported by the National Natural Science Foundation of China(62271485)the SDHS Science and Technology Project(HS2023B044)
文摘Imputation of missing data has long been an important topic and an essential application for intelligent transportation systems(ITS)in the real world.As a state-of-the-art generative model,the diffusion model has proven highly successful in image generation,speech generation,time series modelling etc.and now opens a new avenue for traffic data imputation.In this paper,we propose a conditional diffusion model,called the implicit-explicit diffusion model,for traffic data imputation.This model exploits both the implicit and explicit feature of the data simultaneously.More specifically,we design two types of feature extraction modules,one to capture the implicit dependencies hidden in the raw data at multiple time scales and the other to obtain the long-term temporal dependencies of the time series.This approach not only inherits the advantages of the diffusion model for estimating missing data,but also takes into account the multiscale correlation inherent in traffic data.To illustrate the performance of the model,extensive experiments are conducted on three real-world time series datasets using different missing rates.The experimental results demonstrate that the model improves imputation accuracy and generalization capability.
基金supported by the Yonsei University graduate school Department of Integrative Biotechnology.
文摘Recently,diffusion models have emerged as a promising paradigm for molecular design and optimization.However,most diffusion-based molecular generative models focus on modeling 2D graphs or 3D geom-etries,with limited research on molecular sequence diffusion models.The International Union of Pure and Applied Chemistry(IUPAC)names are more akin to chemical natural language than the simplified molecular input line entry system(SMILES)for organic compounds.In this work,we apply an IUPAC-guided conditional diffusion model to facilitate molecular editing from chemical natural language to chemical language(SMILES)and explore whether the pre-trained generative performance of diffusion models can be transferred to chemical natural language.We propose DiffIUPAC,a controllable molecular editing diffusion model that converts IUPAC names to SMILES strings.Evaluation results demonstrate that our model out-performs existing methods and successfully captures the semantic rules of both chemical languages.Chemical space and scaffold analysis show that the model can generate similar compounds with diverse scaffolds within the specified constraints.Additionally,to illustrate the model’s applicability in drug design,we conducted case studies in functional group editing,analogue design and linker design.
基金co-supported by the National Natural Science Foundation of China(Nos.61806219,61876189 and 61703426)the Young Talent Fund of University Association for Science and Technology in Shaanxi,China(Nos.20190108 and 20220106)the Innvation Talent Supporting Project of Shaanxi,China(No.2020KJXX-065)。
文摘Air target intent recognition holds significant importance in aiding commanders to assess battlefield situations and secure a competitive edge in decision-making.Progress in this domain has been hindered by challenges posed by imbalanced battlefield data and the limited robustness of traditional recognition models.Inspired by the success of diffusion models in addressing visual domain sample imbalances,this paper introduces a new approach that utilizes the Markov Transfer Field(MTF)method for time series data visualization.This visualization,when combined with the Denoising Diffusion Probabilistic Model(DDPM),effectively enhances sample data and mitigates noise within the original dataset.Additionally,a transformer-based model tailored for time series visualization and air target intent recognition is developed.Comprehensive experimental results,encompassing comparative,ablation,and denoising validations,reveal that the proposed method achieves a notable 98.86%accuracy in air target intent recognition while demonstrating exceptional robustness and generalization capabilities.This approach represents a promising avenue for advancing air target intent recognition.
基金supported by the National Natural Science Foundation of China(Grant No.62202210).
文摘The application of generative artificial intelligence(AI)is bringing about notable changes in anime creation.This paper surveys recent advancements and applications of diffusion and language models in anime generation,focusing on their demonstrated potential to enhance production efficiency through automation and personalization.Despite these benefits,it is crucial to acknowledge the substantial initial computational investments required for training and deploying these models.We conduct an in-depth survey of cutting-edge generative AI technologies,encompassing models such as Stable Diffusion and GPT,and appraise pivotal large-scale datasets alongside quantifiable evaluation metrics.Review of the surveyed literature indicates the achievement of considerable maturity in the capacity of AI models to synthesize high-quality,aesthetically compelling anime visual images from textual prompts,alongside discernible progress in the generation of coherent narratives.However,achieving perfect long-form consistency,mitigating artifacts like flickering in video sequences,and enabling fine-grained artistic control remain critical ongoing challenges.Building upon these advancements,research efforts have increasingly pivoted towards the synthesis of higher-dimensional content,such as video and three-dimensional assets,with recent studies demonstrating significant progress in this burgeoning field.Nevertheless,formidable challenges endure amidst these advancements.Foremost among these are the substantial computational exigencies requisite for training and deploying these sophisticated models,particularly pronounced in the realm of high-dimensional generation such as video synthesis.Additional persistent hurdles include maintaining spatial-temporal consistency across complex scenes and mitigating ethical considerations surrounding bias and the preservation of human creative autonomy.This research underscores the transformative potential and inherent complexities of AI-driven synergy within the creative industries.We posit that future research should be dedicated to the synergistic fusion of diffusion and autoregressive models,the integration of multimodal inputs,and the balanced consideration of ethical implications,particularly regarding bias and the preservation of human creative autonomy,thereby establishing a robust foundation for the advancement of anime creation and the broader landscape of AI-driven content generation.
基金supported by the Fundamental Research Funds for the Central Universities(No.226-2024-00038),China.
文摘The internal structures of cells as the basic units of life are a major wonder of the microscopic world.Cellular images provide an intriguing window to help explore and understand the composition and function of these structures.Scientific imagery combined with artistic expression can further expand the potential of imaging in educational dissemination and interdisciplinary applications.
基金sponsored by Tsinghua-Toyota Joint Research Fund
文摘Deep learning has achieved great progress in image recognition,segmentation,semantic recognition and game theory.In this study,a latest deep learning model,a conditional diffusion model was adopted as a surrogate model to predict the heat transfer during the casting process instead of numerical simulation.The conditional diffusion model was established and trained with the geometry shapes,initial temperature fields and temperature fields at t_(i) as the condition and random noise sampled from standard normal distribution as the input.The output was the temperature field at t_(i+1).Therefore,the temperature field at t_(i+1)can be predicted as the temperature field at t_(i) is known,and the continuous temperature fields of all the time steps can be predicted based on the initial temperature field of an arbitrary 2D geometry.A training set with 3022D shapes and their simulated temperature fields at different time steps was established.The accuracy for the temperature field for a single time step reaches 97.7%,and that for continuous time steps reaches 69.1%with the main error actually existing in the sand mold.The effect of geometry shape and initial temperature field on the prediction accuracy was investigated,the former achieves better result than the latter because the former can identify casting,mold and chill by different colors in the input images.The diffusion model has proved the potential as a surrogate model for numerical simulation of the casting process.
基金co-supported by the Aeronautical Science Foundation of China(Nos.2018ZA52002 and 2019ZA052011)。
文摘Multi-scale problems in Computational Fluid Dynamics(CFD)often require numerous simulations across various design parameters.Using a fixed mesh for all cases may fail to capture critical physical features.Moving mesh adaptation provides an optimal resource allocation to obtain high-resolution flow-fields on low-resolution meshes.However,most existing methods require manual experience and the flow posteriori information poses great challenges to practical applications.In addition,generating adaptive meshes directly from design parameters is difficult due to highly nonlinear relationships.The diffusion model is currently the most popular model in generative tasks that integrates the diffusion principle into deep learning to capture the complex nonlinear correlations.A dual diffusion framework,Para2Mesh,is proposed to predict the adaptive meshes from design parameters by exploiting the robust data distribution learning ability of the diffusion model.Through iterative denoising,the proposed dual networks accurately reconstruct the flow-field to provide flow features as supervised information,and then achieve rapid and reliable mesh movement.Experiments in CFD scenarios demonstrate that Para2Mesh predicts similar meshes directly from design parameters with much higher efficiency than traditional method.It could become a real-time adaptation tool to assist engineering design and optimization,providing a promising solution for high-resolution flow-field analysis.
基金supported in part by the National Natural Science Foundation of China(62176059)supported by The Pennsylvania State University.
文摘Finding suitable initial noise that retains the original image’s information is crucial for image-to-image(I2I)translation using text-to-image(T2I)diffusion models.A common approach is to add random noise directly to the original image,as in SDEdit.However,we have observed that this can result in“semantic discrepancy”issues,wherein T2I diffusion models misinterpret the semantic relationships and generate content not present in the original image.We identify that the noise introduced by SDEdit disrupts the semantic integrity of the image,leading to unintended associations between unrelated regions after U-Net upsampling.Building on the widely-used latent diffusion model,Stable Diffusion,we propose a training-free,plugand-play method to alleviate semantic discrepancy and enhance the fidelity of the translated image.By leveraging the deterministic nature of denoising diffusion implicit models(DDIMs)inversion,we correct the erroneous features and correlations from the original generative process with accurate ones from DDIM inversion.This approach alleviates semantic discrepancy and surpasses recent DDIM-inversion-based methods such as PnP with fewer priors,achieving a speedup of 11.2 times in experiments conducted on COCO,ImageNet,and ImageNet-R datasets across multiple I2I translation tasks.
基金supported by the National Natural Science Foundation of China(Nos.81974355 and 82172524)Key Research and Development Program of Hubei Province(No.2021BEA161)+2 种基金National Innovation Platform Development Program(No.2020021105012440)Open Project Funding of the Hubei Key Laboratory of Big Data Intelligent Analysis and Application,Hubei University(No.2024BDIAA03)Free Innovation Preliminary Research Fund of Wuhan Union Hospital(No.2024XHYN047).
文摘Objective This study aimed to explore a novel method that integrates the segmentation guidance classification and the dif-fusion model augmentation to realize the automatic classification for tibial plateau fractures(TPFs).Methods YOLOv8n-cls was used to construct a baseline model on the data of 3781 patients from the Orthopedic Trauma Center of Wuhan Union Hospital.Additionally,a segmentation-guided classification approach was proposed.To enhance the dataset,a diffusion model was further demonstrated for data augmentation.Results The novel method that integrated the segmentation-guided classification and diffusion model augmentation sig-nificantly improved the accuracy and robustness of fracture classification.The average accuracy of classification for TPFs rose from 0.844 to 0.896.The comprehensive performance of the dual-stream model was also significantly enhanced after many rounds of training,with both the macro-area under the curve(AUC)and the micro-AUC increasing from 0.94 to 0.97.By utilizing diffusion model augmentation and segmentation map integration,the model demonstrated superior efficacy in identifying SchatzkerⅠ,achieving an accuracy of 0.880.It yielded an accuracy of 0.898 for SchatzkerⅡandⅢand 0.913 for SchatzkerⅣ;for SchatzkerⅤandⅥ,the accuracy was 0.887;and for intercondylar ridge fracture,the accuracy was 0.923.Conclusion The dual-stream attention-based classification network,which has been verified by many experiments,exhibited great potential in predicting the classification of TPFs.This method facilitates automatic TPF assessment and may assist surgeons in the rapid formulation of surgical plans.
基金supported by the National Natural Science Foundation of China(Grant Nos.42005003 and 41475094)National Key R&D Program of China(Grant No.2018YFC1506704).
文摘With the growing demand for high-precision flow field simulations in computational science and engineering,the super-resolution reconstruction of physical fields has attracted considerable research interest.However,tradi-tional numerical methods often entail high computational costs,involve complex data processing,and struggle to capture fine-scale high-frequency details.To address these challenges,we propose an innovative super-resolution reconstruction framework that integrates a Fourier neural operator(FNO)with an enhanced diffusion model.The framework employs an adaptively weighted FNO to process low-resolution flow field inputs,effectively capturing global dependencies and high-frequency features.Furthermore,a residual-guided diffusion model is introduced to further improve reconstruction performance.This model uses a Markov chain for noise injection in phys-ical fields and integrates a reverse denoising procedure,efficiently solved by an adaptive time-step ordinary differential equation solver,thereby ensuring both stability and computational efficiency.Experimental results demonstrate that the proposed framework significantly outperforms existing methods in terms of accuracy and efficiency,offering a promising solution for fine-grained data reconstruction in scientific simulations.
基金supported by the Aeronautical Science Foundation of China(Nos.2018ZA52002,2019ZA052011)the National Natural Science Foundation of China(No.12472236).
文摘Obtaining unsteady hydrodynamic performance is of great significance for seaplane design.Common methods for obtaining unsteady hydrodynamic performance data include tank test and Computational Fluid Dynamics(CFD)numerical simulation,which are costly and time-consuming.Therefore,it is necessary to obtain unsteady hydrodynamic performance in a low-cost and high-precision manner.Due to the strong nonlinearity,complex data distribution,and temporal characteristics of unsteady hydrodynamic performance,the prediction of it is challenging.This paper proposes a Temporal Convolutional Diffusion Model(TCDM)for predicting the unsteady hydrodynamic performance of seaplanes given design parameters.Under the framework of a classifier-free guided diffusion model,TCDM learns the distribution patterns of unsteady hydrodynamic performance data with the designed denoising module based on temporal convolutional network and captures the temporal features of unsteady hydrodynamic performance data.Using CFD simulation data,the proposed method is compared with the alternative methods to demonstrate its accuracy and generalization.This paper provides a method that enables the rapid and accurate prediction of unsteady hydrodynamic performance data,expecting to shorten the design cycle of seaplanes.
基金supported in part by the Scientific Research Fund of National Natural Science Foundation of China(Grant No.62372168)the Hunan Provincial Natural Science Foundation of China(Grant No.2023JJ30266)+2 种基金the Research Project on teaching reform in Hunan province(No.HNJG-2022-0791)the Hunan University of Science and Technology(No.2022-44-8)the National Social Science Funds of China(19BZX044).
文摘Multi-instance image generation remains a challenging task in the field of computer vision.While existing diffusionmodels demonstrate impressive fidelity in image generation,they often struggle with precisely controlling each object’s shape,pose,and size.Methods like layout-to-image and mask-to-image provide spatial guidance but frequently suffer from object shape distortion,overlaps,and poor consistency,particularly in complex scenes with multiple objects.To address these issues,we introduce PolyDiffusion,a contour-based diffusion framework that encodes each object’s contour as a boundary-coordinate sequence,decoupling object shapes and positions.This approach allows for better control over object geometry and spatial positioning,which is critical for achieving high-quality multiinstance generation.We formulate the training process as a multi-objective optimization problem,balancing three key objectives:a denoising diffusion loss to maintain overall image fidelity,a cross-attention contour alignment loss to ensure precise shape adherence,and a reward-guided denoising objective that minimizes the Fréchet distance to real images.In addition,the Object Space-Aware Attention module fuses contour tokens with visual features,while a prior-guided fusion mechanism utilizes inter-object spatial relationships and class semantics to enhance consistency across multiple objects.Experimental results on benchmark datasets such as COCO-Stuff and VOC-2012 demonstrate that PolyDiffusion significantly outperforms existing layout-to-image and mask-to-image methods,achieving notable improvements in both image quality and instance-level segmentation accuracy.The implementation of Poly Diffusion is available at https://github.com/YYYYYJS/PolyDiffusion(accessed on 06 August 2025).
