State-of-the-art text-to-image diffusion models (DMs) achieve remarkable quality, yet their massive parameter scale (8-11B) poses significant challenges for inferences on resource-constrained devices. In this paper, we present HierarchicalPrune, a novel compression framework grounded in a key observation: DM blocks exhibit distinct functional hierarchies, where early blocks establish semantic structures while later blocks handle texture refinements. HierarchicalPrune synergistically combines three techniques: (1) Hierarchical Position Pruning, which identifies and removes less essential later blocks based on position hierarchy; (2) Positional Weight Preservation, which systematically protects early model portions that are essential for semantic structural integrity; and (3) Sensitivity-Guided Distillation, which adjusts knowledge-transfer intensity based on our discovery of block-wise sensitivity variations. As a result, our framework brings billion-scale diffusion models into a range more suitable for on-device inference, while preserving the quality of the output images. Specifically, combined with INT4 weight quantisation, HierarchicalPrune achieves 77.5-80.4% memory footprint reduction (e.g., from 15.8 GB to 3.2 GB) and 27.9-38.0% latency reduction, measured on server and consumer grade GPUs, with the minimum drop of 2.6% in GenEval score and 7% in HPSv2 score compared to the original model. Finally, our comprehensive user study with 85 participants demonstrates that HierarchicalPrune maintains perceptual quality comparable to the original model while significantly outperforming prior works.
论文检索
输入标题、作者或关键词，从 9,256 篇学术成果中精准定位
Diffusion models deliver state-of-the-art generative performance across diverse modalities but remain computationally expensive due to their inherently iterative sampling process. Existing training-free acceleration methods typically improve numerical solvers for the reverse-time ODE, yet their effectiveness is fundamentally constrained by the inefficiency of the underlying sampling trajectories. We propose A-FloPS (Adaptive Flow Path Sampler), a principled, training-free framework that reparameterizes the sampling trajectory of any pre-trained diffusion model into a flow-matching form and augments it with an adaptive velocity decomposition. The reparameterization analytically maps diffusion scores to flow-compatible velocities, yielding integration-friendly trajectories without retraining. The adaptive mechanism further factorizes the velocity field into a linear drift term and a residual component whose temporal variation is actively suppressed, restoring the accuracy benefits of high-order integration even in extremely low-NFE regimes. Extensive experiments on conditional image generation and text-to-image synthesis show that A-FloPS consistently outperforms state-of-the-art training-free samplers in both sample quality and efficiency. Notably, with as few as 5 function evaluations, A-FloPS achieves substantially lower FID and generates sharper, more coherent images. The adaptive mechanism also improves native flow-based generative models, underscoring its generality. These results position A-FloPS as a versatile and effective solution for high-quality, low-latency generative modeling.
Diffusion models have achieved remarkable success in image and video generation. However, their inherently multiple step inference process imposes substantial computational overhead, hindering real-world deployment. Accelerating diffusion models is therefore essential, yet determining how to combine multiple model acceleration techniques remains a significant challenge. To address this issue, we introduce a framework driven by large language models (LLMs) for automated acceleration code generation and evaluation. First, we present DiffBench, a comprehensive benchmark that implements a three stage automated evaluation pipeline across diverse diffusion architectures, optimization combinations and deployment scenarios. Second, we propose DiffAgent, an agent that generates optimal acceleration strategies and codes for arbitrary diffusion models. DiffAgent employs a closed-loop workflow in which a planning component and a debugging component iteratively refine the output of a code generation component, while a genetic algorithm extracts performance feedback from the execution environment to guide subsequent code refinements. We provide a detailed explanation of the DiffBench construction and the design principles underlying DiffAgent. Extensive experiments show that DiffBench offers a thorough evaluation of generated codes and that DiffAgent significantly outperforms existing LLMs in producing effective diffusion acceleration strategies.
Diff-V2M: A Hierarchical Conditional Diffusion Model with Explicit Rhythmic Modeling for Video-to-Music Generation
PDF ↗Video-to-music (V2M) generation aims to create music that aligns with visual content. However, two main challenges persist in existing methods: (1) the lack of explicit rhythm modeling hinders audiovisual temporal alignments; (2) effectively integrating various visual features to condition music generation remains non-trivial. To address these issues, we propose Diff-V2M, a general V2M framework based on a hierarchical conditional diffusion model, comprising two core components: visual feature extraction and conditional music generation. For rhythm modeling, we begin by evaluating several rhythmic representations, including low-resolution mel-spectrograms, tempograms, and onset detection functions (ODF), and devise a rhythmic predictor to infer them directly from videos. To ensure contextual and affective coherence, we also extract semantic and emotional features. All features are incorporated into the generator via a hierarchical cross-attention mechanism, where emotional features shape the affective tone via the first layer, while semantic and rhythmic features are fused in the second cross-attention layer. To enhance feature integration, we introduce timestep-aware fusion strategies, including feature-wise linear modulation (FiLM) and weighted fusion, allowing the model to adaptively balance semantic and rhythmic cues throughout the diffusion process. Extensive experiments identify low-resolution ODF as a more effective signal for modeling musical rhythm and demonstrate that Diff-V2M outperforms existing models on both in-domain and out-of-domain datasets, achieving state-of-the-art performance in terms of objective metrics and subjective comparisons.
Diffusion models have achieved impressive generative performance across diverse domains such as image, video, and scientific data generation. However, fine-tuning these models for new tasks remains challenging due to their large scale, architectural diversity, and high sensitivity to hyperparameters—particularly learning rates. In this work, we propose Wasserstein-Aware Transfer (WAT), a principled and effective fine-tuning strategy grounded in diffusion trajectory analysis and optimal transport theory. Our key insight is that the distributional discrepancies between diffusion trajectories from different datasets decrease progressively over time and converge near the noise end. Based on this observation, we introduce a class-wise matching mechanism that minimizes the Wasserstein distance between class distributions of source and target datasets. This enables alignment at the class level without modifying the standard fine-tuning pipeline. To further enhance knowledge retention, we propose a novel sampling strategy that linearly combines class-conditional outputs from both pretrained and fine-tuned models. This method is simple yet effective, requiring negligible computational overhead while preserving domain-specific and generalizable knowledge. Extensive experiments across seven diverse benchmarks demonstrate that WAT reliably enhances generation quality under distribution shifts, outperforming competitive baselines. These results underscore its robustness and affirm the potential of optimal transport as a principled basis for knowledge transfer in diffusion models.
Conditional molecular generation, aiming to generate 2D and 3D molecules that satisfy given properties, has achieved remarkable progress, thanks to the advances in deep generative models such as graph diffusion. However, existing methods generally assume that the given conditions for training and testing are consistent, failing to handle the realistic challenge when there exist distribution shifts between training and testing conditions. Invariant learning is a mainstream paradigm for addressing distribution shifts, but fusing invariant learning principles with conditional molecular generation faces three core challenges: (1) existing invariant learning methods focus on discriminative tasks and cannot be directly adapted to molecule generative tasks; (2) how to distinguish between invariant subgraph and variant subgraph of a molecule graph, which is treated as an integrated input; (3) how to fuse invariant subgraphs, variant subgraphs, and property conditions for effective generation. To tackle these challenges, we propose Invariant Conditional MOLecular generation (IC-MOL), a framework that combines invariant learning with graph diffusion to improve the generalization ability of conditional molecular generation under distribution shifts. Specifically, we first disentangle molecular graphs into invariant and variant subgraphs while maintaining SE(3) equivariance, an important inductive bias for molecular generation. On this basis, we further design a two-phase graph diffusion generation model. In the first phase, we generate an invariant molecular consistent with the target property. In the second phase, we propose a cross-attention mechanism to fuse variant subgraph representations and property conditions to guide the generation of complete molecules while maintaining property alignment. Extensive experiments on the benchmark dataset show that IC-MOL consistently outperforms state-of-the-art baselines across six property conditions under distribution shifts.
Venom: Liquid Diffusion-Guided Gradient Inversion for Breaking Differential Privacy in Federated Learning
PDF ↗Gradient perturbation mechanisms, such as differential privacy (DP), aim to defend against gradient inversion attacks (GIA) by injecting noise into the shared gradients. Recent studies have shown that DP-based defenses lack robustness against advanced GIAs. However, existing gradient inversion methods typically rely on iterative refinement and assume static noise, resulting in low efficiency and limited reconstruction fidelity under high-noise conditions. In this paper, we propose Venom, a novel gradient inversion attack method based on a liquid diffusion mechanism. Venom reconstructs private data directly from DP-protected gradients without requiring any prior knowledge of the noise distribution. Specifically, we design a Structural Prior Extraction (SPE) module that analytically extracts deep feature representations from perturbed gradients through energy-based aggregation, enabling stable pre-reconstruction of users' latent data features. We further introduce a Diffusion-driven Liquid Recovery Network (Diff-LRN) for high-fidelity image reconstruction. Unlike traditional diffusion models that rely on iterative sampling with predefined noise schedules, Diff-LRN performs deterministic single-step reconstruction using adaptive liquid neural dynamics to handle spatially heterogeneous noise patterns. Experiments across four benchmarks demonstrate that Venom achieves a speedup of up to 38,343× over state-of-the-art attacks while maintaining high reconstruction fidelity under strong DP settings. These results challenge prevailing assumptions about DP robustness and underscore the need for more resilient privacy-preserving mechanisms in federated learning.
Can we train a 3D molecule generator using data from dense regions to generate samples in sparse regions? This challenge can be framed as an out-of-distribution (OOD) generation problem. While prior research on OOD generation predominantly targets property shifts, structural shifts, such as differences in molecular scaffolds or functional groups, represent an equally critical source of distributional shifts. This work introduces the Geometric OOD Diffusion Model (GODD), a novel diffusion-based framework that enables training on data-abundant molecular distributions while generalizing to data-scarce distributions under distributional structural shifts. Central to our approach is a designated equivariant asymmetric autoencoder to capture distributional structural priors. The asymmetric design allows the model to generalize to unseen structural variations by capturing distributional priors representing distinct distributions. The encoded structural-grained priors guide generation toward sparse regions without requiring explicit training on such data. Evaluated across standard benchmarks encompassing OOD structural shifts (e.g., scaffolds, rings), GODD achieves an improvement of 12.6% in success rate, defined based on molecular validity, uniqueness, and novelty. Furthermore, the framework demonstrates promising performance and generalization on canonical fragment-based drug design tasks, highlighting its utility in learning-based molecular discovery.
A Stage-Aware Mixture of Experts Framework for Neurodegenerative Disease Progression Modelling
PDF ↗The long-term progression of neurodegenerative diseases is commonly conceptualized as a spatiotemporal diffusion process that consists of a graph diffusion process across the structural brain connectome and a localized reaction process within brain regions. However, modeling this progression remains challenging due to 1) the scarcity of longitudinal data obtained through irregular and infrequent subject visits and 2) the complex interplay of pathological mechanisms across brain regions and disease stages, where traditional models assume fixed mechanisms throughout disease progression. To address these limitations, we propose a novel stage-aware Mixture of Experts (MoE) framework that explicitly models how different contributing mechanisms dominate at different disease stages through time-dependent expert weighting. This architecture is a key innovation designed to maximize the utility of small datasets and provide interpretable insights into disease etiology. Data-wise, we utilize an iterative dual optimization method to properly estimate the temporal position of individual observations, constructing a cohort-level progression trajectory from irregular snapshots. Model-wise, we enhance the spatial component with an inhomogeneous graph neural diffusion model (IGND) that allows diffusivity to vary based on node states and time, providing more flexible representations of brain networks. We also introduce a localized neural reaction module to capture complex dynamics beyond standard processes.The resulting IGND-MoE model dynamically integrates these components across temporal states, offering a principled way to understand how stage-specific pathological mechanisms contribute to progression. When used to model tau pathology propagation in human brains, IGND-MoE outperforms purely pathophysiological and purely neural baselines in long-term prediction accuracy. Moreover, its stage-wise weights yield novel clinical insights that align with literature, suggesting that graph-related processes are more influential at early stages, while other unknown physical processes become dominant later on. Our findings highlight the necessity of designing hybrid and expert-constrained models that account for the evolving nature of neurodegenerative processes.
Recent advances in diffusion models have achieved remarkable success in isolated computer vision tasks such as text-to-image generation, depth estimation, and optical flow. However, these models are often restricted by a ``single-task-single-model'' paradigm, severely limiting their generalizability and scalability in multi-task scenarios. Motivated by the cross-domain generalization ability of large language models, we propose a universal visual perception framework based on flow matching that can generate diverse visual representations across multiple tasks. Our approach formulates the process as a universal flow-matching problem from image patch tokens to task-specific representations rather than an independent generation or regression problem. By leveraging a strong self-supervised foundation model as the anchor and introducing a multi-scale, circular task embedding mechanism, our method learns a universal velocity field to bridge the gap between heterogeneous tasks, supporting efficient and flexible representation transfer. Extensive experiments on classification, detection, segmentation, depth estimation, and image-text retrieval demonstrate that our model achieves competitive performance in both zero-shot and fine-tuned settings, outperforming prior generalist and several specialist models. Ablation studies further validate the robustness, scalability, and generalization of our framework. Our work marks a significant step towards general-purpose visual perception, providing a solid foundation for future research in universal vision modeling.
Time series generation is essential for advancing data-driven modeling and decision-making across a wide range of domains. However, existing approaches primarily focus on global patterns, often failing to capture local key patterns such as abrupt changes or anomalies. These key patterns are crucial for interpretability and operational decision making, as they frequently represent intervention points with significant real-world impact. To bridge this gap, we propose Key Prototypes-Guided Diffusion (K-ProtoDiff) for time series generation , a new model that learns the global data distribution while preserving localized key patterns critical for temporal dynamics. In K-ProtoDiff, we first derive time series prototype representations through adaptive self-supervised learning. Then, a key prototype assignment module is used to extract prototype weights, forming key prototype-aware representations that serve as conditional guidance for generation. During sampling, to further enhance the fidelity of key patterns during the denoising process, we propose Reflection Sampling (R-Sampling), a step-wise refinement strategy that encourages the reverse trajectory to better align with key prototype constraints. Experiments on nine real-world datasets demonstrate that K-ProtoDiff significantly outperforms state-of-the-art baselines in key pattern retention, achieving an average 77.6% improvement in key pattern preservation.
Diffusion models have gained prominence as powerful generative tools for solving inverse problems due to their ability to model complex data distributions. However, existing methods typically rely on complete knowledge of the forward observation process to compute gradients for guided sampling, limiting their applicability in scenarios where such information is unavailable. In this work, we introduce *Constrained Particle Seeking (CPS)*, a novel gradient-free approach that leverages all candidate particle information to actively search for the optimal particle while incorporating constraints aligned with high-density regions of the unconditional prior. Unlike previous methods that passively select promising candidates, CPS reformulates the inverse problem as a constrained optimization task, enabling more flexible and efficient particle seeking. We demonstrate that CPS can effectively solve both image and scientific inverse problems, achieving results comparable to gradient-based methods while significantly outperforming gradient-free alternatives.
Diffusion models have recently shown promise in time series forecasting, particularly for probabilistic predictions. However, they often fail to achieve state-of-the-art point estimation performance compared to regression-based methods. This limitation stems from difficulties in providing sufficient contextual bias to track distribution shifts and in balancing output diversity with the stability and precision required for point forecasts. Existing diffusion-based approaches mainly focus on full-distribution modeling under probabilistic frameworks, often with likelihood maximization objectives, while paying little attention to dedicated strategies for high-accuracy point estimation. Moreover, other existing point prediction diffusion methods frequently rely on pre-trained or jointly trained mature models for contextual bias, sacrificing the generative flexibility of diffusion models. To address these challenges, we propose SimDiff, a single-stage, end-to-end framework. SimDiff employs a single unified Transformer network carefully tailored to serve as both denoiser and predictor, eliminating the need for external pre-trained or jointly trained regressors. It achieves state-of-the-art point estimation performance by leveraging intrinsic output diversity and improving mean squared error accuracy through multiple inference ensembling. Key innovations, including normalization independence and the median-of-means estimator, further enhance adaptability and stability. Extensive experiments demonstrate that SimDiff significantly outperforms existing methods in time series point forecasting.
Diffusion models (DMs) memorize training images and can reproduce near-duplicates during generation. Current detection methods identify verbatim memorization but fail to capture two critical aspects: quantifying partial memorization occurring in small image regions, and memorization patterns beyond specific prompt-image pairs. To address these limitations, we propose Foreground Background Memorization (FB-Mem), a novel segmentation-based metric that classifies and quantifies memorized regions within generated images. Our method reveals that memorization is more pervasive than previously understood: (1) individual generations from single prompts may be linked to clusters of similar training images, revealing complex memorization patterns that extend beyond one-to-one correspondences; and (2) existing model-level mitigation methods, such as neuron deactivation and pruning, fail to eliminate local memorization, which persists particularly in foreground regions. Our work establishes an effective framework for measuring memorization in diffusion models, demonstrates the inadequacy of current mitigation approaches, and proposes a stronger mitigation method using a clustering approach.
This paper presents a novel generative framework for learning shared latent representations across multimodal data. Many advanced multimodal methods focus on capturing all combinations of modality-specific details across inputs, which can inadvertently obscure the high-level semantic concepts that are shared across modalities. Notably, Multimodal VAEs with low-dimensional latent variables are designed to capture shared representations, enabling various tasks such as joint multimodal synthesis and cross-modal inference. However, multimodal VAEs often struggle to design expressive joint variational posteriors and suffer from low-quality synthesis. In this work, ShaLa addresses these challenges by integrating a novel architectural inference model and a second-stage expressive diffusion prior, which not only facilitates effective inference of shared latent representation but also significantly improves the quality of downstream multimodal synthesis. We validate ShaLa extensively across multiple benchmarks, demonstrating superior coherence and synthesis quality compared to state-of-the-art multimodal VAEs. Furthermore, ShaLa scales to many more modalities while prior multimodal VAEs have fallen short in capturing the increasing complexity of the shared latent space.
Over-smoothing in Graph Neural Networks (GNNs) causes collapse in distinct node features, particularly on heterophilic graphs where adjacent nodes often have dissimilar labels. Although sheaf neural networks partially mitigate this problem, they typically rely on static or heavily parameterized sheaf structures that hinder generalization and scalability. Existing sheaf-based models either predefine restriction maps or introduce excessive complexity, yet fail to provide rigorous stability guarantees. In this paper, we introduce a novel scheme called SGPC (Sheaf GNNs with PAC-Bayes Calibration), a unified architecture that combines cellular-sheaf message passing with several mechanisms, including optimal transport-based lifting, variance-reduced diffusion, and PAC-Bayes spectral regularization for robust semi-supervised node classification. We establish performance bounds theoretically and demonstrate that end-to-end training in linear computational complexity can achieve the resulting bound-aware objective. Experiments on nine homophilic and heterophilic benchmarks show that SGPC outperforms state-of-the-art spectral and sheaf-based GNNs while providing certified confidence intervals on unseen nodes.
Diffusion models (DMs) have demonstrated to be powerful priors for signal recovery, but their application to 1-bit quantization tasks, such as 1-bit compressed sensing and logistic regression, remains a challenge. This difficulty stems from the inherent non-linear link function in these tasks, which is either non-differentiable or lacks an explicit characterization. To tackle this issue, we introduce Diff-OneBit, which is a fast and effective DM-based approach for signal recovery under 1-bit quantization. Diff-OneBit addresses the challenge posed by non-differentiable or implicit links functions via leveraging a differentiable surrogate likelihood function to model 1-bit quantization, thereby enabling gradient based iterations. This function is integrated into a flexible plug-and-play framework that decouples the data-fidelity term from the diffusion prior, allowing any pretrained DM to act as a denoiser within the iterative reconstruction process. Extensive experiments on the FFHQ, CelebA and ImageNet datasets demonstrate that Diff-OneBit gives high-fidelity reconstructed images, outperforming state-of-the-art methods in both reconstruction quality and computational efficiency across 1-bit compressed sensing and logistic regression tasks.
Graph Neural Networks (GNNs) have received increasing attention due to their ability to handle graph-structured data, yet their explainability remains a significant challenge. An effective solution is to provide the GNN models with counterfactual explanations, which aim to answer “How should the input instance be perturbed to change the model's prediction?". However, existing works mainly focus on generating explanations that can effectively alter model predictions, while neglecting whether the explanations remain aligned with the original data distribution, leading to the distribution shift problem. To address this problem, we propose a novel method called ICExplainer for generating explanations within the original distribution. Specifically, we introduce graph diffusion-based generative model into the counterfactual reasoning, treating it as an optimization objective for graph distribution learning. Taking insights from variational inference, we use it to estimate the true distribution of the input graphs to retain essential structural and semantic information. The inferred distribution is then utilized as prior knowledge to guide the reverse process, ensuring that generated explanations are both counterfactual and distributionally coherent. Extensive experiments conducted on both synthetic and real-world datasets demonstrate the superior performance of ICExplainer over existing methods.
Long-horizon planning is crucial in complex environments, but diffusion-based planners like Diffuser are limited by the trajectory lengths observed during training. This creates a dilemma: long trajectories are needed for effective planning, yet they degrade model performance. In this paper, we introduce this extendable long-horizon planning challenge and propose a two-phase solution. First, Progressive Trajectory Extension incrementally constructs longer trajectories through multi-round compositional stitching. Second, the Hierarchical Multiscale Diffuser enables efficient training and inference over long horizons by reasoning across temporal scales. To avoid the need for multiple separate models, we propose Adaptive Plan Pondering and the Recursive HM-Diffuser, which unify hierarchical planning within a single model. Experiments show our approach yields strong performance gains, advancing scalable and efficient decision-making over long-horizons.
Diffusion-based multimodal large language models (Diffusion MLLMs) have recently demonstrated impressive non-autoregressive generative capabilities across vision-and-language tasks. However, Diffusion MLLMs exhibit substantially slower inference than autoregressive models: Each denoising step employs full bidirectional self-attention over the entire sequence, resulting in cubic decoding complexity that becomes computationally impractical with thousands of visual tokens. To address this challenge, we propose D³ToM, a Decider-guided dynamic token merging method that dynamically merges redundant visual tokens at different denoising steps to accelerate inference in Diffusion MLLMs. At each denoising step, D³ToM uses decider tokens—the tokens generated in the previous denoising step—to build an importance map over all visual tokens. Then it maintains a proportion of the most salient tokens and merges the remainder through similarity-based aggregation. This plug-and-play module integrates into a single transformer layer, physically shortening the visual token sequence for all subsequent layers without altering model parameters. Moreover, D³ToM employs a merge ratio that dynamically varies with each denoising step, aligns with the native decoding process of Diffusion MLLMs, achieving superior performance under equivalent computational budgets. Extensive experiments show that D³ToM accelerates inference while preserving competitive performance.