NeurIPS 2024 论文列表

Text-to-image diffusion models have made significant progress in generating naturalistic images from textual inputs, and demonstrate the capacity to learn and represent complex visual-semantic relationships. While these diffusion models have achieved remarkable success, the underlying mechanisms driving their performance are not yet fully accounted for, with many unanswered questions surrounding what they learn, how they represent visual-semantic relationships, and why they sometimes fail to generalize. Our work presents Diffusion Partial Information Decomposition (DiffusionPID), a novel technique that applies information-theoretic principles to decompose the input text prompt into its elementary components, enabling a detailed examination of how individual tokens and their interactions shape the generated image. We introduce a formal approach to analyze the uniqueness, redundancy, and synergy terms by applying PID to the denoising model at both the image and pixel level. This approach enables us to characterize how individual tokens and their interactions affect the model output. We first present a fine-grained analysis of characteristics utilized by the model to uniquely localize specific concepts, we then apply our approach in bias analysis and show it can recover gender and ethnicity biases. Finally, we use our method to visually characterize word ambiguity and similarity from the model’s perspective and illustrate the efficacy of our method for prompt intervention. Our results show that PID is a potent tool for evaluating and diagnosing text-to-image diffusion models. Link to project page: https://rbz-99.github.io/Diffusion-PID/.

Spectrograms are 2D representations of sound that look very different from the images found in our visual world. And natural images, when played as spectrograms, make unnatural sounds. In this paper, we show that it is possible to synthesize spectrograms that simultaneously look like natural images and sound like natural audio. We call these visual spectrograms *images that sound*. Our approach is simple and zero-shot, and it leverages pre-trained text-to-image and text-to-spectrogram diffusion models that operate in a shared latent space. During the reverse process, we denoise noisy latents with both the audio and image diffusion models in parallel, resulting in a sample that is likely under both models. Through quantitative evaluations and perceptual studies, we find that our method successfully generates spectrograms that align with a desired audio prompt while also taking the visual appearance of a desired image prompt.

In this work, we tackle the question of how to systematically benchmark task-agnostic decoupling methods for privacy-preserving machine learning (ML). Sharing datasets that include sensitive information often triggers privacy concerns, necessitating robust decoupling methods to separate sensitive and non-sensitive attributes. Despite the development of numerous decoupling techniques, a standard benchmark for systematically comparing these methods remains absent. Our framework integrates various decoupling techniques along with synthetic data generation and evaluation protocols within a unified system. Using our framework, we benchmark various decoupling techniques and evaluate their privacy-utility trade-offs. Finally, we release our source code, pre-trained models, datasets of decoupled representations to foster research in this area.

Simulation with realistic and interactive agents represents a key task for autonomous vehicle (AV) software development in order to test AV performance in prescribed, often long-tail scenarios. In this work, we propose SceneDiffuser, a scene-level diffusion prior for traffic simulation. We present a singular framework that unifies two key stages of simulation: scene initialization and scene rollout. Scene initialization refers to generating the initial layout for the traffic in a scene, and scene rollout refers to closed-loop simulation for the behaviors of the agents. While diffusion has been demonstrated to be effective in learning realistic, multimodal agent distributions, two open challenges remain: controllability and closed-loop inference efficiency and realism. To this end, to address controllability challenges, we propose generalized hard constraints, a generalized inference-time constraint mechanism that is simple yet effective. To improve closed-loop inference quality and efficiency, we propose amortized diffusion, a novel diffusion denoising paradigm that amortizes the physical cost of denoising over future simulation rollout steps, reducing the cost of per physical rollout step to a single denoising function evaluation, while dramatically reducing closed-loop errors. We demonstrate the effectiveness of our approach on the Waymo Open Dataset, where we are able to generate distributionally realistic scenes, while obtaining competitive performance in the Sim Agents Challenge, surpassing the state-of-the-art in many realism attributes.

Video generation primarily aims to model authentic and customized motion across frames, making understanding and controlling the motion a crucial topic. Most diffusion-based studies on video motion focus on motion customization with training-based paradigms, which, however, demands substantial training resources and necessitates retraining for diverse models. Crucially, these approaches do not explore how video diffusion models encode cross-frame motion information in their features, lacking interpretability and transparency in their effectiveness. To answer this question, this paper introduces a novel perspective to understand, localize, and manipulate motion-aware features in video diffusion models. Through analysis using Principal Component Analysis (PCA), our work discloses that robust motion-aware feature already exists in video diffusion models. We present a new MOtion FeaTure (MOFT) by eliminating content correlation information and filtering motion channels. MOFT provides a distinct set of benefits, including the ability to encode comprehensive motion information with clear interpretability, extraction without the need for training, and generalizability across diverse architectures. Leveraging MOFT, we propose a novel training-free video motion control framework. Our method demonstrates competitive performance in generating natural and faithful motion, providing architecture-agnostic insights and applicability in a variety of downstream tasks.

Diffusion Transformers have recently demonstrated unprecedented generative capabilities for various tasks. The encouraging results, however, come with the cost of slow inference, since each denoising step requires inference on a transformer model with a large scale of parameters. In this study, we make an interesting and somehow surprising observation: the computation of a large proportion of layers in the diffusion transformer, through introducing a caching mechanism, can be readily removed even without updating the model parameters. In the case of U-ViT-H/2, for example, we may remove up to 93.68% of the computation in the cache steps (46.84% for all steps), with less than 0.01 drop in FID. To achieve this, we introduce a novel scheme, named Learning-to-Cache (L2C), that learns to conduct caching in a dynamic manner for diffusion transformers. Specifically, by leveraging the identical structure of layers in transformers and the sequential nature of diffusion, we explore redundant computations between timesteps by treating each layer as the fundamental unit for caching. To address the challenge of the exponential search space in deep models for identifying layers to cache and remove, we propose a novel differentiable optimization objective. An input-invariant yet timestep-variant router is then optimized, which can finally produce a static computation graph. Experimental results show that L2C largely outperforms samplers such as DDIM and DPM-Solver, alongside prior cache-based methods at the same inference speed.

Recent controllable generation approaches such as FreeControl and Diffusion Self-Guidance bring fine-grained spatial and appearance control to text-to-image (T2I) diffusion models without training auxiliary modules. However, these methods optimize the latent embedding for each type of score function with longer diffusion steps, making the generation process time-consuming and limiting their flexibility and use. This work presents *Ctrl-X*, a simple framework for T2I diffusion controlling structure and appearance without additional training or guidance. Ctrl-X designs feed-forward structure control to enable the structure alignment with a structure image and semantic-aware appearance transfer to facilitate the appearance transfer from a user-input image. Extensive qualitative and quantitative experiments illustrate the superior performance of Ctrl-X on various condition inputs and model checkpoints. In particular, Ctrl-X supports novel structure and appearance control with arbitrary condition images of any modality, exhibits superior image quality and appearance transfer compared to existing works, and provides instant plug-and-play functionality to any T2I and text-to-video (T2V) diffusion model. See our project page for the code and an overview of the results: https://genforce.github.io/ctrl-x

Train-time data poisoning attacks threaten machine learning models by introducing adversarial examples during training, leading to misclassification. Current defense methods often reduce generalization performance, are attack-specific, and impose significant training overhead. To address this, we introduce a set of universal data purification methods using a stochastic transform, $\Psi(x)$, realized via iterative Langevin dynamics of Energy-Based Models (EBMs), Denoising Diffusion Probabilistic Models (DDPMs), or both. These approaches purify poisoned data with minimal impact on classifier generalization. Our specially trained EBMs and DDPMs provide state-of-the-art defense against various attacks (including Narcissus, Bullseye Polytope, Gradient Matching) on CIFAR-10, Tiny-ImageNet, and CINIC-10, without needing attack or classifier-specific information. We discuss performance trade-offs and show that our methods remain highly effective even with poisoned or distributionally shifted generative model training data.

Recent research in tabular data synthesis has focused on single tables, whereas real-world applications often involve complex data with tens or hundreds of interconnected tables. Previous approaches to synthesizing multi-relational (multi-table) data fall short in two key aspects: scalability for larger datasets and capturing long-range dependencies, such as correlations between attributes spread across different tables. Inspired by the success of diffusion models in tabular data modeling, we introduce \textbf{C}luster \textbf{La}tent \textbf{Va}riable guided \textbf{D}enoising \textbf{D}iffusion \textbf{P}robabilistic \textbf{M}odels (ClavaDDPM). This novel approach leverages clustering labels as intermediaries to model relationships between tables, specifically focusing on foreign key constraints. ClavaDDPM leverages the robust generation capabilities of diffusion models while incorporating efficient algorithms to propagate the learned latent variables across tables. This enables ClavaDDPM to capture long-range dependencies effectively. Extensive evaluations on multi-table datasets of varying sizes show that ClavaDDPM significantly outperforms existing methods for these long-range dependencies while remaining competitive on utility metrics for single-table data.

Modern datasets in neuroscience enable unprecedented inquiries into the relationship between complex behaviors and the activity of many simultaneously recorded neurons. While latent variable models can successfully extract low-dimensional embeddings from such recordings, using them to generate realistic spiking data, especially in a behavior-dependent manner, still poses a challenge. Here, we present Latent Diffusion for Neural Spiking data (LDNS), a diffusion-based generative model with a low-dimensional latent space: LDNS employs an autoencoder with structured state-space (S4) layers to project discrete high-dimensional spiking data into continuous time-aligned latents. On these inferred latents, we train expressive (conditional) diffusion models, enabling us to sample neural activity with realistic single-neuron and population spiking statistics. We validate LDNS on synthetic data, accurately recovering latent structure, firing rates, and spiking statistics. Next, we demonstrate its flexibility by generating variable-length data that mimics human cortical activity during attempted speech. We show how to equip LDNS with an expressive observation model that accounts for single-neuron dynamics not mediated by the latent state, further increasing the realism of generated samples. Finally, conditional LDNS trained on motor cortical activity during diverse reaching behaviors can generate realistic spiking data given reach direction or unseen reach trajectories. In summary, LDNS simultaneously enables inference of low-dimensional latents and realistic conditional generation of neural spiking datasets, opening up further possibilities for simulating experimentally testable hypotheses.

Deep learning-based image stitching pipelines are typically divided into three cascading stages: registration, fusion, and rectangling. Each stage requires its own network training and is tightly coupled to the others, leading to error propagation and posing significant challenges to parameter tuning and system stability. This paper proposes the Simple and Robust Stitcher (SRStitcher), which revolutionizes the image stitching pipeline by simplifying the fusion and rectangling stages into a unified inpainting model, requiring no model training or fine-tuning. We reformulate the problem definitions of the fusion and rectangling stages and demonstrate that they can be effectively integrated into an inpainting task. Furthermore, we design the weighted masks to guide the reverse process in a pre-trained large-scale diffusion model, implementing this integrated inpainting task in a single inference. Through extensive experimentation, we verify the interpretability and generalization capabilities of this unified model, demonstrating that SRStitcher outperforms state-of-the-art methods in both performance and stability.

The recent progress in text-to-image models pretrained on large-scale datasets has enabled us to generate various images as long as we provide a text prompt describing what we want. Nevertheless, the availability of these models is still limited when we expect to generate images that fall into a specific domain either hard to describe or just unseen to the models. In this work, we propose DomainGallery, a few-shot domain-driven image generation method which aims at finetuning pretrained Stable Diffusion on few-shot target datasets in an attribute-centric manner. Specifically, DomainGallery features prior attribute erasure, attribute disentanglement, regularization and enhancement. These techniques are tailored to few-shot domain-driven generation in order to solve key issues that previous works have failed to settle. Extensive experiments are given to validate the superior performance of DomainGallery on a variety of domain-driven generation scenarios.

Previous visual object tracking methods employ image-feature regression models or coordinate autoregression models for bounding box prediction. Image-feature regression methods heavily depend on matching results and do not utilize positional prior, while the autoregressive approach can only be trained using bounding boxes available in the training set, potentially resulting in suboptimal performance during testing with unseen data. Inspired by the diffusion model, denoising learning enhances the model’s robustness to unseen data. Therefore, We introduce noise to bounding boxes, generating noisy boxes for training, thus enhancing model robustness on testing data. We propose a new paradigm to formulate the visual object tracking problem as a denoising learning process. However, tracking algorithms are usually asked to run in real-time, directly applying the diffusion model to object tracking would severely impair tracking speed. Therefore, we decompose the denoising learning process into every denoising block within a model, not by running the model multiple times, and thus we summarize the proposed paradigm as an in-model latent denoising learning process. Specifically, we propose a denoising Vision Transformer (ViT), which is composed of multiple denoising blocks. In the denoising block, template and search embeddings are projected into every denoising block as conditions. A denoising block is responsible for removing the noise in a predicted bounding box, and multiple stacked denoising blocks cooperate to accomplish the whole denoising process. Subsequently, we utilize image features and trajectory information to refine the denoised bounding box. Besides, we also utilize trajectory memory and visual memory to improve tracking stability. Experimental results validate the effectiveness of our approach, achieving competitive performance on several challenging datasets. The proposed in-model latent denoising tracker achieve real-time speed, rendering denoising learning applicable in the visual object tracking community.

Diffusion-based purification has demonstrated impressive robustness as an adversarial defense. However, concerns exist about whether this robustness arises from insufficient evaluation. Our research shows that EOT-based attacks face gradient dilemmas due to global gradient averaging, resulting in ineffective evaluations. Additionally, 1-evaluation underestimates resubmit risks in stochastic defenses. To address these issues, we propose an effective and efficient attack named DiffHammer. This method bypasses the gradient dilemma through selective attacks on vulnerable purifications, incorporating $N$-evaluation into loops and using gradient grafting for comprehensive and efficient evaluations. Our experiments validate that DiffHammer achieves effective results within 10-30 iterations, outperforming other methods. This calls into question the reliability of diffusion-based purification after mitigating the gradient dilemma and scrutinizing its resubmit risk.

Conventional diffusion models typically relies on a fixed forward process, which implicitly defines complex marginal distributions over latent variables. This can often complicate the reverse process’ task in learning generative trajectories, and results in costly inference for diffusion models. To address these limitations, we introduce Neural Flow Diffusion Models (NFDM), a novel framework that enhances diffusion models by supporting a broader range of forward processes beyond the standard Gaussian. We also propose a novel parameterization technique for learning the forward process. Our framework provides an end-to-end, simulation-free optimization objective, effectively minimizing a variational upper bound on the negative log-likelihood. Experimental results demonstrate NFDM’s strong performance, evidenced by state-of-the-art likelihood estimation. Furthermore, we investigate NFDM’s capacity for learning generative dynamics with specific characteristics, such as deterministic straight lines trajectories, and demonstrate how the framework may be adopted for learning bridges between two distributions. The results underscores NFDM’s versatility and its potential for a wide range of applications.

Text-guided diffusion models have significantly advanced image editing, enabling high-quality and diverse modifications driven by text prompts. However, effective editing requires inverting the source image into a latent space, a process often hindered by prediction errors inherent in DDIM inversion. These errors accumulate during the diffusion process, resulting in inferior content preservation and edit fidelity, especially with conditional inputs. We address these challenges by investigating the primary contributors to error accumulation in DDIM inversion and identify the singularity problem in traditional noise schedules as a key issue. To resolve this, we introduce the *Logistic Schedule*, a novel noise schedule designed to eliminate singularities, improve inversion stability, and provide a better noise space for image editing. This schedule reduces noise prediction errors, enabling more faithful editing that preserves the original content of the source image. Our approach requires no additional retraining and is compatible with various existing editing methods. Experiments across eight editing tasks demonstrate the Logistic Schedule's superior performance in content preservation and edit fidelity compared to traditional noise schedules, highlighting its adaptability and effectiveness. The project page is available at https://lonelvino.github.io/SYE/.

Motion-to-music and music-to-motion have been studied separately, each attracting substantial research interest within their respective domains. The interaction between human motion and music is a reflection of advanced human intelligence, and establishing a unified relationship between them is particularly important. However, to date, there has been no work that considers them jointly to explore the modality alignment within. To bridge this gap, we propose a novel framework, termed MoMu-Diffusion, for long-term and synchronous motion-music generation. Firstly, to mitigate the huge computational costs raised by long sequences, we propose a novel Bidirectional Contrastive Rhythmic Variational Auto-Encoder (BiCoR-VAE) that extracts the modality-aligned latent representations for both motion and music inputs. Subsequently, leveraging the aligned latent spaces, we introduce a multi-modal diffusion Transformer model and a cross-guidance sampling strategy to enable various generation tasks, including cross-modal, multi-modal, and variable-length generation. Extensive experiments demonstrate that MoMu-Diffusion surpasses recent state-of-the-art methods both qualitatively and quantitatively, and can synthesize realistic, diverse, long-term, and beat-matched music or motion sequences. The generated motion-music samples are available at https://momu-diffusion.github.io/.

Langevin Dynamics is a Stochastic Differential Equation (SDE) central to sampling and generative modeling and is implemented via time discretization. Langevin Monte Carlo (LMC), based on the Euler-Maruyama discretization, is the simplest and most studied algorithm. LMC can suffer from slow convergence - requiring a large number of steps of small step-size to obtain good quality samples. This becomes stark in the case of diffusion models where a large number of steps gives the best samples, but the quality degrades rapidly with smaller number of steps. Randomized Midpoint Method has been recently proposed as a better discretization of Langevin dynamics for sampling from strongly log-concave distributions. However, important applications such as diffusion models involve non-log concave densities and contain time varying drift. We propose its variant, the Poisson Midpoint Method, which approximates a small step-size LMC with large step-sizes. We prove that this can obtain a quadratic speed up of LMC under very weak assumptions. We apply our method to diffusion models for image generation and show that it maintains the quality of DDPM with 1000 neural network calls with just 50-80 neural network calls and outperforms ODE based methods with similar compute.

Graph is a prevalent discrete data structure, whose generation has wide applications such as drug discovery and circuit design. Diffusion generative models, as an emerging research focus, have been applied to graph generation tasks. Overall, according to the space of states and time steps, diffusion generative models can be categorized into discrete-/continuous-state discrete-/continuous-time fashions. In this paper, we formulate the graph diffusion generation in a discrete-state continuous-time setting, which has never been studied in previous graph diffusion models. The rationale of such a formulation is to preserve the discrete nature of graph-structured data and meanwhile provide flexible sampling trade-offs between sample quality and efficiency. Analysis shows that our training objective is closely related to the generation quality and our proposed generation framework enjoys ideal invariant/equivariant properties concerning the permutation of node ordering. Our proposed model shows competitive empirical performance against other state-of-the-art graph generation solutions on various benchmarks while at the same time can flexibly trade off the generation quality and efficiency in the sampling phase.