Generative diffusion models have recently emerged as a leading approach for generating high-dimensional data. In this paper, we show that the dynamics of these models exhibit a spontaneous symmetry breaking that divides the generative dynamics into two distinct phases: 1) A linear steady-state dynamics around a central fixed-point and 2) an attractor dynamics directed towards the data manifold. These two phases'' are separated by the change in stability of the central fixed-point, with the resulting window of instability being responsible for the diversity of the generated samples. Using both theoretical and empirical evidence, we show that an accurate simulation of the early dynamics does not significantly contribute to the final generation, since early fluctuations are reverted to the central fixed point. To leverage this insight, we propose a Gaussian late initialization scheme, which significantly improves model performance, achieving up to 3x FID improvements on fast samplers, while also increasing sample diversity (e.g., racial composition of generated CelebA images). Our work offers a new way to understand the generative dynamics of diffusion models that has the potential to bring about higher performance and less biased fast-samplers.

Abstract--This paper proposes a supervised training algorithm for learning stochastic resource allocation policies with generative diffusion models (GDMs). We formulate the allocation problem as the maximization of an ergodic utility function subject to ergodic Quality of Service (QoS) constraints. Given samples from a stochastic expert policy that yields a near-optimal solution to the constrained optimization problem, we train a GDM policy to imitate the expert and generate new samples from the optimal distribution. We achieve near-optimal performance through the sequential execution of the generated samples. T o enable generalization to a family of network configurations, we parameterize the backward diffusion process with a graph neural network (GNN) architecture.

Sequential Monte Carlo (SMC) methods have recently shown successful results for conditional sampling of generative diffusion models. In this paper we propose a new diffusion posterior SMC sampler achieving improved statistical efficiencies, particularly under outlier conditions or highly informative likelihoods. The key idea is to construct an observation path that correlates with the diffusion model and to design the sampler to leverage this correlation for more efficient sampling. Empirical results conclude the efficiency.

Understanding the neural basis of behavior is a fundamental goal in neuroscience. Current research in large-scale neuro-behavioral data analysis often relies on decoding models, which quantify behavioral information in neural data but lack details on behavior encoding. This raises an intriguing scientific question: "how can we enable in-depth exploration of neural representations in behavioral tasks, revealing interpretable neural dynamics associated with behaviors". However, addressing this issue is challenging due to the varied behavioral encoding across different brain regions and mixed selectivity at the population level. To tackle this limitation, our approach, named ("BeNeDiff"), first identifies a fine-grained and disentangled neural subspace using a behavior-informed latent variable model. It then employs state-of-the-art generative diffusion models to synthesize behavior videos that interpret the neural dynamics of each latent factor.

In the context of the rising share of new energy generation, accurately generating new energy output scenarios is crucial for day-ahead power system scheduling. Deep learning-based scenario generation methods can address this need, but their black-box nature raises concerns about interpretability. To tackle this issue, this paper introduces a method for day-ahead new energy scenario generation based on an improved conditional generative diffusion model. This method is built on the theoretical framework of Markov chains and variational inference. It first transforms historical data into pure noise through a diffusion process, then uses conditional information to guide the denoising process, ultimately generating scenarios that satisfy the conditional distribution. Additionally, the noise table is improved to a cosine form, enhancing the quality of the generated scenarios. When applied to actual wind and solar output data, the results demonstrate that this method effectively generates new energy output scenarios with good adaptability.

Diffusion models suffer from the huge consumption of time and resources to train. For example, diffusion models need hundreds of GPUs to train for several weeks for a high-resolution generative task to meet the requirements of an extremely large number of iterations and a large batch size. Training diffusion models become a millionaire's game. With limited resources that only fit a small batch size, training a diffusion model always fails. In this paper, we investigate the key reasons behind the difficulties of training diffusion models with limited resources. Through numerous experiments and demonstrations, we identified a major factor: the significant variation in the training losses across different timesteps, which can easily disrupt the progress made in previous iterations. Moreover, different prediction types of $x_0$ exhibit varying effectiveness depending on the task and timestep. We hypothesize that using a mixed-prediction approach to identify the most accurate $x_0$ prediction type could potentially serve as a breakthrough in addressing this issue. In this paper, we outline several challenges and insights, with the hope of inspiring further research aimed at tackling the limitations of training diffusion models with constrained resources, particularly for high-resolution tasks.

Generative diffusion models have recently emerged as a leading approach for generating high-dimensional data. In this paper, we show that the dynamics of these models exhibit a spontaneous symmetry breaking that divides the generative dynamics into two distinct phases: 1) A linear steady-state dynamics around a central fixed-point and 2) an attractor dynamics directed towards the data manifold. These two "phases'' are separated by the change in stability of the central fixed-point, with the resulting window of instability being responsible for the diversity of the generated samples. Using both theoretical and empirical evidence, we show that an accurate simulation of the early dynamics does not significantly contribute to the final generation, since early fluctuations are reverted to the central fixed point. To leverage this insight, we propose a Gaussian late initialization scheme, which significantly improves model performance, achieving up to 3x FID improvements on fast samplers, while also increasing sample diversity (e.g., racial composition of generated CelebA images).

Deterministic regression-based downscaling models for climate variables often suffer from spectral bias, which can be mitigated by generative models like diffusion models. To enable efficient and reliable simulation of extreme weather events, it is crucial to achieve rapid turnaround, dynamical consistency, and accurate spatio-temporal spectral recovery. We propose an efficient correction diffusion model, TAUDiff, that combines a deterministic spatio-temporal model for mean field downscaling with a smaller generative diffusion model for recovering the fine-scale stochastic features. We demonstrate the efficacy of this approach on downscaling atmospheric wind velocity fields obtained from coarse GCM simulations. Our approach can not only ensure quicker simulation of extreme events but also reduce overall carbon footprint due to low inference times.

We present an estimate of the Wasserstein distance between the data distribution and the generation of score-based generative models, assuming an $\epsilon$-accurate approximation of the score and a Gaussian-type tail behavior of the data distribution. The complexity bound in dimension is $O(\sqrt{d})$, with a logarithmic constant. Such Gaussian tail assumption applies to the distribution of a compact support target with early stopping technique and the Bayesian posterior with a bounded observation operator. Corresponding convergence and complexity bounds are derived. The crux of the analysis lies in the Lipchitz bound of the score, which is related to the Hessian estimate of a viscous Hamilton-Jacobi equation (vHJ). This latter is demonstrated by employing a dimension independent kernel estimate. Consequently, our complexity bound scales linearly (up to a logarithmic constant) with the square root of the trace of the covariance operator, which relates to the invariant distribution of forward process. Our analysis also extends to the probabilistic flow ODE, as the sampling process.

While high-quality texture maps are essential for realistic 3D asset rendering, few studies have explored learning directly in the texture space, especially on large-scale datasets. In this work, we depart from the conventional approach of relying on pre-trained 2D diffusion models for test-time optimization of 3D textures. Instead, we focus on the fundamental problem of learning in the UV texture space itself. For the first time, we train a large diffusion model capable of directly generating high-resolution texture maps in a feed-forward manner. To facilitate efficient learning in high-resolution UV spaces, we propose a scalable network architecture that interleaves convolutions on UV maps with attention layers on point clouds. Leveraging this architectural design, we train a 700 million parameter diffusion model that can generate UV texture maps guided by text prompts and single-view images. Once trained, our model naturally supports various extended applications, including text-guided texture inpainting, sparse-view texture completion, and text-driven texture synthesis. Project page is at http://cvmi-lab.github.io/TEXGen/.