More Like this

1. Data augmentation via diffusion model to enhance AI fairness

   https://doi.org/10.3389/frai.2025.1530397  

   Hastings_Blow, Christina; Qian, Lijun; Gibson, Camille; Obiomon, Pamela; Dong, Xishuang (March 2025, Frontiers in Artificial Intelligence)

   IntroductionAI fairness seeks to improve the transparency and explainability of AI systems by ensuring that their outcomes genuinely reflect the best interests of users. Data augmentation, which involves generating synthetic data from existing datasets, has gained significant attention as a solution to data scarcity. In particular, diffusion models have become a powerful technique for generating synthetic data, especially in fields like computer vision. MethodsThis paper explores the potential of diffusion models to generate synthetic tabular data to improve AI fairness. The Tabular Denoising Diffusion Probabilistic Model (Tab-DDPM), a diffusion model adaptable to any tabular dataset and capable of handling various feature types, was utilized with different amounts of generated data for data augmentation. Additionally, reweighting samples from AIF360 was employed to further enhance AI fairness. Five traditional machine learning models—Decision Tree (DT), Gaussian Naive Bayes (GNB), K-Nearest Neighbors (KNN), Logistic Regression (LR), and Random Forest (RF)—were used to validate the proposed approach. Results and discussionExperimental results demonstrate that the synthetic data generated by Tab-DDPM improves fairness in binary classification. 

   more » « less
2. Synthetizing acoustic emission data using a diffusion-based generative artificial intelligence model

   https://doi.org/10.1117/12.3090609  

   Bazrafshan, Pedram; Ebrahimkhanlou, Arvin (April 2026, SPIE)

   Niezrecki, Christopher; Farhangdoust, Saman (Ed.)

   This study investigates the use of denoising diffusion probabilistic models (DDPMs) to generate synthetic time-frequency images of acoustic emission (AE) signals. Conventional numerical simulations of AE signals are limited by their deterministic nature, producing identical outputs that fail to capture the stochastic variability observed in real materials. In this work, a dataset of 53,965 AE signals collected from concrete, wood, and recycled plastic specimens was transformed into time-frequency images for model training. The DDPM, implemented with a U-Net backbone, was trained for 600 epochs at 512×512 resolution using mean squared error loss and a linear noise schedule. The model introduces Gaussian noise during the forward process and incrementally removes it during the reverse process to reconstruct high-fidelity images. Evaluation with the Fréchet inception distance (FID) demonstrated a score of 41.2, confirming satisfactory alignment between the generated and real data distributions. The results show that the model effectively captured both coarse and fine features of AE signals, while maintaining variability consistent with real-world loading conditions and material behavior. These findings highlight the capability of DDPMs to provide robust synthetic datasets for AE research, supporting structural health monitoring and nondestructive evaluation. 

   more » « less
3. A generative artificial intelligence benchmarking study: comparing diffusion and GAN-based models for surface crack image generation in civil infrastructure

   https://doi.org/10.1117/12.3090605  

   Bazrafshan, Pedram; Ebrahimkhanlou, Arvin (April 2026, SPIE)

   Rizzo, Piervincenzo; Ricci, Fabrizio; Schaal, Christoph; Peters, Kara J (Ed.)

   This study investigates the generation of synthetic surface crack images in civil infrastructure using two state-of-the-art generative artificial intelligence models: the denoising diffusion probabilistic model (DDPM) and StyleGAN3-ada. The objective is to assess their capability to produce realistic crack patterns for potential use in addressing data scarcity in infrastructure monitoring. Both models were trained on two datasets consisting of 9,887 images of surface cracks in concrete and asphalt pavements, including RGB and masked binary versions. The quality of generated images was evaluated using the Fréchet Inception Distance (FID), a metric for measuring the similarity between real and synthetic images. The DDPM achieved lower FID scores of 30.1 for RGB images and 22.4 for masked images, indicating closer alignment with real distributions. In contrast, StyleGAN3-ada obtained higher FID scores of 45.9 for RGB and 52.2 for masked images, reflecting inferior performance. Visual inspection also showed that DDPM-generated images exhibited more realistic textures, orientations, and crack morphologies, whereas StyleGAN3-ada frequently produced blurred or distorted patterns. These findings demonstrate that diffusion-based approaches outperform GAN-based methods in generating surface crack patterns that closely replicate real-world characteristics. The study highlights the effectiveness of DDPM in capturing complex geometrical and textural features of cracks, underscoring its suitability for generating high-fidelity synthetic data. 

   more » « less
4. TEDi: Temporally-Entangled Diffusion for Long-Term Motion Synthesis

   https://doi.org/10.1145/3641519.3657515  

   Zhang, Zihan; Liu, Richard; Hanocka, Rana; Aberman, Kfir (July 2024, ACM)

   The gradual nature of a diffusion process that synthesizes samples in small increments constitutes a key ingredient of Denoising Diffusion Probabilistic Models (DDPM), which have presented unprecedented quality in image synthesis and been recently explored in the motion domain. In this work, we propose to adapt the gradual diffusion concept (operating along a diffusion time-axis) into the temporal-axis of the motion sequence. Our key idea is to extend the DDPM framework to support temporally varying denoising, thereby entangling the two axes. Using our special formulation, we iteratively denoise a motion buffer that contains a set of increasingly-noised poses, which auto-regressively produces an arbitrarily long stream of frames. With a stationary diffusion time-axis, in each diffusion step we increment only the temporal-axis of the motion such that the framework produces a new, clean frame which is removed from the beginning of the buffer, followed by a newly drawn noise vector that is appended to it. This new mechanism paves the way towards a new framework for long-term motion synthesis with applications to character animation and other domains. 

   more » « less
5. Learning Mixtures of Gaussians Using the DDPM Objective

   Shah, Kulin; Chen, Sitan; Klivans, Adam (December 2023, NeurIPS)

   Recent works have shown that diffusion models can learn essentially any distribution provided one can perform score estimation. Yet it remains poorly understood under what settings score estimation is possible, let alone when practical gradient-based algorithms for this task can provably succeed. In this work, we give the first provably efficient results along these lines for one of the most fundamental distribution families, Gaussian mixture models. We prove that gradient descent on the denoising diffusion probabilistic model (DDPM) objective can efficiently recover the ground truth parameters of the mixture model in the following two settings: 1) We show gradient descent with random initialization learns mixtures of two spherical Gaussians in d dimensions with 1/poly(d)-separated centers. 2) We show gradient descent with a warm start learns mixtures of K spherical Gaussians with Ω(log(min(K,d)))-separated centers. A key ingredient in our proofs is a new connection between score-based methods and two other approaches to distribution learning, the EM algorithm and spectral methods 

   more » « less