The Center for Education and Research in Information Assurance and Security (CERIAS)

The Center for Education and Research in
Information Assurance and Security (CERIAS)

Generative AI: Diffusion Models and Flow Matching

Principal Investigator: Vaneet Aggarwal

Diffusion models and flow matching are the techniques behind most modern image generators. They work by learning to gradually turn random noise into realistic data, and they are increasingly used in science and engineering. Yet a basic question remains open: how much training data and computation do they need to reach a given accuracy? This project answers that question and builds practical tools on top of these models.

Our outcomes include:

  • How much data generative models need: the first optimal data requirements ("sample complexity") for learning a distribution with flow matching, and similar bounds for diffusion models, including models of discrete data, models conditioned on extra inputs, and models trained as data arrives.
  • Handling change: flow-matching methods that stay accurate when the data distribution drifts over time.
  • Privacy: PrivateEdit, a pipeline for editing images with generative AI without exposing private content.
  • Robotics: diffusion-based planning that accounts for uncertainty, so a robot can avoid moving obstacles.
  • Biomedicine: generative models enriched with biomedical knowledge graphs for targeted drug discovery, and GeneFlow, which translates gene-expression measurements into tissue images.

A survey covers how generative models can be used to learn decision policies from logged data.

Representative Publications

  • Jiayu Chen, Bhargav Ganguly, Yang Xu, Yongsheng Mei, Tian Lan, and Vaneet Aggarwal, "Deep Generative Models for Offline Policy Learning: Tutorial, Survey, and Perspectives on Future Directions," Transactions on Machine Learning Research, Aug 2024.

  • Aditya Malusare and Vaneet Aggarwal, "Improving Molecule Generation and Drug Discovery with a Knowledge-enhanced Generative Model," IEEE Transactions on Computational Biology and Bioinformatics, vol. 22, no. 1, pp. 375-381, Jan.-Feb. 2025.

  • Vineet Punyamoorty, Pascal Jutras-Dubé, Ruqi Zhang, Vaneet Aggarwal, Damon Conover, and Aniket Bera, "Dynamic Obstacle Avoidance through Uncertainty-Based Adaptive Planning with Diffusion," in Proc. IROS, Oct 2025.

  • Mengbo Wang, Shourya Verma, Aditya Malusare, Luopin Wang, Yiyang Lu, Mario Sola, Vaneet Aggarwal, Ananth Grama, and Nadia Atallah Lanman, "GeneFlow: Single-cell Gene Expression Translation to Histopathological Images via Rectified Flow," in Proc. NeurIPS, Dec 2025.

  • Mudit Gaur, Prashant Trivedi, Sasidhar Kunapuli, Amrit Singh Bedi, and Vaneet Aggarwal, "Improved Sample Complexity Bounds For Diffusion Model Training Without Empirical Risk Minimizer Access," Transactions on Machine Learning Research, Apr 2026.

  • Aadithya Srikanth, Mudit Gaur, and Vaneet Aggarwal, "Discrete State Diffusion Models: A Sample Complexity Perspective," in Proc. AISTATS, May 2026.

  • Aditya Malusare, Vineet Punyamoorty, and Vaneet Aggarwal, "Augmenting generative models with biomedical knowledge graphs improves targeted drug discovery," IEEE Transactions on Artificial Intelligence, vol. 7, no. 5, pp. 2638-2647, May 2026.

  • Dipesh Tamboli, Vineet Punyamoorty, Atharv Pawar, and Vaneet Aggarwal, "PrivateEdit: A Privacy-Preserving Pipeline for Generative Image Editing," IEEE Transactions on Artificial Intelligence, vol. 7, no. 8, pp. 4726-4735, Aug. 2026.

  • Hari Krishna Sahoo, Mudit Gaur, and Vaneet Aggarwal, "Order-Optimal Sample Complexity for Distribution Learning via Flow Matching," in Proc. NeurIPS, Dec 2026.

  • Mudit Gaur, Prashant Trivedi, Sasidhar Kunapuli, Amrit Singh Bedi, and Vaneet Aggarwal, "Finite-Time Online Sample Complexity for Conditional Diffusion Models," arXiv, Nov 2026.

  • Aadithya Srikanth and Vaneet Aggarwal, "Minimax-Optimal Predictive Flow Matching under Distribution Drift," arXiv, Oct 2026.

Keywords: diffusion models, distribution drift, flow matching, generative AI, privacy-preserving generation, sample complexity