All research directions

Diffusion & Flow Matching

Generative modeling via diffusion, score-based methods, and flow matching for high-fidelity synthesis.

To advance the theory and practice of diffusion and flow-based generative models for controllable, high-quality data synthesis.

Overview

This direction develops generative models based on diffusion processes and flow matching. Research covers score-based methods, training dynamics, and controllable synthesis for images, text, and multimodal data.

Key objectives

  • Develop stable training objectives for diffusion and flow models
  • Enable controllable and personalized generation
  • Establish theoretical guarantees on sample quality
  • Bridge generative modeling with downstream applications

Key topics

  • Diffusion and score-based generative models
  • Flow matching and continuous normalizing flows
  • Controllable and personalized synthesis
  • Generative modeling for vision and language

Papers in this direction

  • 2018

    MGAN: Training Generative Adversarial Nets with Multiple Generators

    Hoang, T, Le, T, Tran, T, Nguyen, QV, Phung, D

    International Conference on Learning Representations (ICLR)

  • 2017

    Dual Discriminator Generative Adversarial Nets

    Nguyen, TD, Le, T, Vu, H, Phung, D

    Advances in Neural Information Processing Systems (NeurIPS)