Hasty Briefsbeta

The Principles of Diffusion Models

13 days ago
  • #deep-learning
  • #diffusion-models
  • #machine-learning
  • Diffusion models are based on gradually corrupting data into noise and then learning to reverse this process.
  • Three main views of diffusion models: variational (noise removal), score-based (gradient learning), and flow-based (velocity field).
  • The common backbone is a time-dependent velocity field that transforms noise into data via a continuous trajectory.
  • Applications include controllable generation, efficient numerical solvers, and flow-map models for direct mappings.
  • The monograph provides a conceptual and mathematically grounded understanding for those with basic deep-learning knowledge.