flow-based-diffusion

points_video.mp4

What:

A superset of DDPM, i.e. a continuous space generalization physically moving probability density fields.

How:

Solves ordinary differential equations (ODE) instead of stochastic differential equations (SDE) to adapt to the continuous space and learn a vector field at each time step from 0 to 1.

Resources:

1. Denoising Diffusion Probabilistic Models
2. Flow Matching For Generative Modeling
3. Flow Matching paper guide
4. Referenced code, heavily adapted