My solutions to the three flow matching / diffusion labs from the Foundation Models course: simulating SDEs, building flow matching from scratch, and training a conditional diffusion transformer on MNIST with classifier-free guidance.
Euler and Euler-Maruyama solvers, Brownian motion, and Ornstein-Uhlenbeck processes.
Conditional probability paths and vector fields, trained on 2D toy distributions (moons, checkerboard).
A diffusion transformer trained on MNIST with classifier-free guidance, plus latent-space interpolation between digits.
pixi installNotebooks: lab_one.ipynb, lab_two.ipynb, lab_three.ipynb.