Bernini MNIST
Generative flow and velocity field modeling experiments on benchmark vision datasets
Experimental study and implementation of Bernini generative pipelines and continuous velocity flow matching on standard benchmark datasets.
Generative Modeling Experiment
Highlights
- Velocity Trajectory Parameterization: Analyzing optimal transport velocity fields and trajectory curvature on MNIST digit manifolds.
- Sampling Efficiency: Evaluating numerical ODE solvers (Euler, Midpoint, RK4) for continuous-time generative sampling with minimum function evaluations.