Correcting Flows with Marginal Matching
The authors report better sample quality on CIFAR-10 and ImageNet-64 with minimal additional computation time.
Their training-based predictor-corrector framework uses the reversibility of the ODE underlying flow matching. At a chosen time step, it corrects forward-generated samples toward the corresponding marginal distribution of the backward ODE. The authors prove that this correction reduces a bound on the distance between the learned and data distributions, motivating two implementations of the correction function.
ArXiv · AI/CL/LG's note
Their training-based predictor-corrector framework uses the reversibility of the ODE underlying flow matching. At a chosen time step, it corrects forward-generated samples toward the corresponding marginal distribution of the backward ODE. The authors prove that this correction reduces a bound on the distance between the learned and data distributions, motivating two implementations of the correction function.
ArXiv · AI/CL/LG's note
score 4