The paper introduces Conditional Föllmer Flow, a deep generative method for learning conditional distributions. The approach defines an ordinary differential equation (ODE) that starts from a standard Gaussian distribution and, as integration time approaches 1, approximates the desired conditional distribution. For implementation, the authors discretize the flow using Euler's method and estimate the velocity field nonparametrically with a deep neural network.
The main theoretical contribution is a convergence result: the Wasserstein-2 distance between the distribution of learned samples and the target conditional distribution is bounded, giving the first comprehensive end-to-end error analysis for conditional distribution learning via ODE flows. This distinguishes the work from earlier methods that lacked such full pipeline guarantees.
Numerical experiments cover standard nonparametric conditional density estimation and more complex image-generation tasks. The authors report that the proposed method is superior to various existing conditional density estimation techniques across these scenarios. Since there is only one source, no independent comparison is available; these claims are as stated in the paper.