Scientific Methodology of Flow Matching Generalization

Observation: Modern generative models, like diffusion and flow matching, generalize remarkably well, but the reasons are unclear.
Question/Problem: Is the stochastic (noisy) nature of the flow matching loss a primary reason for this strong generalization?
Hypothesis: The stochastic nature of the loss is not a primary contributor to generalization, and a non-stochastic (closed-form) loss will perform comparably.
Method (to test the Hyp.): Empirically compare the training losses and final statistical performance of models trained with the stochastic loss versus the closed-form loss.
Experiment: Train state-of-the-art flow matching models on standard image datasets (e.g., CIFAR-10) using both loss variants.
Analysis: Both loss variants yielded nearly equivalent loss values and achieved comparable (or even better) statistical performance.
Conclusion: We ruled out the stochastic nature of the loss as a primary contributor to the effective generalization of flow matching models.