MGFlow is a new method for distributional training in one-step visual generation. It models feature distributions with Gaussian mixtures, separating distribution modeling from matching discrepancy, and uses Wasserstein gradient flow to connect global objectives to pointwise updates. MGFlow supports both optimal transport and score-based matching, and it couples mass-constrained sample assignment with paired component updates to address mode collapse.
On ImageNet 256x256, MGFlow surpasses the FD-Loss baseline, achieving 1.45 FDr^6 on pMF-H and 1.64 on JiT-H. For text-to-image, MGFlow post-trains FLUX.2 4B into a one-step generator that outperforms the original four-step model on both GenEval and PickScore.
This is interesting because it provides a unified framework that recovers prior methods like FD-Loss and Gaussian-kernel Drifting as special cases, and demonstrates that mixture models with careful assignment can significantly improve one-step generation quality.