Unifying Distributional Training for One-Step Visual Generation
MGFlow is the paper’s proposed bridge from distribution-level objectives to one-step image generator updates.
The authors frame FD-Loss and Gaussian-kernel Drifting as cases inside a broader distributional training theory. MGFlow uses Gaussian mixtures to model feature distributions, with optimal-transport or score-based matching and constrained sample assignment to reduce mode collapse. On ImageNet 256x256, they report 1.45 FDr6 on pMF-H and 1.64 on JiT-H, above the FD-Loss baseline. They also say MGFlow post-trains FLUX.2 4B into a one-step text-to-image generator that beats the original four-step model on GenEval and PickScore. ArXiv · AI/CL/LG's note
The authors frame FD-Loss and Gaussian-kernel Drifting as cases inside a broader distributional training theory. MGFlow uses Gaussian mixtures to model feature distributions, with optimal-transport or score-based matching and constrained sample assignment to reduce mode collapse. On ImageNet 256x256, they report 1.45 FDr6 on pMF-H and 1.64 on JiT-H, above the FD-Loss baseline. They also say MGFlow post-trains FLUX.2 4B into a one-step text-to-image generator that beats the original four-step model on GenEval and PickScore. ArXiv · AI/CL/LG's note
score 5