Efficient Projective Drifting Model via Sliced Wasserstein Metrics

Drifting models offer a promising route toward one-step generation by evolving the generated distribution during training rather than solving iterative denoising or flow dynamics at inference time. Recent Sinkhorn-based methods have improved the theoretical foundation of this paradigm by entropy-regularized optimal transport, but their dependence on repeated Sinkhorn iterations introduces additional training costs and is hampered by a trade-off between generation quality and efficiency. We propose ProjDrifting, an efficient projection-based drifting model via sliced Wasserstein metric. Instead of computing high-dimensional Sinkhorn iterations, ProjDrifting reduces distribution matching to sorting-based one-dimensional optimal transport, enabling parallel-friendly drift estimation. We further develop a gradient estimator based approach to directly estimate the drifting field from the gradient of the sliced Wasserstein energy. Further, we introduce an dynamic slicing mechanism that updates projection directions according to their transport energy,thereby focusing on more discriminative directions and reducing signal dilution of uniform averaging. Theoretical analysis and experiments establish that ProjDrifting improves the quality and efficiency of one-step generative modeling.