Generative Modeling via Drifting

Drifting Models propose a new generative modeling paradigm that shifts iterative distribution matching to training time, enabling high-quality sample generation in a single forward pass. This addresses the efficiency bottleneck of diffusion and flow-based models, which require numerous iterative steps during inference. By introducing a drifting field that governs sample movement and achieves equilibrium when distributions match, the method achieves state-of-the-art results on ImageNet at 256x256 resolution, with an FID of 1.54 in latent space and 1.61 in pixel space, outperforming previous single-step approaches and showing effectiveness in robotic control tasks.