Perceptual Drifting for One-Step Image Generation

One-step image generation removes iterative sampling at inference, but learning a single forward mapping that preserves both global semantics and local detail remains difficult, particularly at high resolution. We introduce Perceptual Drifting, a teacher-free training method that defines sample-wise drift directions as gradients of a maximum mean discrepancy in a frozen perceptual feature space. Vector--Jacobian products pull these directions back to pixels, where they form stop-gradient regression targets; an auxiliary LPIPS term promotes local consistency. Under a fixed one-network-evaluation protocol, the method improves matched pixel-drifting baselines on class-conditional ImageNet from 256 to 1024 resolution and on BLIP3o-60k text-to-image generation. Ablations support the use of DINOv3 features, joint global and patch representations, drift normalization, and local perceptual regularization. Perceptual Drifting therefore shifts additional computation to training while retaining a single generator forward pass at inference.