Lookahead Drifting Model Improves One-Step Image Generation

A new machine learning method, the lookahead drifting model, enhances one-step image generation by computing sequential drifting terms during training. This approach builds on the drifting model paradigm, which maps distributions via a single neural functional evaluation (NFE). At each iteration, the lookahead variant calculates multiple drifting terms sequentially, using previously computed terms along with positive samples and model output. Later-stage drifting terms capture higher-order gradient information toward positive samples, improving model optimization. The method achieves state-of-the-art performance on ImageNet, pushing one-step generation quality further. The paper is available on arXiv under ID 2605.04060.