One-Step Generative Modeling

One-step generative modeling aims to produce realistic samples with a single model evaluation instead of repeatedly denoising or integrating a trajectory. Common approaches distill multi-step diffusion or flow teachers, train consistency and MeanFlow models, or learn direct transport maps using distribution matching, drifting, or optimal transport. These methods are tested on image-generation benchmarks such as ImageNet and FID, as well as language, audio, video, scientific forecasting, and robotic control tasks. The central challenge is to retain the quality, diversity, and conditional accuracy of multi-step models while reducing inference cost; trajectory design, data-noise coupling, and evaluation beyond a single metric remain important concerns.