Generative Modeling via Drifting (arxiv.org)

Researchers have introduced a novel approach to generative modeling called Drifting Models, which enhances the efficiency of generating data by evolving the distribution during training. This method focuses on achieving a matching pushforward distribution between the generated and actual data through a unique drifting field that directs sample movement. The key innovation lies in its capability for one-step inference, streamlining the generation process compared to traditional iterative methods used in diffusion and flow-based models. The significance of this research for the AI/ML community is underscored by its state-of-the-art performance on ImageNet at 256 x 256 resolution, achieving a Fréchet Inception Distance (FID) of 1.54 in latent space and 1.61 in pixel space. This advancement not only establishes new benchmarks for generative models but also suggests broader applications for high-quality, efficient one-step generation in various domains. The proposed objective fosters more dynamic neural network optimization, potentially inspiring further innovations in generative modeling techniques.