Normalizing Trajectory Models
Apple researchers have introduced a technique called Normalizing Trajectory Models to improve image generation quality when using only a few sampling steps. By learning to map complex, non-linear diffusion paths into straight trajectories, the method allows models to produce high-fidelity outputs without the extensive processing typically required by traditional denoising. This approach helps bridge the gap between efficiency and image accuracy, potentially enabling faster real-time generation on hardware with limited computational resources.