Normalizing Trajectory Models for Few-Step Image Diffusion

Normalizing Trajectory Models replace diffusion denoising with conditional normalizing flows, achieving four-step image generation with exact likelihood training and self-distillation. This hybrid approach enables faster inference than standard DDPM while maintaining generation quality. Practitioners exploring few-step diffusion alternatives to consistency models or flow matching should evaluate this method.