Normalizing Trajectory Models

Introduces Normalizing Trajectory Models (NTM) which model each reverse step as an expressive conditional normalizing flow with exact likelihood training. NTM combines shallow invertible blocks within each step with a deep parallel predictor across the trajectory. It enables self-distillation to a lightweight denoiser producing high-quality samples in four steps, outperforming baselines on text-to-image benchmarks while retaining exact likelihood.