Normalizing Trajectory Models
Apple Machine Learning Research
Normalizing Trajectory Models (NTM) reframe few-step reverse diffusion as conditional normalizing flows, enabling exact likelihood training for the full generative trajectory. The method produces high-quality text-to-image samples in four sampling steps. It replaces earlier few-step approaches that rely on distillation or other objectives by keeping the likelihood framework and supports self-distillation from the model’s own score function.
Why it matters
Diffusion-based models decompose sampling into many small Gaussian denoising steps, an assumption that breaks down when generation is compressed to a few coarse transitions. Existing few-step methods address this through distillation, consistency training, or adversarial objectives, but sacrifice the likelihood framework in the process. We introduce Normalizing Trajectory Models (NTM), which models each reverse step as an expressive conditional normalizing flow with exact likelihood training. Architecturally, NTM combines shallow invertible blocks within each step with a deep parallel…
