Normalizing Trajectory Models
Apple Research just solved few-step diffusion. Normalizing flows + invertible blocks = faster generation with exact likelihood.
Why it matters
Apple's Normalizing Trajectory Models (NTM) address a known limitation in fast diffusion sampling: existing few-step methods (distillation, consistency training) abandon the likelihood framework. NTM preserves exact likelihood while compressing generation to coarse transitions via expressive conditional normalizing flows—a technical advance that matters for practitioners optimizing inference speed without sacrificing model interpretability or training rigor.
The key facts
5 to knowApple ML Research paper on Normalizing Trajectory Models
Combines shallow invertible blocks per step with deep parallel architecture
Solves likelihood-training tradeoff in few-step diffusion
Existing methods (distillation, consistency training, adversarial) sacrifice likelihood framework
Published October 8, 2026 on machinelearning.apple.com
Go to the source
Apple Machine Learningmachinelearning.apple.com
Publisher excerpt: 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…