Normalizing Trajectory Models

Researchers at Apple, alongside collaborators from the University of Pennsylvania and UIUC, introduced Normalizing Trajectory Models (NTM). The authors are Jiatao Gu, Tianrong Chen, Ying Shen, David Berthelot, Shuangfei Zhai and Josh Susskind. Diffusion-based models break sampling into many small Gaussian denoising steps, an assumption that fails when generation is compressed into a few coarse transitions.

Existing few-step methods handle this through distillation, consistency training or adversarial objectives, but give up the likelihood framework in the process. NTM instead treats each reverse step as an expressive conditional normalizing flow trained with exact likelihood. A normalizing flow is a model that maps data to a simple distribution through invertible transformations, which lets it compute exact likelihoods.

Architecturally, NTM pairs shallow invertible blocks inside each step with a deep parallel predictor spanning the trajectory, forming an end-to-end network. It can be trained from scratch or initialized from pretrained flow-matching models. The exact trajectory likelihood also supports self-distillation: a lightweight denoiser trained on the score function induced by the model itself produces high-quality samples in four steps. On text-to-image benchmarks, NTM matches or outperforms strong image generation baselines in four sampling steps while retaining exact likelihood over the generative trajectory.

Why it matters: Developers working on few-step image generation may gain a method that keeps exact likelihood while matching or beating strong baselines at four sampling steps, with the option to start from existing pretrained flow-matching models rather than train from scratch.