Diffusion-based generation normally spreads sampling across many small Gaussian denoising steps, but that assumption breaks down when the process is compressed to a few coarse transitions. Existing few-step methods rely on distillation, consistency training, or adversarial objectives, and in doing so give up the likelihood framework. A team from Apple, with collaborators at UPenn and UIUC, introduces Normalizing Trajectory Models (NTM), which instead models each reverse step as an expressive conditional normalizing flow trained with exact likelihood.

Architecturally, NTM combines shallow invertible blocks within each step with a deep parallel predictor running across the trajectory. The result is an end-to-end network that can be trained from scratch or initialized from pretrained flow-matching models. Because the model has an exact trajectory likelihood, it also supports self-distillation: a lightweight denoiser is trained on the score function induced by the model itself, and this produces high-quality samples in four steps.

On text-to-image benchmarks, NTM matches or outperforms strong image generation baselines in just four sampling steps while retaining an exact likelihood over the generative trajectory — something the few-step distillation, consistency, and adversarial approaches do not offer. The abstract confines the evaluation to those benchmarks and does not include quantitative tables or ablations in this summary.

The page also links to related work that complements the NTM direction: one note argues that in few-step distillation the multi-step training trajectory, not the student’s capacity, is the bottleneck; another, iterative TARFlow, pushes normalizing flows further on image modeling. Both connect to NTM’s emphasis on rethinking how few-step generative models are trained.