NTM: Exact Likelihood Four-Step Image Generation via Normalizing Flows

1 reports archived1 independent sourcesupdated 10/8/2026, 08:00:00

Synthesis & Latest Updates

1 Sources Cross-Validated

Researchers from UPenn and UIUC introduce Normalizing Trajectory Models (NTM), which use conditional normalizing flows to retain exact likelihood in few-step diffusion sampling. By combining shallow invertible blocks with a deep parallel predictor, NTM achieves state-of-the-art text-to-image results in just four steps while enabling self-distillation through its precise trajectory likelihood.

LATEST/Researchers from UPenn and UIUC introduce Normalizing Trajectory Models (NTM), which use conditional normalizing flows to retain exact likelihood in few-step diffusion sampling. By combining shallow invertible blocks with a deep parallel predictor, NTM achieves state-of-the-art text-to-image results in just four steps while enabling self-distillation through its precise trajectory likelihood.

TIMELINECoverage timeline

Total 1 reports · Latest first

Apple Machine Learning ResearchT1·Oct 8, 08:0078 pts
NTM: Exact Likelihood Four-Step Image Generation via Normalizing Flows
Original: Normalizing Trajectory Models
- NTM models reverse steps as expressive conditional normalizing flows, maintaining exact likelihood training even when compressed to four coarse transitions.
- The architecture combines shallow invertible blocks with a deep parallel predictor, supporting both end-to-end training from scratch and initialization from pretrained flow-matching models.
- A self-distillation mechanism based on exact trajectory likelihood allows lightweight denoisers to generate high-quality samples using the model's own induced score function.
View Evidence Original Source