---
format: "aidr-story-markdown/v1"
id: "a9f660862b14122b754b81c8b73ed12b4ae45c1ac84de6206c010c00eb55daad"
canonical_url: "https://aidr.today/a9f66086?lang=en"
title: "Normalizing Trajectory Models"
lang: "en"
requested_lang: "en"
available_langs: ["en","vi"]
translation_fallback: null
fallback_fields: []
published_at: "2026-10-08T00:00:00.000Z"
category: "Research"
topics: ["inference"]
source_urls: ["https://machinelearning.apple.com/research/normalizing-trajectory-models"]
summary: "Normalizing Trajectory Models (NTM) model each reverse step as a conditional normalizing flow with exact likelihood training, combining shallow invertible blocks with a deep parallel predictor, and enable self-distillation for high-quality samples in four steps while matching or outperforming image generation baselines on text-to-image benchmarks."
---

# Normalizing Trajectory Models

> [Open the canonical story](<https://aidr.today/a9f66086?lang=en>)

**Published:** 2026-10-08T00:00:00.000Z
**Category:** Research
**Topics:** inference

## Summary

Normalizing Trajectory Models \(NTM\) model each reverse step as a conditional normalizing flow with exact likelihood training, combining shallow invertible blocks with a deep parallel predictor, and enable self\-distillation for high\-quality samples in four steps while matching or outperforming image generation baselines on text\-to\-image benchmarks\.

## Sources

- [Story source](<https://machinelearning.apple.com/research/normalizing-trajectory-models>)