AI BriefingApple Researchers Introduce Normalizing Trajectory Models for Exact-Likelihood Few-Step Generation
Key point
Apple researchers introduced Normalizing Trajectory Models (NTM), a framework that matches strong baselines in four sampling steps while retaining exact likelihood over the generative trajectory.
Details
Apple researchers have introduced Normalizing Trajectory Models (NTM), a framework designed to address the breakdown of Gaussian denoising assumptions in compressed generation steps. Unlike existing few-step methods that rely on distillation or adversarial objectives and sacrifice the likelihood framework, NTM models each reverse step as an expressive conditional normalizing flow with exact likelihood training.
Architecture and Training
NTM combines shallow invertible blocks within each step with a deep parallel predictor across the trajectory. This end-to-end network can be trained from scratch or initialized from pretrained flow-matching models. The architecture's exact trajectory likelihood enables self-distillation, where a lightweight denoiser trained on the model's own induced score function produces high-quality samples.
Performance
On text-to-image benchmarks, NTM matches or outperforms strong image generation baselines in just four sampling steps. This approach uniquely retains exact likelihood over the generative trajectory, offering a distinct advantage over methods that abandon this probabilistic framework for speed.
This summary was generated automatically by AI. Check the original for the author's claims and context. Copyright belongs to the original author.
Our guide explains how the AI works. Report summary errors, attribution issues, or removal requests via Contact.