Apple proposes Normalizing Trajectory Models to accelerate sampling in diffusion-based generation

Apple’s Machine Learning Research team proposes Normalizing Trajectory Models (NTM), a method designed to address a core limitation of diffusion-based generative models: the assumption that sampling can be decomposed into many small Gaussian denoising steps breaks down when generation is compressed into a few coarse transitions.

Existing approaches to few-step generation—such as distillation, consistency training, or adversarial objectives—typically sacrifice the likelihood framework, which provides theoretical guarantees about the generative process. NTM avoids this trade-off by modeling each reverse step as an expressive conditional normalizing flow with exact likelihood training.

Architecturally, NTM combines shallow invertible blocks within each reverse step with a deep parallel predictor that operates across the entire trajectory, forming an end-to-end network that can be trained from scratch or initialized from pretrained flow-matching models.

A key feature of NTM is its ability to perform self-distillation: a lightweight denoiser trained on the score function induced by the model itself can produce high-quality samples in as few as four steps, according to the authors.

On text-to-image benchmarks, the team reports that NTM matches or outperforms strong image generation baselines while uniquely retaining exact likelihood over the generative trajectory, a combination previously difficult to achieve in few-step models.