Apple researchers introduce normalizing trajectory models for efficient diffusion sampling

Diffusion-based models decompose sampling into many small Gaussian denoising steps, an assumption that breaks down when generation is compressed to a few coarse transitions.