Diffusion Models Pull Early Noisy Samples Toward Dataset Center

Researchers observe that in the early, mostly-noisy stage of sampling, a diffusion model's best prediction approximates the dataset mean, pulling samples toward the data distribution's center before they bounce back onto the data manifold.

2026-09-08 ~ 2026-09-09 · 2 related posts

• Diffusion models pull early samples toward dataset center even where no data exists — alec_helbling · 2026-09-08
• Why diffusion model samples drift toward the dataset center at high noise levels — YouJiacheng · 2026-09-09