Model Already Knows the Best Noise: Bayesian Active Noise Selection via Attention in Video Diffusion Model

Researchers propose ANSE, a new framework that improves video generation quality in diffusion models by intelligently selecting initial noise seeds based on the model's internal attention patterns. The method uses Bayesian uncertainty quantification to identify high-quality seeds that produce better video quality and temporal coherence with minimal computational overhead.

• →ANSE framework selects optimal noise seeds for video diffusion models by analyzing internal model attention patterns rather than using external priors.
• →BANSA acquisition function measures entropy disagreement across attention samples to estimate model confidence and consistency.
• →Bernoulli-masked approximation enables efficient deployment with single diffusion step estimation from subset of attention layers.
• →Experiments show improved video quality and temporal coherence across diverse text-to-video models with marginal inference overhead.
• →The approach provides a principled and generalizable method for noise selection in video generation without external design constraints.