content-drift

This paper investigates how training-free acceleration can silently change generated content in diffusion-based multimodal large language models, and proposes paired diagnostics and consistency-control methods to mitigate content drift.